Public report of jingyaogong/minimind, published 14 November 2026. It has the same score, grades and number of
findings as the full report. The details of security findings (where they are in the code, which known
vulnerabilities and which passwords or keys were found) are left out, so nobody can use them against the project.
Public
Executive reportMeasured under the Code Assurance Index, an open standard for code quality, using its published scoring rules (rubric version 2026.11.4). Assert signed this survey and filed it in the standard's public register, where anyone can look it up by its survey ID. Check this survey →
Signed and filed in the standard's public register, 14 November 2026, 00:43 UTC · Survey ID cd_365745e0c70c4f7f8f646397a4056554
jingyaogong/minimind
Measured 14 November 2026, 00:42 UTC, at code version 233d00f
CAI scoreAt riskNo change since the last survey
The score is built from these lenses. Each lens grades one area of the code, such as how easy it is to change or how well it is protected. The weakest lens counts most, so one weak area pulls the score down instead of being averaged away.
A small codebase: 4,125 lines of product code in 5 projects, tests not counted · Estimated rebuild effort: under 0.1 person-years · Weakest lens: Production Readiness, 15
Findings by how serious they are
7 critical69 serious37 minor
19 dimensions are left out of the score. For each one, we either couldn't measure it in this code, or found no evidence of the safeguard we look for, such as a tested backup restore. They count neither for nor against the score, and Limitations lists each one with its reason.
This survey was produced by
Canine Code Analyzer
Producer
Canine Development
Analyzer
Canine Code Analyzer engine 1.0.0
Measured
14 November 2026, 00:42 UTC
A measurement, not a certificate. The Code Assurance Index records a number
together with the evidence it was worked out from, so anyone can recalculate it, but it doesn't certify or guarantee
the code. The standard was written by Canine Development, which also builds the Canine Code Analyzer. It is open to any tool that
implements it, and today the Canine Code Analyzer is the only one that does.
A critical security finding is open, such as a library with a known critical vulnerability or a password or key found in the code. The score itself is unchanged, but its band shows At risk until the finding is fixed. See the Security and Compliance lens.
The score is deterministic, which means the same code always gets the same score. We work it out with tools that follow published rules, and an AI model never decides it, so anyone who applies the same rules to the same code gets the same number. How it is done is explained under Methodology ▸
34of 34 dimensions measured by tools.
77of 113 findings point to an exact file and line, so you can open each one in the code. The other 36 concern the repository as a whole, such as missing documentation.
34dimensions across 6 lenses, read from 4,125 lines of product code in 5 projects.
jingyaogong/minimind is rated At risk because critical security findings are open, all of them in the packages it uses. Its score is held back because the project has no automated tests and we found no working pipeline that builds or tests changes to the code. It scores 45 out of 100, which the At risk rating leaves unchanged.
Production Readiness, at 15, is the weakest lens and counts most in the score. The best first step is adding automated tests, worth about 26 points. Raising Production Readiness to 70 would lift the score to about 66.
Six known security holes in the packages the code depends on can be used by attackers. The project has no automated tests, so a change can break it without warning. Adding automated tests and updating the vulnerable packages would close both risks.
The product code is 4,125 lines, about €3,600 to rebuild.
A lens is capped at Adequate while one of its dimensions scores below 4 out of 10, however well the rest of the lens does. That way a serious problem in one place always shows in the lens's rating.
How the score is built: what each lens counts forThe lenses don't count equally. The weakest lens counts most, the next weakest a little less, and so on, so a weak area pulls the score down instead of being averaged away. In the bar below, each part's width shows how much that lens counts and its colour shows its rating. Improving the widest part moves the score most, and as that lens improves, it counts for less, because another lens becomes the weakest.
Exception handling · Swallowed exception (caught, then discarded) trainer/train_agent.py
We compare every finding with the previous survey's. A finding that has only moved to another line in the same file counts as the same finding, so only new findings appear here. Advisory notes from the AI model are left out, because the model words them afresh in every survey.
Fixed since the last survey: 16 findings
These 16 critical or serious findings were in the previous survey, on 25 October 2026, and are gone from this one.
€1,200 to €6,100, for 20 to 64 hours of work, under 0.1 person-years for one engineer
Difficulty of the problem
Standard, judged from the kind of software and how it is built. Code that solves a harder problem, such as payments or medical bookings, costs more per line to write.
Quality factor
×0.7 at a score of 45. Better code takes more work to write, so the factor is ×1.0 at a score of 70, lower below it and higher above it.
Size
Small: 4,125 lines of product code
What it is not
Not a quotation: nobody has offered to build it for that. Not a valuation: it says nothing about what the business is worth. Not a certified figure: no professional has signed it as one.
How the estimate is made: We price the hours a team would need to write this code again, at €60 to €95 an hour. Declarations and boilerplate are priced at a fast rate and logic at a slower one, and the logic is then multiplied by the difficulty of the problem and by the quality factor. The estimate can be about 30% off either way, and it moves most with the hourly rate and the difficulty. For this code, every line is priced as straight logic, which keeps the estimate on the low side.
Top priorities
The three actions that raise the score most for the work they take. The full ranked list is in the Roadmap.
1
Resolve the 1 No automated tests finding(s) in Code Coverage.
Raises the score by about 26 points · High effort · Code Coverage
2
Add a CI workflow that builds and runs the test suite on every push/PR.
Raises the score by about 6 points · Medium effort · CI/CD gates
3
Write a diagnostic trail the user can hand back with a bug report: a log file under the platform's per-user data directory, switchable with a `--verbose`/`--debug` flag.
Start with the error paths that emit nothing — a discarded error is the failure the user cannot describe and you cannot reproduce.
Raises the score by about 6 points · Medium effort · Observability
Diagnosis: what the measurements show together
Each conclusion below combines several dimensions into one point that no single dimension shows, and lists the measurements it rests on.
Where the investment is most exposed · Value at risk
This code would cost about €3,600 to build again, under 0.1 person-years of work. Its weakest lens is Production Readiness, at 15 (Critical). Production Readiness is about whether the product is ready to run in production and to change without breaking. Because the weakest lens counts most in the score, work on Production Readiness also moves the score most.
Based on: the rebuild estimate · the Production Readiness grade · how much each lens counts
What to do: Start with the Production Readiness actions in Top priorities.
If you do one thing first · First action
Resolve the 1 No automated tests finding(s) in Code Coverage. Of all the actions in this report, it raises the score most for the work it takes: about 26 points for high effort, because it clears the 1 Code Coverage findings. The next survey will show the change in the score.
Based on: the ranked actions in Top priorities, by score gain and effort
What hard-to-read code costs each year · Cost
Four dimensions measure how easy the code is to work in: two measure how tangled its logic is, one how much of it is copied, and one how well each part sticks to a single job. In this codebase they average 6.8 out of 10. Code at that level is slower to change than clean code, and we estimate the extra work at 3% to 7% on every change. The history of the code shows how much it changes: about 2,346 lines over the last 90 days, which is about 9,514 lines a year. At that pace, the extra work comes to an estimated 0.7 to 4.2 engineer-days a year, and it grows as more code changes.
How this is estimated: The number of changed lines is measured from the history of the code. The rest rests on two assumptions of our own: code scoring 8 out of 10 or more adds no extra work, and each point below 8 adds 2.5% to 5.6% to every change, never less than 2% to 5% and never more than 40% to 60% in all; and an engineer changes 150 to 400 lines in a day. The result is a range for planning, not a measurement.
Based on: Cyclomatic Complexity · Cognitive Complexity · Code Duplication · Cohesion (LCOM4) · the code's change history
What to do: Start with the files listed under Churn × Complexity Hotspots, which change often and are hard to read, because that is where the extra work is paid most often.
Architecture — module dependency graph
Project dependencies, layered top-to-bottom; arrows show direction. Any dashed red edge points upward or sideways, which only happens where the projects form a dependency cycle. A graph without cycles has none.
Architecture: which parts of the code depend on which
The code is divided into modules, groups of code that each handle one part of the product, and a module depends on another when it uses code from it. The table lists the modules both down the side and across the top, in an order where each module should only use modules listed above it. A number in a cell means the module on that row uses the module in that column, and it counts the pairs of building blocks (types) that connect them. The one thing to look for is a mark above the diagonal line. It means two or more modules use each other in a loop (a dependency cycle), so none of them can be changed, tested or released on its own.
12 modules, with 3 dependencies between them. Every module uses only modules listed above it, so there are no loops.
Which modules use which. The module on each row uses the module in each column, and the number counts the pairs of types that connect them. A cell above the diagonal is part of a loop.
scripts uses scripts.serve_openai_api, so a change to scripts.serve_openai_api can break scripts, while a change to scripts can't break scripts.serve_openai_api.
Position
Below the diagonal: scripts uses a module listed above it, which is the healthy direction.
trainer uses trainer.rollout_engine, so a change to trainer.rollout_engine can break trainer, while a change to trainer can't break trainer.rollout_engine.
Position
Below the diagonal: trainer uses a module listed above it, which is the healthy direction.
trainer.train_ppo uses model.model_minimind, so a change to model.model_minimind can break trainer.train_ppo, while a change to trainer.train_ppo can't break model.model_minimind.
Position
Below the diagonal: trainer.train_ppo uses a module listed above it, which is the healthy direction.
At a glance — Code Health · 60% · Adequate · gated by D2, X3 ·
Findings mapped to OWASP categories; the specific CVEs/secrets are in the Security dimension cards below and findings.md (redacted only on the public version of this report).
OWASP category
Findings
Severity
A03:2025 — Software Supply Chain Failures
11
High / Critical
Roadmap
Top priorities: Resolve the 1 No automated tests finding(s) in Code Coverage; Add a CI workflow that builds and runs the test suite on every push/PR; Write a diagnostic trail the user can hand back with a bug report: a log file under the platform's per-user data directory, switchable with a `--verbose`/`--debug` flag. Start with the error paths that emit nothing — a discarded error is the failure the user cannot describe and you cannot reproduce.
Ranked by impact ÷ effort. "Helps" is the estimated gain on the 0–100 health score.
Do this
Helps
Effort
Dimension
Resolve the 1 No automated tests finding(s) in Code Coverage.
Write a diagnostic trail the user can hand back with a bug report: a log file under the platform's per-user data directory, switchable with a `--verbose`/`--debug` flag. Start with the error paths that emit nothing — a discarded error is the failure the user cannot describe and you cannot reproduce.
Record your significant design decisions. Write one short document per decision, giving its date, the situation, the decision and its consequences, and keep them together with your other design documents. A docs/adr folder with each file named by number and title is the easiest place to find them.
Bring the 7 bodies over 30 down to 30 or less in Cognitive Complexity — start with train_ppo.ppo_train_epoch (cognitive 73), train_agent.calculate_rewards (cognitive 71), train_agent.rl_train_epoch (cognitive 52).
Every finding carries one of four grades. Three say how serious it is. The fourth says this
survey could not settle it — and it is a grade, not a gap.
Critical — 7
A definite problem that already costs you something and drags the score down: a
missing authorisation check, a dependency with a known exploit, a build that does not reproduce. Failure here
tends to cause failures elsewhere.
Serious — 69
Likely wrong, but not failing yet. It degrades
the codebase over a longer horizon and can cause failures elsewhere — not urgent this week, not something to
carry for two years either.
Minor — 37
Recorded, with no effect on how the codebase functions.
Present so the survey is complete, not because it needs doing.
Could not be resolved — 19
Something this survey could not settle
from the outside, and which could be critical or serious. Either a control was required and no
positive evidence of it exists in the repository — a backup job that nothing shows was ever restored from proves
nothing about restores — or our own analysis could not run over that part of the tree. This is not a clean
result. These are excluded from the score rather than awarded a pass, so the number on the cover neither
rewards nor penalises them: if you act on this survey without resolving them, you carry that risk yourself. Each
one is named under Limitations.
Methodology & how to trust this report
This survey is a deep, periodic assessment — run each sprint, monthly, or quarterly, taking the time to go wider and deeper than a quick check and surfacing in one coherent report what you'd otherwise piece together from a dozen separate tools. It scores deterministically: the same commit yields the same score, every run. All 34 evaluated dimensions are computed by tools and static analysis; no language model sets or judges any of them. Overall confidence is 0.9, the weighted average across measured dimensions.
Every figure here is one of two kinds, and we label which: ✓ Measured — a deterministic fact (LoC, complexity, coverage); ~ Modeled — an estimate from a stated model (cost, effort, value-at-risk), always a range with its assumptions, never a precise fact. We never present a modelled estimate as if it were measured. Perfect or absent scores carry their provenance too (ADR-0011): ✓ Tool-verified means the property itself was measured across the surface; ○ Nothing flagged means the probes came back clean — a claim bounded by what a repository can show; ⊘ Not evidenced means a working control (a tested restore, an automated rollback) showed no positive evidence — absence of evidence is not evidence of a control, so it's excluded from the score rather than awarded a spurious 10.
What we checked — 34 dimensions across the health lenses
Each chip is a dimension scored from real signals across architecture, testing, dependencies, security & compliance, documentation, git-history and code quality — in one coherent pass. A surface report typically covers a handful.
How to trust any code-health report — three questions
Can you open the finding? Real findings cite a repo-relative file and line you can open at the cited line — never an absolute scratch path. Here, 77 of 113 do; the remainder are repo-wide signals — a dimension-level measurement, not a single line. (Every path in this report is repo-relative by construction: paths are normalized at the producer and the report is rejected if any rooted path leaks through.)
Is there a tool behind the number? Every score below names the method that produced it: Roslyn, git or a scanner, not a narrative.
Does re-running give the same result? Run it again on the same commit and the score — and this report, byte for byte — is identical. A report whose numbers move between runs is describing the run, not the code.
This report answers yes to all three. That's the bar to hold any assessment to.
Tools & methods
The actual versions used this run (captured at analysis time) — re-run on the same commit for the identical score.
Method
Backs
Version
Evaluator
Roslyn static analysis
Complexity, cohesion, coupling, dead code, API surface, layering
What ran differently this time — a tool absent, degraded, or that fell back to an estimate. Named openly, not folded silently into the scores. A degraded run also records its exact cause in diagnostics.md.
D16 Bus Factor — not measured this run — This is a true statement about the repository that carries nothing for its owner to act on, so it is reported here rather than as a defect in their code. Single-maintainer repository — bus factor is not applicable (24 contributor(s) across 421 commit(s) sampled, automation and bot accounts excluded). One of them holds 90% of the history; the other 23 hold 0.4% each on average, below the 5% at which there is somebody to hand the work to. That is a single maintainer with drive-by contributors, not a team whose knowledge has concentrated — so the bus factor is not applicable and there is nothing here for the owner to act on.
D20 ADR Quality — not measured this run — We could not measure this here. That is a gap on our side — a collector, parser or image we have not built yet — and it is neither a defect in this repository nor evidence that the check would have passed. The ADR expectation is gated on whether a repository ships a deployable product. Our repo-kind classifier reads .NET project files and the host and package markers of the other ecosystems it knows (a Go `package main`, a Cargo binary, a package.json `bin`, `start` script or application dev server such as Vite or Next, Python console scripts or a Django/WSGI entry point, a Spring Boot / WAR / Gradle application build, a Rack or Rails app, a Composer project, an OTP release or escript, a Dockerfile, a Procfile, a Platform.sh `.platform/routes.yaml`, and the corresponding published-package manifests, including an OTP `*.app.src`, a Drupal module's `*.info` and a Claude Code `.claude-plugin/plugin.json`), and found none here, so we could not tell what kind of repository this is and did not assess its ADR log. This is a gap in our analyzer, not a finding about this repository. No manifest, project file or entry point of any ecosystem it reads exists in this tree.
D22 Internal API Consistency — not measured this run — This is a true statement about the repository that carries nothing for its owner to act on, so it is reported here rather than as a defect in their code. This repository's source is written only in languages whose packages are published through a manifest, and no manifest in the tree makes a publication act: either there is no package manifest of any ecosystem (no package.json, pyproject.toml, setup.py, gemspec, Cargo.toml, pom.xml, build.gradle, composer.json, pubspec.yaml or mix.exs), or every one present declares in its package manager's own terms that it is not published (an npm package marked "private": true, a Cargo package with publish = false, a pubspec with publish_to: none, a Composer "project", a Poetry project with package-mode = false, or the "Private :: Do Not Upload" classifier). With no publication act there is no intentionally-exposed surface for API consistency to be judged over.
D24 Comment Value — not measured this run — We could not measure this here. That is a gap on our side — a collector, parser or image we have not built yet — and it is neither a defect in this repository nor evidence that the check would have passed. This check's reader does not cover the language this repository's product is written in, so it had nothing of the product to read. That is a gap in this analyzer's language reach — not a finding about this repository.
AX6 Interface segregation — not measured this run — We could not measure this here. That is a gap on our side — a collector, parser or image we have not built yet — and it is neither a defect in this repository nor evidence that the check would have passed. This check is computed over the public interfaces this run's compilations declare, and none was loaded, so it had nothing to read. That is a gap in this analyzer's language reach — not a finding that the repository is free of what this check looks for.
AXR1 Runtime accessibility — not measured this run — We could not measure this here. That is a gap on our side — a collector, parser or image we have not built yet — and it is neither a defect in this repository nor evidence that the check would have passed. Render strategy: boot via ToolchainServer (.). the service did not build — `python3 -m venv /tmp/wd-venv && . /tmp/wd-venv/bin/activate && pip install -q -r REDACTED && pip install -q uvicorn` exited 1: ct-overflow', '-Wsign-compare', '-DNDEBUG', '-g', '-O2', '-Wall', '-fPIC', '-I./python', '-I./lib', '-I./deps/double-conversion/double-conversion', '-I/tmp/wd-venv/include', '-I/usr/include/python3.12…; runtime evidence skipped (launch chosen because scripts/serve_openai_api.py defines the FastAPI application 'app' at module level).
C1 Data Protection — not measured this run — We could not measure this here. That is a gap on our side — a collector, parser or image we have not built yet — and it is neither a defect in this repository nor evidence that the check would have passed. These personal data controls are read from declarative annotations, request middleware, entity/column names and guard methods in a C# source model, and none was loaded on this run, so there was nothing to gather. That is a gap in this analyzer's language reach — not a finding that the repository lacks personal data controls.
C3 Audit Trail — not measured this run — We could not measure this here. That is a gap on our side — a collector, parser or image we have not built yet — and it is neither a defect in this repository nor evidence that the check would have passed. 'dataset.lm_dataset' is imported (trainer/train_agent.py:25), but no entity, schema, table or model declaration this check reads was found beside it, so whether personal data is stored could not be decided.
C4 Data Retention — not measured this run — We could not measure this here. That is a gap on our side — a collector, parser or image we have not built yet — and it is neither a defect in this repository nor evidence that the check would have passed. 'dataset.lm_dataset' is imported (trainer/train_agent.py:25), but no entity, schema, table or model declaration this check reads was found beside it, so whether personal data is stored could not be decided.
C5 Data-Subject Rights — not measured this run — We could not measure this here. That is a gap on our side — a collector, parser or image we have not built yet — and it is neither a defect in this repository nor evidence that the check would have passed. 'dataset.lm_dataset' is imported (trainer/train_agent.py:25), but no entity, schema, table or model declaration this check reads was found beside it, so whether personal data is stored could not be decided.
GD1 Unfinished & placeholder code — not measured this run — We could not measure this here. That is a gap on our side — a collector, parser or image we have not built yet — and it is neither a defect in this repository nor evidence that the check would have passed. This check reads C# syntax, F#, JavaScript/TypeScript, Dart, Ruby and Scala source only, and no C#, F#, JavaScript/TypeScript, Dart, Ruby or Scala was found, and this repository's Python is not read yet, so it had nothing of this repository's product to read. That is a gap in this analyzer's language reach — not a finding that the repository is free of what this check looks for.
IC1 Incompleteness & stubs — not measured this run — We could not measure this here. That is a gap on our side — a collector, parser or image we have not built yet — and it is neither a defect in this repository nor evidence that the check would have passed. This check reads C# syntax, JavaScript/TypeScript, Java, Dart, Ruby and Scala source only, and no C#, JavaScript/TypeScript, Java, Dart, Ruby or Scala was found, and this repository's Python is not read yet, so it had nothing of this repository's product to read. That is a gap in this analyzer's language reach — not a finding that the repository is free of what this check looks for.
M4 Documentation accuracy — measured, with a gap in what it reached — We measured this, but not all of it. What we did not reach is a gap on our side — a collector, parser or image we have not built yet — so the numbers on that dimension cover less than the repository, and the part left out is not evidence that it would have passed. M4 checks only the claim kinds adjudicated at or above 95 % precision: links and versions (PREREGISTRATION version 9), and artefact counts an order of magnitude off. Paths, command arguments, installs, environment variables, settings, projects, packages, routes, cd targets, infrastructure products and removed modules are read but not checked, and none of them is reported or counted.
P5 DR & Backup — not measured this run — This is a true statement about the repository that carries nothing for its owner to act on, so it is reported here rather than as a defect in their code. No backup/snapshot/replication config, RTO/RPO or restore-procedure documentation was found — and no production persistence was detected either (no data-access packages, no data-store services, no database resources), so there is nothing in this repository whose loss a DR control would recover. If this system's data lives in a platform or ops repo we can't see, that's where the DR evidence belongs.
S1 Web-Security Posture — not measured this run — We could not measure this here. That is a gap on our side — a collector, parser or image we have not built yet — and it is neither a defect in this repository nor evidence that the check would have passed. These web-security controls are read from declarative annotations, request middleware, entity/column names and guard methods in a C# source model, and none was loaded on this run, so there was nothing to gather. That is a gap in this analyzer's language reach — not a finding that the repository lacks web-security controls.
X16 Unfloored truncation loop — not measured this run — We could not measure this here. That is a gap on our side — a collector, parser or image we have not built yet — and it is neither a defect in this repository nor evidence that the check would have passed. This check reads C# syntax, and JavaScript/TypeScript, Java, PHP, Dart, Ruby, Scala and Kotlin source only, and no C# was loaded and no JavaScript/TypeScript, Java, PHP, Dart, Ruby, Scala or Kotlin was found in this repository, so it had nothing of this repository's product to read. That is a gap in this analyzer's language reach — not a finding that the repository is free of what this check looks for.
X18 Disposal-pattern correctness — not measured this run — We could not measure this here. That is a gap on our side — a collector, parser or image we have not built yet — and it is neither a defect in this repository nor evidence that the check would have passed. This check reads C# syntax, and JavaScript/TypeScript, Java, Kotlin, Dart, Ruby, Scala and PHP source only, and no C# was loaded and no JavaScript/TypeScript, Java, Kotlin, Dart, Ruby, Scala or PHP was found in this repository, so it had nothing of this repository's product to read. That is a gap in this analyzer's language reach — not a finding that the repository is free of what this check looks for.
X29 Per-element action decided by a fixed element — not measured this run — We could not measure this here. That is a gap on our side — a collector, parser or image we have not built yet — and it is neither a defect in this repository nor evidence that the check would have passed. This check reads C# syntax, and JavaScript/TypeScript (Vue and Svelte components included), Java, Rust, Dart, Ruby, Scala and Kotlin source only, and no C# was loaded and no JavaScript/TypeScript, Java, Rust, Dart, Ruby, Scala or Kotlin was found in this repository, so it had nothing of this repository's product to read. That is a gap in this analyzer's language reach — not a finding that the repository is free of what this check looks for.
X4 Structured logging — not measured this run — We could not measure this here. That is a gap on our side — a collector, parser or image we have not built yet — and it is neither a defect in this repository nor evidence that the check would have passed. This check reads C# syntax, and JavaScript/TypeScript, Java, Kotlin, Python, PHP, Dart, Ruby, Scala and Rust source only, and no C# was loaded, and this repository's JavaScript/TypeScript, Python and PHP make no call through a named logger, so it had nothing of this repository's product to read. That is a gap in this analyzer's language reach — not a finding that the repository is free of what this check looks for.
X5 Nullable reference types — not measured this run — We could not measure this here. That is a gap on our side — a collector, parser or image we have not built yet — and it is neither a defect in this repository nor evidence that the check would have passed. This check reads C# syntax and MSBuild project configuration, TypeScript source and its tsconfig files, Java source and its package and module declarations, PHP source and its PHPStan and Psalm configuration, Dart source and its pubspec language version, Ruby source and its Sorbet and Steep configuration, Scala source and its build's explicit-nulls, WartRemover and Scalafix settings, and Kotlin source only, and no C# was loaded, no TypeScript was found, and this repository's Python is not read yet, so it had nothing of this repository's product to read. That is a gap in this analyzer's language reach — not a finding that the repository is free of what this check looks for.
Repo exclusion declarations (.gitattributes linguist-generated/vendored, .editorconfig generated_code): none declared — every source file was scored.
Limitations & what we did not check
We assess the repository exactly as committed, and only the repository. By design we do not reach outside the source tree: the live cloud account, the running CI/CD pipeline, the host's branch-protection and approval rules, the production configuration, or a restore actually exercised against a backup are all out of scope. That boundary is a feature, not a gap — a repo-relative, deterministic scan re-runs identically on any commit and every finding opens at a real file and line, where a live audit can neither be reproduced nor traced. The visible consequence is that controls which leave no in-repo evidence are reported as "not evidenced" and excluded from the score rather than awarded a number a static scan cannot justify.
Per-dimension blind spots
For each dimension that was measured, what a static, repo-only scan structurally cannot see — the honest edge of the measurement, not a failure of it.
D1 Cyclomatic Complexity: Cyclomatic complexity counts branches statically — it cannot tell an essential decision tree from accidental tangle, nor see complexity that lives in data or configuration (large switch-case token tables, DSL lexers/parsers, data-as-code rule tables) rather than control flow: a tokenizer's many single-character cases read as high complexity though each branch is trivial.
D2 Cognitive Complexity: Cognitive-complexity heuristics approximate how hard code is to follow; genuine domain difficulty and well-named intent that eases reading are not captured.
D3 God Classes: "God class" is sized by members and responsibilities visible in the type — a deliberately broad facade over a coherent subsystem can read the same as an accidental grab-bag. For front-end JS the file-length check is cohesion-aware (a single-responsibility module — one class/IIFE — earns a 3× threshold), but cohesion is approximated from top-level declarations, not true dependency structure.
D4 Code Duplication: Duplication is token-similarity — an in-process token-stream comparison over sliding windows, with type-aware normalization — so it finds copy-paste, not semantic duplication expressed differently. Committed machine-written code (scaffolded migrations, designer/codegen output, protobuf/OpenAPI stubs, model snapshots) is EXCLUDED — its repetition is the tool's, not the team's — so the score reflects hand-written duplication only.
D5 Coupling: Coupling is measured between projects/assemblies — runtime coupling through DI, reflection, messaging or shared databases is invisible to a static reference graph.
D6 Cohesion (LCOM4): LCOM4 cohesion is syntactic — it infers connectivity from which methods touch which fields/methods by name, not from real runtime behaviour or intent.
D8 Code Coverage: Coverage is measured by building and running the test suite inside our isolated image — the target repo is never modified, and nothing on your systems runs. So coverage exists only when the suite builds and runs within the inline time budget; one that needs external services, can't build, or exceeds the budget yields no coverage (D8 then degrades to not-measured, not a low score). Line coverage also says nothing about assertion quality.
D12 Dependency Hygiene: Dependency health reads manifests and lockfiles — a vulnerability in a vendored/copied dependency, or risk from how a dependency is actually used, is outside this view.
D13 Secret Scanning: Secret detection is signature- and entropy-based on the current tree — a secret that does not match a known pattern, or one already rotated, will not be flagged (a clean scan is "nothing matched", not "no secrets exist").
D14 License Compliance: License compatibility is checked against declared package metadata and a policy — mislabelled or missing license metadata, and obligations that depend on how you distribute, are not resolved here.
D15 Churn × Complexity Hotspots: Churn hotspots come from git history — a freshly imported or squashed repository has no churn signal, and recent rewrites can mask a historically risky file.
D17 Explicit Debt: Acknowledged-debt signals (TODO/FIXME, suppressions, dead code) are textual — undocumented debt that nobody marked, and debt that lives in design rather than annotations, is invisible. Committed machine-written code (scaffolded migrations, designer/codegen output, generated stubs) is excluded — it is never the team's dead code to delete.
D28 Secrets (history): Secret-history scanning sweeps the git log for known patterns — a secret that predates the available history, or never matched a signature, is not found (clean means "nothing matched in the history we can see").
D29 Static Analysis (SAST): SAST findings are pattern-based (semgrep) — it finds classes of bug it has rules for; logic flaws, auth/authorization gaps and issues needing runtime context are out of reach (and clean means "no rule matched").
D30 Dependency Vulnerabilities: CVE matching depends on accurate package/version metadata and on the advisory databases — a vulnerability with no published advisory, or in code not declared as a dependency, is not seen. Coverage needs a RESOLVED graph: an unpinned REDACTED, or a pom without a resolved build, yields partial coverage rather than a clean verdict. An ecosystem the analyzer cannot scan is reported as unmeasured, never as clean.
D34 Knowledge Freshness: Freshness is decayed commit RECENCY, not comprehension — code read often but rarely committed reads as orphaned, and stable code that genuinely needs no changes is penalised the same as forgotten code; bot/squash commits distort it like the bus factor.
D35 Change Coupling: Change coupling is co-change in COMMITS — files split across separate commits, or coupled only through a shared config/build step, read as uncoupled, and a sweeping commit (rename/format) is excluded so it doesn't couple everything. It shows that files change together, not WHY: a high coupling can be a healthy cohesive pair as readily as a hidden leak.
D43 Malicious Dependencies: Only packages some vulnerability database has already NAMED as malicious are seen — a compromise published in the last hours, or never reported at all, is invisible here, and this dimension reading 10 is not evidence that a dependency is trustworthy. There is no typosquat or dependency-confusion analysis: a package nobody has reported is simply absent from the feeds. Coverage is the dependency scan's: an ecosystem that could not be scanned is disclosed as unmeasured, never as clean.
D44 Platform End-of-Life: The support table is FROZEN, so it goes out of date by losing RECALL: a release that ended support after the table was written is missed until the table is refreshed, and this dimension reading 10 is not evidence that a platform is current. Only platforms the repository DECLARES in a place this pass reads are seen — a runtime named only in a Dockerfile (D31's subject), in a CI workflow (D29's), or in a file this pass does not parse (go.mod, a Gemfile ruby directive) is invisible here, which is why a repository declaring none of them abstains rather than scoring. Only frameworks with a PUBLISHED support policy are tracked: React, Flask and Express publish none, so their age cannot be judged and their absence from a report is not a statement that they are supported.
AX10 Code composition: Role is inferred from namespace/folder convention, not semantics — a domain concept living in a folder named "Services" reads as application, and the split is lines-of-code, not business value. The business-logic-share score is a SOFT, FLOORED signal: it contributes to the Architecture lens but is floored at the Critical gate, so an infrastructure-heavy design (a gateway, an ETL, a driver) is legitimately low without being nuked to zero.
M4 Documentation accuracy: Documentation accuracy checks links, versions and artefact counts against the tree: other claims (paths, commands, settings, routes) are read but not checked, and it cannot run the onboarding or measure how long a real new joiner takes.
What it measures: How tangled the control flow is — methods with many branches are hard to test and change.
Method: Cyclomatic complexity per method (1 + decision points), computed exhaustively across production source; test projects separated by convention. Deterministic.
+ 1 more group(s) — more in Appendix A; the complete list is findings.md.
What to do
Bring the 6 bodies over 15 down to 15 or less in Cyclomatic Complexity — start with train_ppo.ppo_train_epoch (cyclomatic 29), train_agent.calculate_rewards (cyclomatic 26), train_agent.rl_train_epoch (cyclomatic 26). — This score is capped by its worst body, so a finding fixed alone moves it by almost nothing — the next one down takes its place. Refactoring these 6 together lifts Cyclomatic Complexity from 7.8 to about 10.0/10, projected with the scoring formula itself and assuming each lands exactly at 15.
Detailed fixes: d1_recommendation.md · top locations in Appendix A, every location in findings.md.
What it measures: How hard the code is for a person to follow, beyond raw branching.
Method: Cognitive complexity per method (Sonar-style nesting-penalized score), computed exhaustively over production code, excluding test projects. Deterministic.
+ 11 more group(s) — more in Appendix A; the complete list is findings.md.
What to do
Bring the 7 bodies over 30 down to 30 or less in Cognitive Complexity — start with train_ppo.ppo_train_epoch (cognitive 73), train_agent.calculate_rewards (cognitive 71), train_agent.rl_train_epoch (cognitive 52). — This score is capped by its worst body, so a finding fixed alone moves it by almost nothing — the next one down takes its place. Refactoring these 7 together lifts Cognitive Complexity from 3.2 to about 6.3/10, projected with the scoring formula itself and assuming each lands exactly at 30; a cleaner split scores higher.
Detailed fixes: d2_recommendation.md · top locations in Appendix A, every location in findings.md.
Do you agree with this assessment?
D3 · God Classes10.0 / 10Exemplary✓ Tool-verified
What it measures: Over-large classes that try to do too much ("god classes").
Method: God-class detection by line and method-count thresholds per logical type (partial classes unified), filtered for generated code and registration/contract false positives. Deterministic.
What it measures: How much of the production code is copy-pasted: the share of production lines inside a reported copy (every copy of a duplicated block or of an edited member, each line once), 10 at none and 0 at a quarter of the lines.
Method: Code duplication via token-stream sliding windows with type-aware normalization (locals masked, type names preserved), scored on the share of production lines inside a reported copy (each line once; 10 at none, 0 at a quarter). Deterministic.
20 duplicated block group(s) detected. 9.3 % of the production lines (385 of 4,125) are in a reported copy. A further 8 rows report members as variants of one another; the share counts each copied line once, however many rows cite it. 2 of the 28 are in trees this repository does not ship — vendored, example/demo, project-template, fixture and benchmark code — and are ranked below the shipped groups rather than excluded from them: the duplication there is real and is still counted in this dimension's score. The dimensions that publish a production-file census leave those trees out of theirs, so this count is deliberately drawn over the wider population.
Near-duplicate member family (3 members, 32 shared lines)trainer/train_full_sft.py:25
+ 16 more group(s) — more in Appendix A; the complete list is findings.md.
What to do
Resolve the 4 Duplicated block (6 lines × 2) finding(s) in Code Duplication — start with train_agent.py (2), lm_dataset.py, web_demo.py. — One of this dimension's main actionable groups (4 warning-level).
Resolve the 3 Duplicated block (5 lines × 2) finding(s) in Code Duplication — start with train_agent.py, eval_llm.py, convert_model.py. — One of this dimension's main actionable groups (3 warning-level).
Resolve the 2 Duplicated block (8 lines × 2) finding(s) in Code Duplication — start with train_ppo.py, train_distillation.py. — One of this dimension's main actionable groups (2 warning-level).
Detailed fixes: d4_recommendation.md · top locations in Appendix A, every location in findings.md.
Do you agree with this assessment?
D5 · Coupling6.8 / 10Adequate✓ Tool-verified
What it measures: Whether volatile projects sit underneath others that depend on them (so their churn ripples upward), and whether project dependencies form cycles. A widely-depended-on but stable shared/kernel project is healthy, not penalised.
Method: Dependency cycles via elementary-DFS over real .csproj references, plus Martin instability (afferent/efferent) per project. Exhaustive over the reference graph, deterministic.
Coverage: Exhaustive · type-level: afferent/efferent coupling + cycles computed over every production type — the population is all types, not a name convention.
5 production modules (source-import), 0 dependency cycle(s), 0 unstable depended-on module(s). No manifest declares this repository's modules, so the modules are its top-level source directories and the edges the imports between them; 2 module(s) off the main sequence.
Off the main sequence: dataset · ×2
What to do
Resolve the 2 Off the main sequence finding(s) in Coupling. — One of this dimension's main actionable groups (2 warning-level).
Detailed fixes: d5_recommendation.md · top locations in Appendix A, every location in findings.md.
What it measures: Whether a class's methods are focused on a single responsibility.
Method: LCOM4 cohesion per production class with at least two methods: connected components of methods sharing state or calls, computed syntactically. Deterministic, not a proxy.
Coverage: Exhaustive · type-level: LCOM4 cohesion computed over every production class — the population is all types, not a name convention.
What it measures: How much of the code is actually exercised by tests.
Method: Coverage from the analyzer's own run of the suite, computed per-file with structured exclusions for generated, trivial, and glue code. A coverage report found in the repository is never read. When the suite can't be built/run in-image, coverage is reported NOT-MEASURED (excluded from the score) with the precondition to make it measurable — never a LoC-ratio proxy folded in as if measured. Deterministic.
What it measures: Whether dependencies are current, not deprecated, and not bloated. Known vulnerabilities are D30's.
Method: Manifest scan via dotnet list package across all projects; worst-signal-per-package deduction (capped-linear for deprecation, prerelease and outdated) per KLoC. Known-vulnerable packages are D30's and are not charged here (rubric-2026.11.2). Exhaustive.
28 outdated, 0 yanked pinned Python distributions. Only EXACT pins are graded for currency: a floor or a range already installs the newest release it admits, so reporting one would report this repository for being current. Whether a deliberate upper cap has itself gone stale is a different and weaker question, and is not asked. Whether any distribution is UNMAINTAINED is not graded — PyPI publishes no maintenance status, and release age does not stand in for one. Lockfiles are not read (poetry.lock, uv.lock, Pipfile.lock, pdm.lock), so a lock-resolved install is outside this verdict. Known CVEs in this dependency graph are D30's question.
Outdated: datasets · ×28
Detailed fixes: d12_recommendation.md · top locations in Appendix A, every location in findings.md.
What it measures: Whether any secrets (keys, tokens, passwords) have leaked into the code.
Method: In-process native secret scanner (entropy plus signature patterns) across all tracked files; no external tool. A clean result is a measured 10, not no-data zero. Deterministic.
What it measures: Whether the licenses of third-party packages are compatible with your policy.
Method: Third-party package licenses resolved from declared package metadata and checked against the configured policy (allow/deny/copyleft). Deterministic; clean = no incompatible license found at metadata depth.
0 of 146 shipped Python distribution(s) use a banned license. Licences were resolved from PyPI over the distributions a consumer installs — this repository's 31 declared runtime requirement(s) closed transitively over each distribution's published `requires_dist` (115 reached that way). Requirements it states ONLY under an extra, a PEP 735 dependency group, a Poetry dev group or a dev-named requirements file are excluded: pip does not install any of them for a consumer. ★ This repository commits no dependency lockfile that this pass reads, so each licence is the one PyPI publishes for the distribution's CURRENT release rather than for a pinned version. 15 of them publish no licence on PyPI this pass can read; that is missing data, not a violation, and none of them is charged.
What it measures: Files that change often and are also complex — the riskiest hotspots.
Method: Per production file churn times cyclomatic complexity over a rolling window, computed from git and Roslyn/JS/Razor analysis. Exhaustive, deterministic per commit date.
Resolve the 3 Hotspot finding(s) in Churn × Complexity Hotspots — start with train_agent.py, train_ppo.py, train_grpo.py. — One of this dimension's main actionable groups (3 warning-level).
Resolve the 1 Repeated repair finding(s) in Churn × Complexity Hotspots — start with eval_llm.py. — One of this dimension's main actionable groups (1 warning-level).
Detailed fixes: d15_recommendation.md · top locations in Appendix A, every location in findings.md.
What it measures: Acknowledged debt left in the code — TODOs, dead code, suppressed warnings.
Method: Roslyn syntactic debt markers (suppressions/TODO/FIXME/HACK/empty-catch/commented-code/Obsolete) plus SymbolFinder dead-code analysis; weighted-debt-per-KLoC density deducted 2.0x per unit. Deterministic, exhaustive.
0 deducted task-comment markers across 4125 LoC (0.0/KLoC) → score 10.0. Task comments only: this repository's language is read without a compiler, so D17's suppression, dead-code and commented-out-code arms did not run and this score counts fewer marker kinds than a .NET repository's would.
What it measures: Whether any secrets were ever committed — scanned across the full git history, not just now.
Method: Secret scan via TWO gitleaks detect passes in an isolated checkout — the full git history, then a second --no-git pass over the working tree as it stands — merged and de-duplicated by (rule, file, line); each match flagged High. Both invocations are recorded in the audit trail. Exhaustive; when the tool is absent, or when its output cannot be parsed into the expected shape, the dimension is WITHHELD as an explicit measurement gap on our side — unscored and excluded from the lens, never a hedged middling score.
What it measures: Real static-analysis (SAST) findings — likely security bugs in the code, any language.
Method: Polyglot static analysis via semgrep across the repo using the pinned, image-baked p/security-audit + p/owasp-top-ten rulesets (no scan-time registry fetch); severity rules (ERROR/WARNING/INFO) map to a full-band severity-weighted score. Exhaustive, deterministic; degrades on parse failure.
Coverage: semgrep pattern rules over all files — exhaustive for the rule set, blind to classes of bug without a rule (clean = no rule matched).
What it measures: Whether any dependency has a known published vulnerability (CVE), direct or transitive, in ANY ecosystem the repository declares — Dart pub, Elixir and Erlang via Hex, Go modules, Java and Kotlin via Maven/Gradle, JavaScript/npm, .NET/NuGet, PHP/Composer, Python/PyPI, RubyGems, Rust/Cargo and Swift.
Method: Dependency-CVE scan across every ecosystem the repository declares, scored ONCE. Three sources are unioned and deduplicated by advisory identity (rule id + alias closure, CVE<->GHSA) scoped to package+version, keeping the worst severity: `osv-scanner --recursive` over osv.dev for Dart pub, Elixir/Hex (and Erlang, whose `rebar.lock` syft first converts to a CycloneDX SBOM the scanner reads, with rows attributed back to the lock), Go, Java and Kotlin via Maven/Gradle (and Scala, whose sbt build's pinned direct declarations are written into a CycloneDX SBOM the scanner reads, with rows attributed back to the build file), npm, PHP/Composer, Python/PyPI, RubyGems, Rust/Cargo and Swift; `trivy fs --scanners vuln` for npm lockfiles; and `dotnet list package --vulnerable --include-transitive` for NuGet (with per-advisory collapse of the project x target-framework fan-out), plus a DECLARED-dependency arm that resolves a published gem's gemspec against rubygems.org where no Gemfile.lock is committed. `SeverityScore(c,h,m,l, normalizer 8.0)`. NotApplicable only when NO ecosystem is readable; if any applicable ecosystem could not be scanned the findings are REPORTED and the score is withheld. Supersedes the npm and OSV arms, retired 2026-09-05. Advisory since rubric-2026.11.2 (CAI decision 0006): every row is a match against a vulnerability database that changes daily, so the reading is shown and never moves a scored dimension, lens or the headline.
Resolve the 5 Medium CVE finding(s) in Dependency Vulnerabilities — start with REDACTED (5). — One of this dimension's main actionable groups (5 warning-level).
Resolve the 4 High CVE finding(s) in Dependency Vulnerabilities — start with REDACTED (4). — One of this dimension's main actionable groups (4 issue-level).
Resolve the 2 Critical CVE finding(s) in Dependency Vulnerabilities — start with REDACTED (2). — One of this dimension's main actionable groups (2 issue-level).
Detailed fixes: d30_recommendation.md · top locations in Appendix A, every location in findings.md.
What it measures: Whether anyone still has living knowledge of each file, or it has been orphaned — last understood long ago by someone now gone quiet. The sibling of the bus factor: D16 asks who owns it, D34 asks whether anyone still knows it.
Method: File orphaning as total living-knowledge decay below one focused-commit's worth within a year, computed per-file from the D16 decay model. Exhaustive, deterministic over fixed history.
Every significant source file has living knowledge — recently and meaningfully worked. Counted over 19 of the 22 production source files in this repository: the rest are under the ~2,400-byte size floor this dimension measures over.
What it measures: Whether files that change together actually belong together — pairs that repeatedly co-change in git history despite having no explicit code dependency, surfacing the hidden/logical coupling (and boundaries in the wrong place) a static scan can't see.
Method: Pairwise co-occurrence over the per-commit file sets in git history (production source only — tests and generated dropped): Degree-of-Coupling = shared ÷ min individual revisions, reported above noise floors (each file ≥10 revisions, ≥5 shared commits, ≥50% strength); sweeping commits excluded. Deterministic over fixed history.
Coverage: Population: PRODUCTION source files only — test and generated files are dropped before pairing, so a class co-changing with its own test (trivially ~100%) can't drown the real production↔production coupling. Pairs ranked by Degree-of-Coupling. A non-source file is never a coupling PARTICIPANT either: documentation, schemas, config and data files are dropped with the rest, so a code↔docs pair — a command and the reference page that restates it — is not reported however strongly the two co-change; nor is coupling that runs THROUGH a build step or config file.
Resolve the 1 Change coupling clique finding(s) in Change Coupling — start with train_agent.py. — One of this dimension's main actionable groups (1 warning-level).
Detailed fixes: d35_recommendation.md · top locations in Appendix A, every location in findings.md.
What it measures: Whether any dependency the repository declares is published as MALICIOUS rather than merely vulnerable — a package that is an attacker's work, in any ecosystem osv-scanner reads. Reported apart from D30 because the answer is binary: there is no safe version to upgrade to, and the fix is to remove the package and rotate every credential it could have read.
Method: The same dependency scan D30 reads, partitioned on the scanner's own classification rather than rescanned: a row is MALICIOUS when its id is in the `MAL-` space (the ossf/malicious-packages feed) OR its `database_specific.cwe_ids` carries `CWE-506` ("Embedded Malicious Code"). Both channels are structural; the summary text is deliberately NOT read, because a malicious-package record whose summary says only "Critical severity vulnerability" is a real shape ([GHSA redacted]) and a text matcher misses it. Scored BINARY: any surviving row is 0, whatever its severity and however many CVEs sit beside it — a hostile dependency is not a quantity. Applicability and degradation are D30's: NotApplicable only when no ecosystem is readable, and an unscannable ecosystem degrades rather than reading clean. Advisory since rubric-2026.11.2 (CAI decision 0006): the malicious-package feed is a database that changes daily, so the reading is shown and never moves a scored dimension, lens or the headline.
What it measures: Whether anyone still ships security patches for the platform this repository RUNS ON — the runtime it pins and the framework majors its own constraints hold it to. Separate from D12 because the question differs: a current Django on an end-of-life Python is perfectly up to date and completely unsupported, and the fix is a migration rather than a version bump. What the repository says it merely SUPPORTS is never charged.
Method: End-of-life PLATFORM read from the repository's own declarations and graded against a FROZEN, dated table of vendor support dates — no network, no feed, no API, so this dimension answers identically inside a closed scan fence. Two subjects: a RUNTIME the project pins (a single or all-end-of-life TargetFramework, a .nvmrc or .python-version, a requires-python CAP) and a FRAMEWORK major a dependency constraint cannot move off (a caret, tilde or exact version; `vue@^2.7.16` pins Vue 2). A FLOOR is deliberately never charged — `requires-python = ">=3.8"` states what a package SUPPORTS, not what it runs on — and a multi-target project is charged only when EVERY target is out of support. Runtime 4.0/product capped 8.0, framework 1.5 capped 4.5. The table is safe to freeze because a statement about support that ended in the past cannot become false: it loses recall as it ages, never precision, and a test asserts every entry predates the freeze date. Disjoint from D31 (a container image's OS layer) and D29 (the toolchain a CI workflow installs). Abstains when the repository declares no platform this pass reads — never scores it clean.
0 end-of-life runtime(s) and 0 end-of-life framework(s), read from 0 platform declaration(s) and 31 dependency declaration(s). This dimension reads what the repository says about ITSELF — a pinned target framework, a version file, a capped requires-python, a Rust toolchain pin, a framework major a constraint cannot move off. A FLOOR is deliberately never charged: `requires-python = ">=3.8"` states what the package SUPPORTS, not what it runs on, and a well-maintained library declares exactly that while running its own CI on a current release. The end-of-life facts are FROZEN and dated, so this dimension needs no network and answers identically inside a closed scan fence; as the table ages it loses recall and never precision, because a statement about support that ended in the past cannot become false. The OS layer of a container image is D31's question and the toolchain a CI workflow installs is D29's; this row is neither.
Other · Architecture — How the codebase splits by code ROLE — domain, application, infrastructure, test, generated. The significance map behind the knowledge/coupling weighting, and a DDD signal in its own right: a thin domain core under fat infrastructure is the anemic-domain smell, quantified. How each file's role is decided, because the split is only as good as that: a generated name or a build-output tree makes it Generated, a test project makes it Test, and otherwise the file's NAMESPACE and PATH words are matched against fixed vocabularies in a fixed ORDER — domain, then infrastructure, then application — so a file whose words hit two layers is counted under the earlier one. A production file matching none of them counts as application, so that share reads 'application or unclassified' rather than a measured application layer. Roles come from naming convention, never from what the code does. On this repository the split was taken from the source tree on disk rather than from a loaded .NET workspace, so a file's role is decided by its PATH segments alone — no declared namespace was available to add to the evidence — and generated output is excluded from the census entirely rather than counted as a generated share.
Method: Roslyn line-count by code ROLE: every source file classified Domain/Application/Infrastructure/Test/Generated by namespace + path convention (the shared CodeRoleClassifier), then significant lines summed per role. Deterministic; the advisory score is the business-logic (domain+application) share of production code.
Coverage: Population: ALL source files, each bucketed into ONE of five roles (Domain/Application/Infrastructure/Test/Generated) by namespace + path convention — a file whose layer isn't named in the convention falls to Application (the neutral default), and the split is line-count, not semantic depth or business value.
Other · Architecture — Whether the project-reference graph is acyclic (cycles block independent build/deploy and signal eroding boundaries).
Method: Project reference cycles via elementary-DFS over real .csproj references, using the engine shared with D5/D7; cyclic versus acyclic. Exhaustive, deterministic.
Other · Architecture — Whether dependencies point inward (Domain ← Application ← Infrastructure/Web) — the clean-architecture dependency rule, checked across the project graph.
Method: Layer violations by name-segment inference (Domain/Core to Application to Infrastructure/Web) over the project-reference graph. Exhaustive over all projects, deterministic.
Maturity · Maturity — Whether the repository and its projects have a README, and whether it is substantive.
Method: Filesystem scan: README presence, word count, and headings for depth; git history for staleness. Exhaustive across root and project dirs, deterministic.
Maturity · Maturity — Whether key decisions (ADRs) and the high-level shape (C4/diagrams) are written down.
Method: Filesystem scan: ADR folder/naming conventions or content, plus Mermaid/PlantUML/C4/architecture.md discovery. Exhaustive, deterministic.
No Architecture Decision Records found — no conventional ADR directory, no numbered `NNNN-title` documents in any markup this check reads, and nothing ADR-shaped by content. Design rationale recorded elsewhere (a design-notes tree, a mailing list, pull-request discussion) is not visible to this check and is not re-findable per decision, so a future maintainer cannot ask why one choice was made and get an answer.
What to do
Record your significant design decisions. Write one short document per decision, giving its date, the situation, the decision and its consequences, and keep them together with your other design documents. A docs/adr folder with each file named by number and title is the easiest place to find them.
Maturity · Maturity — Whether the repository is organised deliberately: production code and tests are set apart in a way its ecosystem or the repository itself declares.
Method: Filesystem scan: src/test folder separation and namespace-prefix consistency (majority RootNamespace agreement). Exhaustive across projects, deterministic.
Most of the production code is in directories of its own, but 1 application file(s) sit at the repository root beside the build and tooling files (`eval_llm.py`).
Maturity · Maturity — Whether the project's current-state documentation describes the code that exists.
Method: Every current-state document's checkable claims looked up in the tree (links, versions, and artefact counts an order of magnitude off), plus two README checks: an advertised infrastructure product with no footprint, and a module the CHANGELOG records as removed. Deterministic, no model. Advisory.
Do you agree with this assessment?
P1 · CI/CD gates0.0 / 10Critical✓ Tool-verified
Readiness · Readiness — Whether an automated pipeline builds and tests every change.
Method: Filesystem scan: CI workflow files (.github/workflows, .gitlab-ci.yml, etc.) for build and test stages. Exhaustive, deterministic.
No CI workflow found (.github/workflows, azure-pipelines.yml, .gitlab-ci.yml, …) — changes aren't gated by an automated build/test.
What to do
Add a CI workflow that builds and runs the test suite on every push/PR.
Do you agree with this assessment?
P2 · Observability0.0 / 10Critical✓ Tool-verified
Readiness · Readiness — Whether the code is diagnosable in production — structured logging, tracing/metrics, health checks.
No logging or diagnostics emission found. This repository ships a binary its users launch rather than a service you operate, so there is no operator to page — but when a run fails on a user's machine, neither they nor you have anything to read.
What to do
Write a diagnostic trail the user can hand back with a bug report: a log file under the platform's per-user data directory, switchable with a `--verbose`/`--debug` flag. Start with the error paths that emit nothing — a discarded error is the failure the user cannot describe and you cannot reproduce.
Readiness · Readiness — Whether SAST, secret/dependency scanning and performance benchmarking are wired in (presence, not runtime).
Method: Filesystem scan: SAST configuration, dependency-update automation, secret scanning, and a benchmark harness or benchmark step — in this repository's own ecosystem. Exhaustive, deterministic.
No static application security testing detected. For this repository's stack, add bandit, `semgrep --config=p/python`, or CodeQL's python pack — this repository has no CI pipeline yet, so run it locally to clear the existing findings, then make it a step of the first workflow you add so a regression fails the build. What was searched, so you can tell an absence from a miss: the 0 CI workflow file(s) in this repository, and the scanner and linter configuration checked in beside them. A scan that runs outside CI, one configured in your forge's web UI rather than in a committed file, or a tool whose name is none of those this check carries, is not seen — if that is your case the row is wrong, and saying so is more useful than adding a second scanner.
What to do
Run what this repository's stack ships: bandit, `semgrep --config=p/python`, or CodeQL's python pack — locally for now, since there is no CI pipeline here yet, and as a step of the first workflow you add so a security regression fails the build instead of landing.
Enable Dependabot/Renovate or a dependency-review gate.
Add gitleaks/trufflehog in CI to block PRs that introduce committed secrets.
Readiness · Performance — Whether asynchronous code stays responsive — it avoids sync-over-async blocking (a .NET .Wait()/.GetAwaiter().GetResult(), a time.sleep or blocking HTTP call inside a Python coroutine, a *Sync call inside an async JavaScript function, block_on inside a Rust async fn, runBlocking inside a Kotlin suspend function, block() inside a Reactor publisher) that stalls a thread or event loop and risks deadlock, and, where the code is a reusable library on .NET, awaits with ConfigureAwait(false) so it never captures and stalls its caller's context.
Method: Production-source scan: sync-over-async blocking counted everywhere — .Wait()/.GetAwaiter().GetResult() in .NET; off .NET, read from each language's code model, a blocking call inside an async function (Python, TS/JS, Rust, Kotlin, Swift, Dart) or inside a Java method returning a Reactor Mono/Flux or a Scala method returning a Future or effect — and, for a .NET library with ≥5 awaits, the share of awaits using ConfigureAwait(false). Deterministic, syntax/text detection.
Other · Code Health — Whether any branch is dead by construction — a switch arm whose label can never equal a case-normalised subject, or an `else if` whose predicate the arm above has already swallowed.
Method: Roslyn syntax + semantics: switch labels compared against the subject's own case normaliser, and if/else-if chains checked for a literal an earlier arm's containment test already swallows. Deterministic, provable per finding. Advisory.
Other · Code Health — Whether a value the caller is invited to supply is the value the type actually uses — a constructor parameter stored in a private field that nothing ever reads while the default it was given is spelled out a second time at the site that should have read it, a keyed lookup that falls back to a different setting than the one its key names while the same type falls back to the matching one for that same key, or a culture-sensitive parse given no format provider by a type that feeds its own settable culture to the same kind of parse elsewhere, or a number an entry point writes to standard output under the host's culture while the same entry point formats with the invariant culture elsewhere. Either way, every caller who supplies a value silently gets something else.
Method: Roslyn syntax: private instance fields of a non-partial type assigned in a constructor from one of its own parameters with a `??` fallback, checked for whether anything in the type body reads the field and whether that same fallback expression is spelled out again outside the constructor; and `??` fallbacks onto a member access from a lookup call carrying exactly one string literal, grouped by that key across the type and checked for a fallback member whose folded name disagrees with the key while a sibling site for the same key agrees with it. On a repository with no .NET source the first two arms read JavaScript/TypeScript off the engine’s own token stream (tests included, bundles and vendored paths not): a `#x`, `private` or `private` parameter-property instance field filled in the constructor from a parameter (or one member of one) through `??`/`||` or a parameter default, never read anywhere in the file by name, whose constructed default is spelled again in the class body; and `lookup("key") ?? s.member` grouped by key per class, or per module outside every class. The culture arm has no JavaScript counterpart: its parses take no locale. Deterministic, provable per finding. Advisory.
Do you agree with this assessment?
X28 · Index access outside its own emptiness guard10.0 / 10Exemplary○ Nothing flagged
Other · Code Health — Whether a condition that tests a value for emptiness indexes that same value only where the test holds — an `||` written one parenthesis too far to the left leaves an index access outside the guard beside it, so the empty case the guard exists to anticipate reaches the index and throws.
Method: Roslyn syntax only, no semantic model: the OUTERMOST `&&`/`||` of every boolean condition, read for a symbol the condition tests for emptiness (`string.IsNullOrEmpty`/`IsNullOrWhiteSpace`, a `Length`/`Count` comparison against a literal, `Any()`, a `Length`/`Count` pattern, or a comparison against `""`) and ALSO indexes. Each `symbol[...]` access is placed by a boolean-reachability walk from the access up to the outermost connective: an access is COVERED when some enclosing step has it in the right operand and the left operand, under the truth value that step forces, proves the symbol non-empty — a recursion over `&&`/`||` whose true- and false-directions are asymmetric. A finding needs BOTH an uncovered access and a covered one on the same symbol in the same condition, which is the agreeing twin that separates a misplaced parenthesis from an unrelated length test. Bare index accesses with no emptiness test in the condition are neither counted nor reported; a non-identifier receiver and a lambda nested inside the condition are outside the population. Deterministic, provable per finding. Advisory.
Other · Code Health — Whether exceptions are handled rather than silently swallowed or rethrown with lost stack traces.
Method: Roslyn syntax scan: every catch clause counted; empty catches and bare rethrows flagged. Population is all catch clauses, not estimated. Deterministic, hard fact.
`except` takes every exception and records none of it — the body neither logs it, rethrows it, nor even names it, so the failure is discarded as completely as by an empty catch and only the substituted value survives. Log it through whatever this codebase already uses to report problems, narrow the catch to the exception this call can actually raise, or say in a comment on the catch why the failure genuinely cannot matter. (×4) — trainer/rollout_engine.py:204, trainer/train_agent.py:80, trainer/train_agent.py:135, …
`except Exception` takes every exception and records none of it — the body neither logs it, rethrows it, nor even names it, so the failure is discarded as completely as by an empty catch and only the substituted value survives. Log it through whatever this codebase already uses to report problems, narrow the catch to the exception this call can actually raise, or say in a comment on the catch why the failure genuinely cannot matter. — trainer/train_agent.py:89
What to do
Swallowed exception (caught, then discarded)
Do you agree with this assessment?
Reference — by lens
The score is the rank-weighted fold of these lenses (worst-heaviest), each including its meta-dimensions; a lens with a Critical contributor is capped at Fair (its band reads "gated by …") and is never the strongest area however high its average.
Unscored — 1 check(s) recorded observations but carry no score
These checks ran and found something, but they do not carry a score — either by design (an advisory check reports evidence rather than grading it) or because they could not be scored here. They are excluded from the score for that reason, not because there was nothing to see.
X10 Duplicated predicate — 6 observation(s) recorded · Advisory — this card reports evidence and never carries a score.
Not evidenced — 6 control(s) we could not find positive evidence for
These checks grade a working control, and the repository shows no evidence of one. That is deliberately not scored as a zero: a repository cannot show an ops runbook, a database TTL or an infrastructure-side audit log, so absence of evidence here is not evidence the control is missing. It is also not a statement that the check is irrelevant to this codebase — the thing it grades applies; we just could not see it. Excluded from the score either way.
C3 Audit Trail — Not assessed: this repository stores data through 'dataset.lm_dataset' (trainer/train_agent.py:25) but declares no persisted model this check reads (Django, SQLAlchemy, SQLModel, Peewee, Tortoise and MongoEngine model fields, Alembic and Django migration columns, CREATE TABLE and ALTER TABLE … ADD columns in SQL files). Personal data may be written through raw SQL or a hosted backend, so whether its changes need an audit trail could not be established, and the trail's absence is not scored.
C4 Data Retention — Not assessed: this repository stores data through 'dataset.lm_dataset' (trainer/train_agent.py:25) but declares no persisted model this check reads (Django, SQLAlchemy, SQLModel, Peewee, Tortoise and mongoengine model fields, Alembic and Django migration columns, SQL schemas). Personal data may be written through raw SQL or a hosted backend, so whether a retention limit applies here could not be established, and its absence is not scored.
C5 Data-Subject Rights — Not assessed: this repository stores data through 'dataset.lm_dataset' (trainer/train_agent.py:25) but declares no persisted model this check reads (Django, SQLAlchemy, SQLModel, Peewee, Tortoise and mongoengine model fields, Alembic and Django migration columns, CREATE TABLE and ALTER TABLE … ADD columns in SQL files). Personal data may be written through raw SQL or a hosted backend, so whether data-subject rights apply here could not be established, and their absence is not scored.
P4 Deployment & Rollback — not evidenced — no deploy/rollback/approval signal in the repo; absence of evidence is not evidence of a manual release
P5 DR & Backup — not evidenced — repo shows no backup/RTO/RPO controls; absence of evidence is not evidence of a working control
P6 Release Hygiene — not evidenced — no changelog, version stamp or semver release tag in the repo
Not included — 75 check(s) not relevant to this codebase
These checks had nothing to measure here (no tests, no git history, the codebase is small, or the architecture style doesn't apply), so they're omitted above rather than scored low.
AC1 Text alternatives — No web markup found — accessibility is not applicable to this repository.
AC2 Forms & labels — No web markup found — accessibility is not applicable to this repository.
AC3 Page structure — No web markup found — accessibility is not applicable to this repository.
AC4 Keyboard semantics — No web markup found — accessibility is not applicable to this repository.
AC5 ARIA correctness — No web markup found — accessibility is not applicable to this repository.
AC6 Visual & motion safety — No web markup found — accessibility is not applicable to this repository.
AX1 Captive dependencies — Not applicable: this repository's Python imports no dependency-injection container and defines no container of its own — nothing that both registers and resolves bindings, and nothing that names two lifetimes — so nothing holds one lifetime's instance while handing out another's.
AX2 Stateful singletons — No container singleton was found, so there is no shared instance for concurrent requests to race on. No function in this Python code is served by a threaded web framework (a Flask, Bottle or FastAPI route, a Django or Pyramid view), so no module is shared between request threads.
AX5 Architecture & structure — not assessed — architecture style/structure is computed from a project graph (projects, types, module namespaces) that was not loaded for this repository, because the repository is written in a language this check does not yet model, or because its projects failed to load. This is a gap in the analyzer, not a finding about this repository
AX6 Interface segregation — not assessed — interface segregation is computed over a type surface that was not loaded for this repository, because the repository is written in a language this check does not yet model, or because its projects failed to load. This is a gap in the analyzer, not a finding about this repository
AX7 Slice cohesion — not applicable — not a vertical-slice architecture
AX8 Test isolation — no test/production split to check
AX9 CQS / query purity — no CQRS query handlers detected — query purity is not applicable to this codebase
AXB2 Runtime readiness — Advisory — this card reports evidence and never carries a score, so there is nothing missing here.
AXR1 Runtime accessibility — the service did not build — `python3 -m venv /tmp/wd-venv && . /tmp/wd-venv/bin/activate && pip install -q -r REDACTED && pip install -q uvicorn` exited 1: ct-overflow', '-Wsign-compare', '-DNDEBUG', '-g', '-O2', '-Wall', '-fPIC', '-I./python', '-I./lib', '-I./deps/double-conversion/double-conversion', '-I/tmp/wd-venv/include', '-I/usr/include/python3.12…; runtime evidence skipped (launch chosen because scripts/serve_openai_api.py defines the FastAPI application 'app' at module level) This is a statement about this run, not a statement about your application: nothing here says the surface is inaccessible, only that it was never rendered.
C1 Data Protection — Not assessed: these personal data controls are read from a source model (declarative annotations, request middleware, entity/column names, guard methods) that was not loaded for this repository — because the repository is written in a language this check does not yet model, or because its projects failed to load. Absence of an idiom this check recognises is NOT evidence that this repository lacks personal data controls: it may implement them entirely in its own ecosystem. This is a gap in the analyzer's language coverage, not a finding about this repository.
C2 Access Controls — No access-control surface detected in the analyzed source — this repository registers no route and starts no server another machine can reach (no Flask/FastAPI/Starlette route, Django urlpatterns or class-based view, Tornado handler). Access control is therefore N/A here — a library or command-line tool is authorized by the application that serves it, not by itself. If this codebase grows request handlers, the dimension reactivates and a default-deny posture is expected then.
D10 Test Quality — No tests were found in the analyzed repository to assess for quality.
D11 Test Reliability — No test suite was found to re-run, so reliability couldn't be assessed. Two searches produced that zero and both came back empty: the classifier that reads the loaded workspace recognised no suite it could run, and a walk of the source on disk — which covers the JS/TS `*.test.*` and `*.spec.*` conventions and probes for a Pester suite — found no test source in any other ecosystem either. Neither search reaches a suite that is missing from the loaded workspace and carries no name either walk recognises, so this is 'no suite found by those two searches', not a verdict that the repository has none.
D16 Bus Factor — single-maintainer repository — bus factor is not applicable
D18 Solution Shape — D18 scores the shape of a .NET solution; this repository has no .NET solution or project files, so the dimension does not apply.
D19 Documentation Quality — No runnable or published unit to look for an example in, and no multi-project root README to look for a map in; the other documentation kinds are withheld (see the gaps), so documentation quality is not applicable here.
D20 ADR Quality — Repository kind not determined
D21 Naming Consistency — N/A — too few peers to establish a naming convention (571 declarations, no unit with a dominant convention).
D22 Internal API Consistency — No intentionally-exposed public API to evaluate for consistency.
D23 Boundary Type-Coupling — No bounded-context organisation was detected either — neither a context-shaped layout, nor 2+ sibling source directories each declaring an aggregate root, nor 2+ sibling projects each declaring an entity of their own or taking another project's integration events. Declaring this codebase's bounded contexts (≥2) would let cross-boundary type coupling be assessed. Declare them in `.codehealth/config.yaml` at the repository root (create it if absent), mapping each context name to the module-path or namespace prefixes that belong to it — e.g. `architecture:` → `contexts:` → `Billing: ["src/billing", "Acme.Billing"]`, `Catalog: ["src/catalog", "Acme.Catalog"]`.
D24 Comment Value — D24 reads C#, Visual Basic, TypeScript and JavaScript comments only — this repository's production source is .py, which was left unread. Not scored: this is a gap in the analyzer, not a verdict about this repository.
D25 ADR Conformance — no ADRs to check
D26 Project Cohesion — This repository is one build unit, `(repository root)` (4125 LoC, 42 module-visible types across 5 directories), within the size bounds. Not scored: a single unit of this size gives project cohesion nothing to grade.
D27 Navigability — symbol resolution incomplete — navigability not assessed
D31 IaC & Container Security — No Infrastructure-as-Code or container manifests found (Dockerfile, Docker Compose, Terraform, Kubernetes/Helm, CloudFormation, ARM, Bicep, Ansible); nothing to scan.
D32 Data Compliance (PII/GDPR) — No personal data was found crossing a boundary the PII/GDPR ruleset checks — nothing written to a log or console sink, placed in a URL or query string, or persisted to browser storage. That is a clean result for the LEAK surface only: this ruleset detects personal data escaping, it does not inventory the personal data a repository holds, so it is not evidence that this repository has no personal-data surface. The personal-data map (Appendix C) and the C1-C5 compliance cards are what speak to that.
D36 Supply-chain Provenance & Signing — No CI/build pipeline found (.github/.forgejo/.gitea workflows, .circleci, .buildkite, .woodpecker, .teamcity, .gitlab-ci.yml, .travis.yml, bitbucket-pipelines.yml, .drone.yml, .cirrus.yml, .woodpecker.yml, appveyor.yml, azure-pipelines*.yml, .pipelines/, .vsts-ci/, .azuredevops/, Jenkinsfile); there is no build to attest provenance for.
D37 Vulnerability-disclosure Policy — No vulnerability-disclosure policy file found (SECURITY.md/.markdown/.rst/.txt at root or under .github/.forgejo/.gitea/docs, .well-known/security.txt). A coordinated-disclosure policy may live off-repo, so this is not evidenced rather than failed.
D40 Network Egress Confinement — No Kubernetes/orchestration workloads found in the repository manifests; network egress policy is a cluster-native control that may live at the platform/firewall layer, so there is nothing to assess here.
D41 Kernel & Syscall Confinement — No Kubernetes/orchestration workloads found in the repository manifests; seccomp/AppArmor/SELinux confinement is a workload-level control, so there is nothing to assess here.
D42 Runtime Threat Enforcement — No Kubernetes/orchestration workloads found in the repository manifests; runtime threat-detection and admission-control policy are cluster-level controls, so there is nothing to assess here.
D7 Architectural Integrity — no checkable ADRs, and no project-reference graph for the cycle pass to read — so this dimension makes no claim about dependency cycles in either direction (where this repository's language has an import-cycle lens, cycles are reported there). Architectural integrity not assessed
DM1 Domain Modelling — not scored — this repository shows only 1 of the 3 signals this lens looks for (1 value object(s))
ED1 Event-Driven — not scored — this repository shows none of the 3 signals this lens looks for
ED5 Idempotency — no mutating command handlers or message consumers detected — idempotency check not applicable
ES1 Event Sourcing — not scored — this repository shows none of the 3 signals this lens looks for
GD1 Unfinished & placeholder code — no source files were read — this check reads C#, F#, JavaScript/TypeScript, Dart, Ruby and Scala, and none was read for this repository's product. That is a limit of the analyzer, not a finding about your code.
IC1 Incompleteness & stubs — not analysed — these correctness checks read a source model that was not loaded for this repository, because the repository is written in a language this check does not yet model, or because its projects failed to load. This is a gap in the analyzer, not a finding about this repository
P12 CI test-gate honesty — no CI workflow found
P7 Outbound HTTP resilience — not applicable — no HTTP server, API framework or worker entry point was found in the Python source, so there is no service whose uptime a failing dependency could take down
P8 Schema migrations — no ORM, schema-migration tool or schema auto-create was found in this repository's dependency manifests or source, so there is no database schema for this check to judge
P9 Domain vs controller coverage — coverage not measured: this check compares the coverage of the analyzer's own test run across the domain and web layers, and that run measured none here — Test coverage (D8) says why
PF1 Benchmark discipline — Not applicable: no benchmark suite was found. This check searched for a pytest-benchmark or pytest-codspeed dependency with tests taking the `benchmark` fixture, an asv.conf.json with its `time_`/`mem_`/`peakmem_` benchmarks, or a pyperf Runner, and for a `*benchmark*` script that this repository's CI runs. Benchmarks are credited as a bonus, so their absence is neither scored nor deducted.
PF2 Allocation hygiene — Not scored: this check counted allocation-aware idioms (Span, pools, stackalloc) per line, which describes what kind of code a repository is rather than whether it has an allocation problem, so it was withdrawn. Nothing here is a finding about your code.
S1 Web-Security Posture — Not assessed: these web-security controls are read from a source model (declarative annotations, request middleware, entity/column names, guard methods) that was not loaded for this repository — because the repository is written in a language this check does not yet model, or because its projects failed to load. Absence of an idiom this check recognises is NOT evidence that this repository lacks web-security controls: it may implement them entirely in its own ecosystem. This is a gap in the analyzer's language coverage, not a finding about this repository.
X1 Async correctness — not applicable — the code was read and holds no site of this kind, so there is nothing for this check to judge
X13 Undrained process stream — not applicable — the code was read and holds no site of this kind, so there is nothing for this check to judge
X14 Bypassable address classification — not applicable — the code was read and holds no site of this kind, so there is nothing for this check to judge
X15 Unvalidated length from an untrusted reader — not applicable — the code was read and holds no site of this kind, so there is nothing for this check to judge
X16 Unfloored truncation loop — not analysed — these correctness checks read a source model that was not loaded for this repository, because the repository is written in a language this check does not yet model, or because its projects failed to load. This is a gap in the analyzer, not a finding about this repository
X17 Uncapped recursion over a caller-supplied document — not applicable — the code was read and holds no site of this kind, so there is nothing for this check to judge
X18 Disposal-pattern correctness — not analysed — these correctness checks read a source model that was not loaded for this repository, because the repository is written in a language this check does not yet model, or because its projects failed to load. This is a gap in the analyzer, not a finding about this repository
X19 Unrestored process-global state — not applicable — the code was read and holds no site of this kind, so there is nothing for this check to judge
X2 Cancellation propagation — no C# was loaded; this repository's Python was read, and no request in it can be cancelled by a caller — a Python request takes no token: asyncio cancels the task and trio or anyio the scope that made it, from outside the call, and requests, httpx, aiohttp and gRPC take a timeout, not a cancellation
X20 Mistyped argument guard — This check asks whether an argument guard reports an EMPTY value as a NULL one, which needs a language that throws a null-specific argument exception — .NET's ArgumentNullException, the JVM's NullPointerException, Dart's ArgumentError.notNull. This repository contains none of those languages: the ones it is written in have a single exception for a bad argument, so there is no pair of exceptions to confuse and nothing here for it to assess. Not a gap in the analyzer and not a finding about your code.
X21 Side-effecting pattern guard — not applicable — the code was read and holds no site of this kind, so there is nothing for this check to judge
X22 Contradicted release guard — not applicable — the code was read and holds no site of this kind, so there is nothing for this check to judge
X23 Unguarded diagnostic materialisation — not applicable — the code was read and holds no site of this kind, so there is nothing for this check to judge
X24 Document value interpolated into markup unescaped — not applicable — the code was read and holds no site of this kind, so there is nothing for this check to judge
X26 Unsynchronised callback handoff — not applicable — the code was read and holds no site of this kind, so there is nothing for this check to judge
X27 Collection changed while being enumerated — not applicable — the code was read and holds no site of this kind, so there is nothing for this check to judge
X29 Per-element action decided by a fixed element — not analysed — these correctness checks read a source model that was not loaded for this repository, because the repository is written in a language this check does not yet model, or because its projects failed to load. This is a gap in the analyzer, not a finding about this repository
X30 Support guard that admits what it rejects — not applicable — the code was read and holds no site of this kind, so there is nothing for this check to judge
X32 Type resolved by simple name across every loaded assembly — This check is about how a .NET program searches the assemblies loaded into its process for a type, and this repository contains no .NET source, so there is nothing here for it to assess. Not a gap in the analyzer and not a finding about your code.
X4 Structured logging — not analysed — these correctness checks read a source model that was not loaded for this repository, because the repository is written in a language this check does not yet model, or because its projects failed to load. This is a gap in the analyzer, not a finding about this repository
X5 Nullable reference types — not analysed — these correctness checks read a source model that was not loaded for this repository, because the repository is written in a language this check does not yet model, or because its projects failed to load. This is a gap in the analyzer, not a finding about this repository
X6 Hand-rolled structured-format parsing — Advisory — this card reports evidence and never carries a score, so there is nothing missing here.
X7 Silent fallback defaults — Advisory — this card reports evidence and never carries a score, so there is nothing missing here.
X9 Subsumed condition operand — Advisory — this card reports evidence and never carries a score, so there is nothing missing here.
Appendix A — Findings (grouped)
The findings behind the scores, grouped by severity, then by dimension and kind. The high-severity issues are enumerated in full below; items per group are capped at 25 with any overflow stated explicitly per group, never silently truncated. The complete machine-readable list of every finding (all severities) is the companion findings.md in this report's bundle.
No automated tests — No automated tests — no test code was found in this repository. Untested code is the largest single risk to changing it safely. Start with the code you change most often: add a suite in a framework a runner can collect (pytest or unittest), and run it in CI so the gap cannot reopen.
Swallowed exception (caught, then discarded) trainer/rollout_engine.py:204— `except` takes every exception and records none of it — the body neither logs it, rethrows it, nor even names it, so the failure is discarded as completely as by an empty catch and only the substituted value survives. Log it through whatever this codebase already uses to report problems, narrow the catch to the exception this call can actually raise, or say in a comment on the catch why the failure genuinely cannot matter.
Swallowed exception (caught, then discarded) trainer/train_agent.py:80— `except` takes every exception and records none of it — the body neither logs it, rethrows it, nor even names it, so the failure is discarded as completely as by an empty catch and only the substituted value survives. Log it through whatever this codebase already uses to report problems, narrow the catch to the exception this call can actually raise, or say in a comment on the catch why the failure genuinely cannot matter.
Swallowed exception (caught, then discarded) trainer/train_agent.py:89— `except Exception` takes every exception and records none of it — the body neither logs it, rethrows it, nor even names it, so the failure is discarded as completely as by an empty catch and only the substituted value survives. Log it through whatever this codebase already uses to report problems, narrow the catch to the exception this call can actually raise, or say in a comment on the catch why the failure genuinely cannot matter.
Swallowed exception (caught, then discarded) trainer/train_agent.py:135— `except` takes every exception and records none of it — the body neither logs it, rethrows it, nor even names it, so the failure is discarded as completely as by an empty catch and only the substituted value survives. Log it through whatever this codebase already uses to report problems, narrow the catch to the exception this call can actually raise, or say in a comment on the catch why the failure genuinely cannot matter.
Swallowed exception (caught, then discarded) trainer/train_agent.py:229— `except` takes every exception and records none of it — the body neither logs it, rethrows it, nor even names it, so the failure is discarded as completely as by an empty catch and only the substituted value survives. Log it through whatever this codebase already uses to report problems, narrow the catch to the exception this call can actually raise, or say in a comment on the catch why the failure genuinely cannot matter.
Duplicated block (6 lines × 2) dataset/lm_dataset.py:75— dataset/lm_dataset.py:75-80 | dataset/lm_dataset.py:244-249 — both copies are in the same file, so extract the block into one function there and call it from each site — the copies drift apart the first time only one of them is edited.
Duplicated block (6 lines × 2) trainer/train_agent.py:131— trainer/train_agent.py:131-136 | trainer/train_agent.py:225-230 — both copies are in the same file, so extract the block into one function there and call it from each site — the copies drift apart the first time only one of them is edited.
Duplicated block (6 lines × 2) trainer/train_agent.py:312— trainer/train_agent.py:312-317 | trainer/train_grpo.py:119-124 — before extracting anything, compare `trainer/train_agent.py` and `trainer/train_grpo.py` as WHOLE FILES: this scan already matched 4 separate duplicated blocks between them, totalling at least 31 lines, which is the signature of one file having been copied from the other rather than of a helper waiting to be extracted. If that is what happened, the fix is to keep one copy and have the other call it (or delete it), which resolves this row and its siblings together — extracting one helper per block leaves the fork in place.
Duplicated block (6 lines × 2) scripts/web_demo.py:178— scripts/web_demo.py:178-183 | scripts/web_demo.py:192-197 — both copies are in the same file, so extract the block into one function there and call it from each site — the copies drift apart the first time only one of them is edited. Note that the copies do not run to the end of the range shown: their LAST lines are different code, not the same code under different names — the matched region ends inside that line. Extract the lines above it, and read the last line of each site separately.
Hotspot: trainer/train_agent.py trainer/train_agent.py:245— trainer/train_agent.py changed 10 times in last 90 days, and the most complex body those changes touched has cyclomatic complexity 26 in train_agent.rl_train_epoch at line 245. 5 of those changes were fix/bug commits, and the other 5 changed it for other reasons — this file is under both repair and feature pressure. Before the next change lands here, make sure the area it touches is under test, then split that area out of the file so the following change is smaller than this one — a file this often edited pays the complexity back every time. Counted over 2026-07-29..2026-10-27, the 90 days ending at the analysed commit. Reproduce with `git log --since='2026-07-29 14:57:05 +08:00' --until='2026-10-27 14:57:05 +08:00' --full-history --no-merges -- trainer/train_agent.py`: merges are excluded because a merge re-states changes already counted at their own commits, and history is NOT path-simplified because a change that reached the file through a merged branch is still a change to it. That command counts raw commits and can read HIGHER than this row, which counts a cherry-picked re-land, and a revert together with the commit it undoes, once each — a difference of several commits on a file whose history was re-landed or reverted inside the window.
Hotspot: trainer/train_ppo.py trainer/train_ppo.py:78— trainer/train_ppo.py changed 6 times in last 90 days, and the most complex body those changes touched has cyclomatic complexity 29 in train_ppo.ppo_train_epoch at line 78. 2 of those changes were fix/bug commits, and the other 4 changed it for other reasons — this file is under both repair and feature pressure. Before the next change lands here, make sure the area it touches is under test, then split that area out of the file so the following change is smaller than this one — a file this often edited pays the complexity back every time. Counted over 2026-07-29..2026-10-27, the 90 days ending at the analysed commit. Reproduce with `git log --since='2026-07-29 14:57:05 +08:00' --until='2026-10-27 14:57:05 +08:00' --full-history --no-merges -- trainer/train_ppo.py`: merges are excluded because a merge re-states changes already counted at their own commits, and history is NOT path-simplified because a change that reached the file through a merged branch is still a change to it. That command counts raw commits and can read HIGHER than this row, which counts a cherry-picked re-land, and a revert together with the commit it undoes, once each — a difference of several commits on a file whose history was re-landed or reverted inside the window.
Hotspot: trainer/train_grpo.py trainer/train_grpo.py:71— trainer/train_grpo.py changed 7 times in last 90 days, and the most complex body those changes touched has cyclomatic complexity 23 in train_grpo.grpo_train_epoch at line 71. 3 of those changes were fix/bug commits, and the other 4 changed it for other reasons — this file is under both repair and feature pressure. Before the next change lands here, make sure the area it touches is under test, then split that area out of the file so the following change is smaller than this one — a file this often edited pays the complexity back every time. Counted over 2026-07-29..2026-10-27, the 90 days ending at the analysed commit. Reproduce with `git log --since='2026-07-29 14:57:05 +08:00' --until='2026-10-27 14:57:05 +08:00' --full-history --no-merges -- trainer/train_grpo.py`: merges are excluded because a merge re-states changes already counted at their own commits, and history is NOT path-simplified because a change that reached the file through a merged branch is still a change to it. That command counts raw commits and can read HIGHER than this row, which counts a cherry-picked re-land, and a revert together with the commit it undoes, once each — a difference of several commits on a file whose history was re-landed or reverted inside the window.
Duplicated block (5 lines × 2) trainer/train_agent.py:332— trainer/train_agent.py:332-336 | trainer/train_grpo.py:144-150 — before extracting anything, compare `trainer/train_agent.py` and `trainer/train_grpo.py` as WHOLE FILES: this scan already matched 4 separate duplicated blocks between them, totalling at least 31 lines, which is the signature of one file having been copied from the other rather than of a helper waiting to be extracted. If that is what happened, the fix is to keep one copy and have the other call it (or delete it), which resolves this row and its siblings together — extracting one helper per block leaves the fork in place. Note that the copies do not run to the end of the range shown: their LAST lines are different code, not the same code under different names — the matched region ends inside that line. Extract the lines above it, and read the last line of each site separately.
Duplicated block (5 lines × 2) eval_llm.py:25— eval_llm.py:25-29 | scripts/serve_openai_api.py:41-45 — the copies span different directories, so extracting a shared function means choosing where it lives: put it somewhere both call sites can already reach — a location they all depend on today, or a new shared one if there is none — and call it from each site; until then, every change has to be made twice.
Duplicated block (5 lines × 2) scripts/convert_model.py:27— scripts/convert_model.py:27-31 | scripts/convert_model.py:86-91 — both copies are in the same file, so extract the block into one function there and call it from each site — the copies drift apart the first time only one of them is edited. ★ These copies have DRIFTED, and that is worth reading before extracting anything: just before the matched lines, `scripts/convert_model.py:26` calls `save_pretrained` and `scripts/convert_model.py:85` does not — after which the two agree again for 2 more lines. One of those two behaviours is the intended one and the other is what a copy-paste left behind, so decide which BEFORE unifying them: extracting the shared part will silently settle it, and if the copy that skips the call is the wrong one, that bug is already live.
Duplicated block (8 lines × 2) trainer/train_ppo.py:237— trainer/train_ppo.py:237-244 | trainer/train_ppo.py:247-254 — both copies are in the same file, so extract the block into one function there and call it from each site — the copies drift apart the first time only one of them is edited.
Duplicated block (8 lines × 2) trainer/train_distillation.py:122— trainer/train_distillation.py:122-129 | trainer/train_dpo.py:106-113 — before extracting anything, compare `trainer/train_distillation.py` and `trainer/train_dpo.py` as WHOLE FILES: this scan already matched 4 separate duplicated blocks between them, totalling at least 41 lines, which is the signature of one file having been copied from the other rather than of a helper waiting to be extracted. If that is what happened, the fix is to keep one copy and have the other call it (or delete it), which resolves this row and its siblings together — extracting one helper per block leaves the fork in place.
Duplicated block (7 lines × 2) trainer/train_grpo.py:106— trainer/train_grpo.py:106-112 | trainer/train_ppo.py:102-108 — the copies sit in sibling files of one directory, so a shared home is within easy reach: extract the block into a single shared function the call sites can all reach — a file they already depend on, or a new one alongside them — and call it from both call sites, so a change lands once.
Duplicated block (7 lines × 2) scripts/web_demo.py:307— scripts/web_demo.py:307-313 | trainer/trainer_utils.py:57-63 — the copies span different directories, so extracting a shared function means choosing where it lives: put it somewhere both call sites can already reach — a location they all depend on today, or a new shared one if there is none — and call it from each site; until then, every change has to be made twice.
Off the main sequence: dataset — dataset: abstractness 0.00, instability 0.00, distance 1.00 — zone of pain — concrete and depended on by 1 project(s), so it's rigid to change.
Off the main sequence: model — model: abstractness 0.00, instability 0.00, distance 1.00 — zone of pain — concrete and depended on by 3 project(s), so it's rigid to change.
train_ppo.ppo_train_epoch (cyclomatic 29) trainer/train_ppo.py:78— train_ppo.ppo_train_epoch has cyclomatic complexity 29 (threshold 15). To reduce it, separate the cases: extract each independent branch into its own named function, and where the body has guards that only reject input, fold those into early returns at the top.
train_agent.calculate_rewards (cyclomatic 26) trainer/train_agent.py:190— train_agent.calculate_rewards has cyclomatic complexity 26 (threshold 15). To reduce it, separate the cases: extract each independent branch into its own named function, and where the body has guards that only reject input, fold those into early returns at the top.
train_agent.rl_train_epoch (cyclomatic 26) trainer/train_agent.py:245— train_agent.rl_train_epoch has cyclomatic complexity 26 (threshold 15). To reduce it, separate the cases: extract each independent branch into its own named function, and where the body has guards that only reject input, fold those into early returns at the top.
train_grpo.grpo_train_epoch (cyclomatic 23) trainer/train_grpo.py:71— train_grpo.grpo_train_epoch has cyclomatic complexity 23 (threshold 15). To reduce it, separate the cases: extract each independent branch into its own named function, and where the body has guards that only reject input, fold those into early returns at the top.
train_distillation.train_epoch (cyclomatic 19) trainer/train_distillation.py:39— train_distillation.train_epoch has cyclomatic complexity 19 (threshold 15). To reduce it, separate the cases: extract each independent branch into its own named function, and where the body has guards that only reject input, fold those into early returns at the top.
MiniMindForCausalLM.generate (cyclomatic 18) model/model_minimind.py:262— MiniMindForCausalLM.generate has cyclomatic complexity 18 (threshold 15). To reduce it, separate the cases: extract each independent branch into its own named function, and where the body has guards that only reject input, fold those into early returns at the top.
Repeated repair: eval_llm.py eval_llm.py:33— eval_llm.py changed 7 times in last 90 days and 4 of those changes were fix/bug commits, so repair is the majority of this file's churn. Its max cyclomatic complexity is 8 (its worst body is eval_llm.main at line 33), AT OR UNDER the 15 threshold, so this is deliberately not filed as a churn × complexity hotspot — the difficulty here is in the behaviour the file has to get right, not in its control flow, and refactoring it for complexity would be the wrong move. The repairs counted were: “[fix] decide native-vs-transformers loading by directory name, not substring”; “fix: align LoRA inference paths with training output”; “[fix] unify Ctrl+C handling”; “fix: eval_llm.py Ctrl+C 退出异常修复”. Each one is a case this code did not handle. Before the next change lands here, check that every one of them is pinned by a test that fails without its fix; where the same area keeps coming back, the durable fix is usually at the interface that keeps being misused rather than at the line that was last corrected. Counted over 2026-07-29..2026-10-27, the 90 days ending at the analysed commit. Reproduce with `git log --since='2026-07-29 14:57:05 +08:00' --until='2026-10-27 14:57:05 +08:00' --full-history --no-merges -- eval_llm.py`: merges are excluded because a merge re-states changes already counted at their own commits, and history is NOT path-simplified because a change that reached the file through a merged branch is still a change to it. That command counts raw commits and can read HIGHER than this row, which counts a cherry-picked re-land, and a revert together with the commit it undoes, once each — a difference of several commits on a file whose history was re-landed or reverted inside the window.
train_ppo.ppo_train_epoch (cognitive 73) trainer/train_ppo.py:78— train_ppo.ppo_train_epoch has cognitive complexity 73 (threshold 15). Drivers by points: if/else 14 (42 pts), ternaries 5 (15 pts), loops 5 (11 pts), boolean chains 5 (nesting depth added 44). To reduce it, split the body into named stages: move each independent step or branch into its own named function so the body reads as a short sequence of named calls rather than one long body.
train_agent.calculate_rewards (cognitive 71) trainer/train_agent.py:190— train_agent.calculate_rewards has cognitive complexity 71 (threshold 15). Drivers by points: ternaries 13 (36 pts), if/else 8 (23 pts), loops 3 (6 pts), error handling 1 (5 pts), boolean chains 1 (nesting depth added 45). To reduce it, split the body into named stages: move each independent step or branch into its own named function so the body reads as a short sequence of named calls rather than one long body.
train_agent.rl_train_epoch (cognitive 52) trainer/train_agent.py:245— train_agent.rl_train_epoch has cognitive complexity 52 (threshold 15). Drivers by points: if/else 11 (25 pts), loops 4 (10 pts), ternaries 4 (10 pts), boolean chains 7 (nesting depth added 26). To reduce it, split the body into named stages: move each independent step or branch into its own named function so the body reads as a short sequence of named calls rather than one long body.
train_grpo.grpo_train_epoch (cognitive 44) trainer/train_grpo.py:71— train_grpo.grpo_train_epoch has cognitive complexity 44 (threshold 15). Drivers by points: if/else 10 (21 pts), loops 3 (8 pts), ternaries 3 (8 pts), boolean chains 7 (nesting depth added 21). To reduce it, split the body into named stages: move each independent step or branch into its own named function so the body reads as a short sequence of named calls rather than one long body.
trainer_utils.lm_checkpoint (cognitive 40) trainer/trainer_utils.py:65— trainer_utils.lm_checkpoint has cognitive complexity 40 (threshold 15). Drivers by points: if/else 10 (21 pts), ternaries 6 (17 pts), loops 1 (2 pts) (nesting depth added 23). To reduce it, flatten the nesting: this score is depth rather than breadth — most of its points come from checks stacked inside one another, so the work sits several levels in. Invert each enclosing check into an early exit (a return, or the language's equivalent) so the happy path stays at one level, and where a level cannot be exited early, lift the block it encloses into its own named function.
train_distillation.train_epoch (cognitive 38) trainer/train_distillation.py:39— train_distillation.train_epoch has cognitive complexity 38 (threshold 15). Drivers by points: if/else 10 (18 pts), ternaries 4 (13 pts), boolean chains 4, loops 2 (3 pts) (nesting depth added 18). To reduce it, split the body into named stages: move each independent step or branch into its own named function so the body reads as a short sequence of named calls rather than one long body.
MiniMindForCausalLM.generate (cognitive 31) model/model_minimind.py:262— MiniMindForCausalLM.generate has cognitive complexity 31 (threshold 15). Drivers by points: if/else 10 (18 pts), ternaries 5 (9 pts), loops 2 (4 pts) (nesting depth added 14). To reduce it, split the body into named stages: move each independent step or branch into its own named function so the body reads as a short sequence of named calls rather than one long body.
train_agent.rollout_single (cognitive 29) trainer/train_agent.py:93— train_agent.rollout_single has cognitive complexity 29 (threshold 15). Drivers by points: if/else 7 (16 pts), error handling 1 (4 pts), ternaries 2 (4 pts), loops 2 (3 pts), boolean chains 2 (nesting depth added 15). To reduce it, split the body into named stages: move each independent step or branch into its own named function so the body reads as a short sequence of named calls rather than one long body.
train_full_sft.train_epoch (cognitive 25) trainer/train_full_sft.py:24— train_full_sft.train_epoch has cognitive complexity 25 (threshold 15). Drivers by points: if/else 4 (9 pts), ternaries 3 (9 pts), boolean chains 4, loops 2 (3 pts) (nesting depth added 12). To reduce it, split the body into named stages: move each independent step or branch into its own named function so the body reads as a short sequence of named calls rather than one long body.
train_pretrain.train_epoch (cognitive 25) trainer/train_pretrain.py:24— train_pretrain.train_epoch has cognitive complexity 25 (threshold 15). Drivers by points: if/else 4 (9 pts), ternaries 3 (9 pts), boolean chains 4, loops 2 (3 pts) (nesting depth added 12). To reduce it, split the body into named stages: move each independent step or branch into its own named function so the body reads as a short sequence of named calls rather than one long body.
train_dpo.train_epoch (cognitive 22) trainer/train_dpo.py:53— train_dpo.train_epoch has cognitive complexity 22 (threshold 15). Drivers by points: if/else 4 (9 pts), ternaries 2 (6 pts), boolean chains 4, loops 2 (3 pts) (nesting depth added 10). To reduce it, split the body into named stages: move each independent step or branch into its own named function so the body reads as a short sequence of named calls rather than one long body.
train_lora.train_epoch (cognitive 22) trainer/train_lora.py:25— train_lora.train_epoch has cognitive complexity 22 (threshold 15). Drivers by points: if/else 4 (9 pts), ternaries 2 (6 pts), boolean chains 4, loops 2 (3 pts) (nesting depth added 10). To reduce it, split the body into named stages: move each independent step or branch into its own named function so the body reads as a short sequence of named calls rather than one long body.
train_tokenizer.get_texts (cognitive 18) trainer/train_tokenizer.py:13— train_tokenizer.get_texts has cognitive complexity 18 (threshold 15). Drivers by points: if/else 6 (11 pts), ternaries 1 (3 pts), error handling 1 (2 pts), boolean chains 1, loops 1 (nesting depth added 8). To reduce it, split the body into named stages: move each independent step or branch into its own named function so the body reads as a short sequence of named calls rather than one long body.
SFTDataset.generate_labels (cognitive 17) dataset/lm_dataset.py:91— SFTDataset.generate_labels has cognitive complexity 17 (threshold 15). Drivers by points: if/else 3 (7 pts), loops 3 (7 pts), ternaries 1 (3 pts) (nesting depth added 10). To reduce it, split the body into named stages: move each independent step or branch into its own named function so the body reads as a short sequence of named calls rather than one long body. This shape REPEATS in the file: one other method here (DPODataset.generate_loss_mask) has the same decision points, in the same order, at the same nesting depths — so this is one pattern written twice rather than two separate problems. Splitting this body alone leaves the other exactly as it is. Where these are variations on one operation, the change that clears both is the shared one: lift the common shape into a single routine the variants call, parameterised by whatever genuinely differs between them, and keep in each method only the part that is not shared.
DPODataset.generate_loss_mask (cognitive 17) dataset/lm_dataset.py:180— DPODataset.generate_loss_mask has cognitive complexity 17 (threshold 15). Drivers by points: if/else 3 (7 pts), loops 3 (7 pts), ternaries 1 (3 pts) (nesting depth added 10). To reduce it, split the body into named stages: move each independent step or branch into its own named function so the body reads as a short sequence of named calls rather than one long body. This shape REPEATS in the file: one other method here (SFTDataset.generate_labels) has the same decision points, in the same order, at the same nesting depths — so this is one pattern written twice rather than two separate problems. Splitting this body alone leaves the other exactly as it is. Where these are variations on one operation, the change that clears both is the shared one: lift the common shape into a single routine the variants call, parameterised by whatever genuinely differs between them, and keep in each method only the part that is not shared.
train_grpo.calculate_rewards (cognitive 17) trainer/train_grpo.py:37— train_grpo.calculate_rewards has cognitive complexity 17 (threshold 15). Drivers by points: ternaries 3 (11 pts), if/else 1 (3 pts), loops 2 (3 pts) (nesting depth added 11). To reduce it, split the body into named stages: move each independent step or branch into its own named function so the body reads as a short sequence of named calls rather than one long body.
Change coupling clique: train_agent.py, train_grpo.py, train_ppo.py trainer/train_agent.py— 3 files — `trainer/train_agent.py`, `trainer/train_grpo.py`, `trainer/train_ppo.py` — all change together with no explicit dependency: a fully-connected co-change clique, not 3 separate couplings. They share one concern (thin parallel siblings over a common abstraction), so extract the shared part into ONE unit and the whole clique's coupling clears at once — you do not need to break each pair individually.
D4 · Code Duplication· Near-duplicate member family (3 members, 32 shared lines) · ×1
Near-duplicate member family (3 members, 32 shared lines) trainer/train_full_sft.py:25— trainer/train_full_sft.py:25-72 | trainer/train_lora.py:26-67 | trainer/train_pretrain.py:25-71 — These 3 members are variants of one another: a block of 32 lines reported below appears in every one of them, and the pairwise near-duplicate rows they would otherwise produce are collapsed into this row. Read them as one construct written 3 times. The repair is at the members' grain — factor the shared implementation out once and have all of them call it with their differences as parameters or as an injected step, or, where the difference is systematic, generate them from one template. Extracting the individual blocks below is not the same fix: it leaves every body in place and the next edit still has to be made 3 times.
Near-duplicate member pair (29 shared lines) trainer/train_agent.py:246— trainer/train_agent.py:246-367 | trainer/train_grpo.py:72-196 — These two members are variants of one another: 29 of their lines are already reported as duplicated blocks below, spread through both bodies rather than gathered into one. Read them as a single construct written twice. The repair is at the members' grain — factor the shared pipeline into one implementation the two call with their differences as parameters or as an injected step, or, where the difference is systematic (sync against async, one transport against another), generate one from the other. Extracting the individual blocks below is not the same fix: it leaves the two bodies in place and the next edit still has to be made twice.
Near-duplicate member pair (22 shared lines) trainer/train_distillation.py:40— trainer/train_distillation.py:40-134 | trainer/train_dpo.py:54-119 — These two members are variants of one another: 22 of their lines are already reported as duplicated blocks below, spread through both bodies rather than gathered into one. Read them as a single construct written twice. The repair is at the members' grain — factor the shared pipeline into one implementation the two call with their differences as parameters or as an injected step, or, where the difference is systematic (sync against async, one transport against another), generate one from the other. Extracting the individual blocks below is not the same fix: it leaves the two bodies in place and the next edit still has to be made twice.
D4 · Code Duplication· Edited copy of a member (24 corresponding lines) · ×1
Edited copy of a member (24 corresponding lines) trainer/train_grpo.py:38— trainer/train_grpo.py:38-68 | trainer/train_ppo.py:52-75 — These two members are one piece of code written twice and then edited apart: 24 consecutive lines correspond almost exactly, broken only by small local edits. Most of that correspondence is NOT reported as duplicated blocks below — the edits cut it into fragments and only the largest of them clear the block floor, so the rows below understate it. The repair is at the members' grain — factor the shared implementation into one the two call with their differences as parameters or as an injected step, or, where the difference is systematic (an extra return value, one transport against another), generate one from the other. Left alone, the next edit has to be made twice and the two will drift further apart.
D4 · Code Duplication· Edited copy of a member (18 corresponding lines) · ×1
Edited copy of a member (18 corresponding lines) eval_llm.py:14— eval_llm.py:14-31 | scripts/serve_openai_api.py:29-47 — These two members are one piece of code written twice and then edited apart: 18 consecutive lines correspond almost exactly, broken only by small local edits. Most of that correspondence is NOT reported as duplicated blocks below — the edits cut it into fragments and only the largest of them clear the block floor, so the rows below understate it. The repair is at the members' grain — factor the shared implementation into one the two call with their differences as parameters or as an injected step, or, where the difference is systematic (an extra return value, one transport against another), generate one from the other. Left alone, the next edit has to be made twice and the two will drift further apart.
D4 · Code Duplication· Edited copy of a member (16 corresponding lines) · ×1
Edited copy of a member (16 corresponding lines) dataset/lm_dataset.py:92— dataset/lm_dataset.py:92-107 | dataset/lm_dataset.py:181-196 — These two members are one piece of code written twice and then edited apart: 16 consecutive lines correspond almost exactly, broken only by small local edits. Most of that correspondence is NOT reported as duplicated blocks below — the edits cut it into fragments and only the largest of them clear the block floor, so the rows below understate it. The repair is at the members' grain — factor the shared implementation into one the two call with their differences as parameters or as an injected step, or, where the difference is systematic (an extra return value, one transport against another), generate one from the other. Left alone, the next edit has to be made twice and the two will drift further apart.
D4 · Code Duplication· Edited copies of one member (3 members, 3+ corresponding lines) · ×1
Edited copies of one member (3 members, 3+ corresponding lines) trainer/train_agent.py:34— trainer/train_agent.py:34-36 | trainer/train_grpo.py:32-34 | trainer/train_ppo.py:30-32 — These 3 members are one piece of code written 3 times and then edited apart: every two of them correspond over at least 3 lines, broken only by small local edits. Most of that correspondence is NOT reported as duplicated blocks below: the edits cut it into fragments and only the largest of them clear the block floor, so the rows below understate it. The repair is at the members' grain. Factor the shared implementation out once and have all of them call it with their differences as parameters or as an injected step, or, where the difference is systematic, generate them from one template. Left alone, the next edit has to be made 3 times and the copies will drift further apart.
D4 · Code Duplication· Edited copy of a member (7 corresponding lines) · ×1
Edited copy of a member (7 corresponding lines) dataset/lm_dataset.py:127— dataset/lm_dataset.py:127-133 | dataset/lm_dataset.py:201-207 — These two members are one piece of code written twice and then edited apart: 7 consecutive lines correspond almost exactly, broken only by small local edits. Most of that correspondence is NOT reported as duplicated blocks below — the edits cut it into fragments and only the largest of them clear the block floor, so the rows below understate it. The repair is at the members' grain — factor the shared implementation into one the two call with their differences as parameters or as an injected step, or, where the difference is systematic (an extra return value, one transport against another), generate one from the other. Left alone, the next edit has to be made twice and the two will drift further apart.
Duplicated block (17 lines × 3) trainer/train_full_sft.py:25— trainer/train_full_sft.py:25-41 | trainer/train_lora.py:26-42 | trainer/train_pretrain.py:25-41 — before extracting anything, compare `trainer/train_full_sft.py` and `trainer/train_lora.py` as WHOLE FILES: this scan already matched 4 separate duplicated blocks between them, totalling at least 41 lines, which is the signature of one file having been copied from the other rather than of a helper waiting to be extracted. If that is what happened, the fix is to keep one copy and have the other call it (or delete it), which resolves this row and its siblings together — extracting one helper per block leaves the fork in place.
Duplicated block (12–14 lines × 4) trainer/train_distillation.py:93— trainer/train_distillation.py:93-104 | trainer/train_dpo.py:85-96 | trainer/train_full_sft.py:38-51 | trainer/train_pretrain.py:38-51 — before extracting anything, compare `trainer/train_distillation.py` and `trainer/train_dpo.py` as WHOLE FILES: this scan already matched 4 separate duplicated blocks between them, totalling at least 41 lines, which is the signature of one file having been copied from the other rather than of a helper waiting to be extracted. If that is what happened, the fix is to keep one copy and have the other call it (or delete it), which resolves this row and its siblings together — extracting one helper per block leaves the fork in place.
Duplicated block (2–11 lines × 5) trainer/train_distillation.py:107— trainer/train_distillation.py:107-108 | trainer/train_dpo.py:99-100 | trainer/train_full_sft.py:44-54 | trainer/train_lora.py:44-53 | trainer/train_pretrain.py:44-54 — before extracting anything, compare `trainer/train_distillation.py` and `trainer/train_dpo.py` as WHOLE FILES: this scan already matched 4 separate duplicated blocks between them, totalling at least 41 lines, which is the signature of one file having been copied from the other rather than of a helper waiting to be extracted. If that is what happened, the fix is to keep one copy and have the other call it (or delete it), which resolves this row and its siblings together — extracting one helper per block leaves the fork in place. Note that the copies do not run to the end of the range shown: their LAST lines are different code, not the same code under different names — the matched region ends inside that line. Extract the lines above it, and read the last line of each site separately.
Duplicated block (19 lines × 2) trainer/train_grpo.py:50— trainer/train_grpo.py:50-68 | trainer/train_ppo.py:57-75 — the copies sit in sibling files of one directory, so a shared home is within easy reach: extract the block into a single shared function the call sites can all reach — a file they already depend on, or a new one alongside them — and call it from both call sites, so a change lands once. The `return` at the foot of the matched lines is the enclosing body's own terminal exit, not an early one: it moves with them unchanged, and each site calls the extracted unit from the position that `return` occupied — no decision has to be handed back and re-acted on.
Duplicated block (10 lines × 2) dataset/lm_dataset.py:93— dataset/lm_dataset.py:93-102 | dataset/lm_dataset.py:182-191 — both copies are in the same file, so extract the block into one function there and call it from each site — the copies drift apart the first time only one of them is edited. The matched lines also transfer control out of the body holding them, which cannot survive a move into a called unit unchanged: have the extracted unit return that decision and let each site act on it.
Duplicated block (7 lines × 5) trainer/train_distillation.py:98— trainer/train_distillation.py:98-104 | trainer/train_dpo.py:90-96 | trainer/train_full_sft.py:44-51 | trainer/train_lora.py:44-50 | trainer/train_pretrain.py:44-51 — before extracting anything, compare `trainer/train_distillation.py` and `trainer/train_dpo.py` as WHOLE FILES: this scan already matched 4 separate duplicated blocks between them, totalling at least 41 lines, which is the signature of one file having been copied from the other rather than of a helper waiting to be extracted. If that is what happened, the fix is to keep one copy and have the other call it (or delete it), which resolves this row and its siblings together — extracting one helper per block leaves the fork in place.
Duplicated block (11 lines × 2) trainer/train_agent.py:319— trainer/train_agent.py:319-329 | trainer/train_grpo.py:132-142 — before extracting anything, compare `trainer/train_agent.py` and `trainer/train_grpo.py` as WHOLE FILES: this scan already matched 4 separate duplicated blocks between them, totalling at least 31 lines, which is the signature of one file having been copied from the other rather than of a helper waiting to be extracted. If that is what happened, the fix is to keep one copy and have the other call it (or delete it), which resolves this row and its siblings together — extracting one helper per block leaves the fork in place.
Duplicated block (5–7 lines × 2) trainer/train_agent.py:360— trainer/train_agent.py:360-366 | trainer/train_grpo.py:189-193 — before extracting anything, compare `trainer/train_agent.py` and `trainer/train_grpo.py` as WHOLE FILES: this scan already matched 4 separate duplicated blocks between them, totalling at least 31 lines, which is the signature of one file having been copied from the other rather than of a helper waiting to be extracted. If that is what happened, the fix is to keep one copy and have the other call it (or delete it), which resolves this row and its siblings together — extracting one helper per block leaves the fork in place. Read the line range as the matched WINDOW rather than a finished unit: at `trainer/train_agent.py:360` it begins part-way through the construct above it, so those exact lines cannot be lifted as they stand — widen the region to the smallest complete statement or declaration that contains it, and extract that. Note that the copies do not run to the end of the range shown: their LAST lines are different code, not the same code under different names — the matched region ends inside that line. Extract the lines above it, and read the last line of each site separately.
Duplicated block (5 lines × 3) trainer/train_full_sft.py:57— trainer/train_full_sft.py:57-61 | trainer/train_lora.py:56-60 | trainer/train_pretrain.py:57-61 — before extracting anything, compare `trainer/train_full_sft.py` and `trainer/train_lora.py` as WHOLE FILES: this scan already matched 4 separate duplicated blocks between them, totalling at least 41 lines, which is the signature of one file having been copied from the other rather than of a helper waiting to be extracted. If that is what happened, the fix is to keep one copy and have the other call it (or delete it), which resolves this row and its siblings together — extracting one helper per block leaves the fork in place.
No structured logging — No logging or diagnostics emission found. This repository ships a binary its users launch rather than a service you operate, so there is no operator to page — but when a run fails on a user's machine, neither they nor you have anything to read.
Duplicated predicate trainer/train_agent.py:295— `args.debug_mode and is_main_process() and step % args.debug_interval == 0` appears character-identically in 3 files — trainer/train_agent.py, trainer/train_grpo.py, trainer/train_ppo.py. It is one line, so the duplication detector's token window never sees it; the copies drift when only one is corrected. Give the condition a name and one home.
Duplicated predicate trainer/rollout_engine.py:177— `not dist.is_initialized() or dist.get_rank() == 0` appears character-identically in 2 files — trainer/rollout_engine.py, trainer/trainer_utils.py. It is one line, so the duplication detector's token window never sees it; the copies drift when only one is corrected. Give the condition a name and one home.
Duplicated predicate trainer/train_agent.py:335— `step % args.accumulation_steps == 0 or step == iters` appears character-identically in 7 files — trainer/train_agent.py, trainer/train_distillation.py, trainer/train_dpo.py, trainer/train_full_sft.py. It is one line, so the duplication detector's token window never sees it; the copies drift when only one is corrected. Give the condition a name and one home.
Duplicated predicate trainer/train_agent.py:339— `step % args.log_interval == 0 or step == iters` appears character-identically in 7 files — trainer/train_agent.py, trainer/train_distillation.py, trainer/train_dpo.py, trainer/train_full_sft.py. It is one line, so the duplication detector's token window never sees it; the copies drift when only one is corrected. Give the condition a name and one home.
Duplicated predicate trainer/train_agent.py:351— `step % args.save_interval == 0 or step == iters` appears character-identically in 8 files — trainer/train_agent.py, trainer/train_distillation.py, trainer/train_dpo.py, trainer/train_full_sft.py. It is one line, so the duplication detector's token window never sees it; the copies drift when only one is corrected. Give the condition a name and one home.
Duplicated predicate trainer/train_agent.py:475— `train_sampler and train_sampler.set_epoch(epoch)` appears character-identically in 8 files — trainer/train_agent.py, trainer/train_distillation.py, trainer/train_dpo.py, trainer/train_full_sft.py. It is one line, so the duplication detector's token window never sees it; the copies drift when only one is corrected. Give the condition a name and one home.
No ADRs — No Architecture Decision Records found — no conventional ADR directory, no numbered `NNNN-title` documents in any markup this check reads, and nothing ADR-shaped by content. Design rationale recorded elsewhere (a design-notes tree, a mailing list, pull-request discussion) is not visible to this check and is not re-findable per decision, so a future maintainer cannot ask why one choice was made and get an answer.
No src/ separation — Most of the production code is in directories of its own, but 1 application file(s) sit at the repository root beside the build and tooling files (`eval_llm.py`).
No SAST — No static application security testing detected. For this repository's stack, add bandit, `semgrep --config=p/python`, or CodeQL's python pack — this repository has no CI pipeline yet, so run it locally to clear the existing findings, then make it a step of the first workflow you add so a regression fails the build. What was searched, so you can tell an absence from a miss: the 0 CI workflow file(s) in this repository, and the scanner and linter configuration checked in beside them. A scan that runs outside CI, one configured in your forge's web UI rather than in a committed file, or a tool whose name is none of those this check carries, is not seen — if that is your case the row is wrong, and saying so is more useful than adding a second scanner.
Outdated: datasets — `datasets` is pinned to 3.6.0, but 5.2.0 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 3.6.0 until the declaration is edited. This crosses a major version, so it is a migration rather than a bump.
Outdated: datasketch — `datasketch` is pinned to 1.6.4, but 2.0.0 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 1.6.4 until the declaration is edited. This crosses a major version, so it is a migration rather than a bump.
Outdated: einops — `einops` is pinned to 0.8.1, but 0.8.2 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 0.8.1 until the declaration is edited.
Outdated: fastapi — `fastapi` is pinned to 0.120.1, but 0.162.7 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 0.120.1 until the declaration is edited.
Outdated: flask — `flask` is pinned to 3.0.3, but 3.1.3 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 3.0.3 until the declaration is edited.
Outdated: flask-cors — `flask-cors` is pinned to 4.0.0, but 6.0.5 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 4.0.0 until the declaration is edited. This crosses a major version, so it is a migration rather than a bump.
Outdated: jinja2 — `jinja2` is pinned to 3.1.2, but 3.1.6 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 3.1.2 until the declaration is edited.
Outdated: marshmallow — `marshmallow` is pinned to 3.22.0, but 4.4.0 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 3.22.0 until the declaration is edited. This crosses a major version, so it is a migration rather than a bump.
Outdated: modelscope — `modelscope` is pinned to 1.37.0, but 1.40.3 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 1.37.0 until the declaration is edited.
Outdated: ngrok — `ngrok` is pinned to 1.4.0, but 1.7.0 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 1.4.0 until the declaration is edited.
Outdated: nltk — `nltk` is pinned to 3.8, but 3.10.4 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 3.8 until the declaration is edited.
Outdated: numpy — `numpy` is pinned to 1.26.4, but 2.5.4 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 1.26.4 until the declaration is edited. This crosses a major version, so it is a migration rather than a bump.
Outdated: openai — `openai` is pinned to 1.59.6, but 3.51.0 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 1.59.6 until the declaration is edited. This crosses a major version, so it is a migration rather than a bump.
Outdated: protobuf — `protobuf` is pinned to 6.33.5, but 7.36.2 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 6.33.5 until the declaration is edited. This crosses a major version, so it is a migration rather than a bump.
Outdated: psutil — `psutil` is pinned to 5.9.8, but 7.2.2 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 5.9.8 until the declaration is edited. This crosses a major version, so it is a migration rather than a bump.
Outdated: pydantic — `pydantic` is pinned to 2.11.5, but 2.14.3 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 2.11.5 until the declaration is edited.
Outdated: rich — `rich` is pinned to 13.7.1, but 15.0.0 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 13.7.1 until the declaration is edited. This crosses a major version, so it is a migration rather than a bump.
Outdated: scikit-learn — `scikit-learn` is pinned to 1.5.1, but 1.9.1 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 1.5.1 until the declaration is edited.
Outdated: sentence-transformers — `sentence-transformers` is pinned to 2.3.1, but 6.3.0 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 2.3.1 until the declaration is edited. This crosses a major version, so it is a migration rather than a bump.
Outdated: sentencepiece — `sentencepiece` is pinned to 0.2.0, but 0.2.3 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 0.2.0 until the declaration is edited.
Outdated: streamlit — `streamlit` is pinned to 1.50.0, but 1.67.0 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 1.50.0 until the declaration is edited.
Outdated: swanlab — `swanlab` is pinned to 0.9.8, but 0.10.5 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 0.9.8 until the declaration is edited.
Outdated: tiktoken — `tiktoken` is pinned to 0.10.0, but 0.14.0 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 0.10.0 until the declaration is edited.
Outdated: transformers — `transformers` is pinned to 4.57.6, but 5.23.0 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 4.57.6 until the declaration is edited. This crosses a major version, so it is a migration rather than a bump.
Outdated: trl — `trl` is pinned to 0.13.0, but 1.18.0 is the current stable release on PyPI. An exact pin never moves on its own, so this repository installs 0.13.0 until the declaration is edited. This crosses a major version, so it is a migration rather than a bump.
+ 3 more in this group — see findings.md.
Appendix B — Reproduction & audit trail
Every external tool invocation behind a deep-scan dimension — the tool, its captured version, the exact command, how many findings it yielded, and a link to the retained raw output. To reproduce any finding: check out the same commit and run the command shown (repo-relative — never an absolute scratch path). The complete raw scanner output is retained verbatim under artifacts/raw/ (indexed in artifacts/raw/index.json); per-invocation exit codes and wall-clock durations are in sidecar.json — kept out of this table so the rendered report stays byte-identical across runs of the same commit.
trivy: not applicable — No Infrastructure-as-Code or container manifests found (Dockerfile, Docker Compose, Terraform, Kubernetes/Helm, CloudFormation, ARM, Bicep, Ansible); nothing to scan.
semgrep: not applicable — No personal data was found crossing a boundary the PII/GDPR ruleset checks — nothing written to a log or console sink, placed in a URL or query string, or persisted to browser storage. That is a clean result for the LEAK surface only: this ruleset detects personal data escaping, it does not inventory the personal data a repository holds, so it is not evidence that this repository has no personal-data surface. The personal-data map (Appendix C) and the C1-C5 compliance cards are what speak to that.
provenance: not applicable — No CI/build pipeline found (.github/.forgejo/.gitea workflows, .circleci, .buildkite, .woodpecker, .teamcity, .gitlab-ci.yml, .travis.yml, bitbucket-pipelines.yml, .drone.yml, .cirrus.yml, .woodpecker.yml, appveyor.yml, azure-pipelines*.yml, .pipelines/, .vsts-ci/, .azuredevops/, Jenkinsfile); there is no build to attest provenance for.
disclosure: not applicable — No vulnerability-disclosure policy file found (SECURITY.md/.markdown/.rst/.txt at root or under .github/.forgejo/.gitea/docs, .well-known/security.txt). A coordinated-disclosure policy may live off-repo, so this is not evidenced rather than failed.
runtime-hardening: not applicable — No Kubernetes/orchestration workloads found in the repository manifests; network egress policy is a cluster-native control that may live at the platform/firewall layer, so there is nothing to assess here.
runtime-hardening: not applicable — No Kubernetes/orchestration workloads found in the repository manifests; seccomp/AppArmor/SELinux confinement is a workload-level control, so there is nothing to assess here.
runtime-hardening: not applicable — No Kubernetes/orchestration workloads found in the repository manifests; runtime threat-detection and admission-control policy are cluster-level controls, so there is nothing to assess here.
Run 01a1d781-59e2-7c8c-ab39-3f7f3eef5e55 · every finding is also locatable in findings.md, and the complete scoring record (with exit codes + durations) in sidecar.json.
Issues: 7 · Warnings: 69 · Recommendations: 9 · Info: 28 — Appendix A · all findings · full markdown report.
Generated by the Canine Code Analyzer: deterministic code-health analysis. 14-11-2026 @ 00:42 UTC.
Downloadable artifacts
Machine-readable and reproducible from this commit + frozen rubric — drop them straight into a contract appendix, a CRA dossier, or a downstream SCA / VEX tool.