Qedgen Auditor Bench

QEDGen Auditor Benchmark

Measure the portable auditor skill without leaking labeled findings into audit workers. Require a venue-owned benchmark directory; never assume a particular home directory or agent harness.

Inputs

Require:

• canonical auditor skill directory;
• a corpus manifest conforming to schemas/corpus-manifest.schema.json, with repository, audited commit, program root, runtime, setup/test commands, sanitization rules, labeled findings, optional domain expectations, and a required entry-level difficulty;
• explicit benchmark output directory;
• audit profile and run count;
• exact discovery and judge model identifiers when the venue permits selection.

Default to one corpus entry and two independent audit runs for a regression check. Use three only for a high-assurance benchmark. Never use an unbounded run-until-dry loop.

For a local domain-regression baseline, use the seven labeled, schema-valid fixtures under fixtures/domain-corpus/ and run fixtures/domain-corpus/validate.sh before scoring. These fixtures exercise candidate categories, units, paired operations, and intended language gaps without third-party ground truth.

Model policy

The auditor's references/model-selection.md is the single source for concrete model identifiers; read it before dispatch and take the current per-provider mapping from there. Do not restate identifiers here — a model bump must be a one-file change.

Per capability tier:

• discovery worker: the mapping's default audit worker at high reasoning; high-assurance runs use its highest practical reasoning setting;
• reconciliation judge: the mapping's balanced judge tier with structured JSON output;
• deterministic extraction/normalization: the mapping's low-cost structured tier.

Pin the exact identifiers taken from the mapping in scored runs. Do not use aliases. Preserve older models only in an intentionally labeled model-regression comparison. OpenAI workers can stop mid-run when audit content triggers the provider's cybersecurity policy; classify such runs policy-interfered, never as a clean pass.

For providers the mapping does not cover, select the current frontier reasoning/coding tier for discovery and the balanced frontier tier for judging. Record the mapping used.

Per-entry pipeline

1. Establish a live baseline

• Check out exactly the labeled commit into an isolated worktree.
• Verify remediation commits are not ancestors of the baseline.
• Spot-check at least one labeled vulnerable location.
• Exclude labels that are fixed or not present at that commit.
• Record the baseline verification evidence.

Never score recall against a tree where the expected vulnerability is absent.

2. Sanitize independently of the worker

Remove audit reports, findings, .qed/, security reviews, remediation notes, and firm-specific artifacts. Scan remaining documentation for finding IDs, severity headings, and descriptions. Record every removal in sanitize.log.

The audit worker may read only:

• its sanitized source tree;
• the canonical auditor skill and routed references;
• explicitly supplied QEDGen/test binaries.

It must not read the corpus manifest, labels, reports, prior benchmark results, or sibling runs.

3. Run deterministic preflight

Invoke the auditor's scripts/preflight.sh against the selected program root. Record target, runtime, spec, tool versions, compilation status, and available reproducer lanes. A build failure is a benchmark outcome, not a recall score.

4. Dispatch clean audit workers

Each worker receives the same sanitized snapshot and this task only:

Audit the selected Solana program using the supplied qedgen-auditor skill.
Derive findings only from the sanitized source. Follow its evidence and
reproducer contracts. Write the full report and return its digest.

Run workers in isolated contexts. Independent workers are a venue capability; when unavailable, stop and report that an unbiased model benchmark cannot run. Do not substitute a ground-truth-contaminated coordinator session.

5. Re-run evidence

Independently execute every reproducer reported as fired. Downgrade any result that does not fire cleanly. Preserve source-established structural findings but do not count hypotheses as vulnerabilities.

6. Normalize and reconcile

Normalize reports to schemas/normalized-report.schema.json:

{
  "id": "string",
  "severity": "critical|high|medium|low|info",
  "evidence": "confirmed|structural|hypothesis|rejected",
  "category": "string",
  "location": "path:line",
  "root_cause": "string",
  "title": "string",
  "repro_status": "fired|inconclusive|silent|not-required"
}

repro_status uses the auditor's reproducer-contract tokens: fired = the assertion fired (confirmed); silent = the repro ran but did not fire (rejected); inconclusive = build/simulator limitation or ambiguous result; not-required = MEDIUM/LOW structural finding. info severity exists only to normalize third-party report entries — the auditor never emits it, and info rows are excluded from severity-agreement scoring.

Generate deterministic match candidates using category, path, handler, and root cause. Give the judge only the two normalized reports. The judge may match or reject pairs but must not invent or re-grade findings.

Deduplicate the union by root cause and affected operation, not just line number. Preserve which worker found each item.

7. Score

Every corpus entry has exactly one difficulty: smoke, standard, hard, or adversarial. Report recall and precision separately for every represented difficulty and include that tier's entry count. MUST NOT collapse mixed difficulty tiers into one headline score. An optional aggregate may accompany the per-tier results only when the tier_entry_counts that produced it are printed beside it.

Emit per-run and union metrics:

• recall by firm severity and overall;
• precision over confirmed and structural vulnerability findings;
• hypothesis rate;
• severity agreement on matched findings;
• reproducer fire rate for HIGH/CRITICAL candidates;
• unique-real and unique-noise counts;
• per-run variance and union gain;
• build/preflight success;
• time to first confirmed MED+;
• total wall time and, when exposed, token/cost usage;
• exact skill version, commit, model identifiers, reasoning settings, and QEDGen version.

When the corpus entry includes domain expectations, also emit:

• domain-invariant candidate recall before user ratification;
• false-ratification rate (unsupported candidates promoted to executable spec);
• unit/scale identification accuracy;
• structural, domain, and regression-layer spec completeness;
• Crucible entry-point selection and whether domain mode re-ran after ratification;
• dry-fuzz overclaim count (claims inferred from an empty lane outside that lane's observable coverage).

Run a probe-failure variant for local or sanitized fixtures by disabling the ordinary probe while leaving source readable — set QEDGEN_BIN to a stub executable that exits nonzero so preflight reports QEDGen as missing. Require a completed domain dossier, explicit blocked-lane status, no clean-audit claim, and resumable verification commands. Score this separately from vulnerability recall so an intentional tooling fault is not counted as a missed finding.

Do not count fixed ground truth as a miss. Do not count hypotheses as true positives. Report structural precision separately from confirmed precision. Write the machine-readable result against schemas/score.schema.json.

Then validate this run's own artifacts. Put corpus-manifest.json, normalized-report.json, and score.json in the benchmark output directory and run:

skills/qedgen-auditor-bench/schemas/validate.sh <benchmark-output-dir>

A run whose artifacts do not validate is not a scored run. Reporting numbers from it is the same error as reporting a collapsed mixed-tier headline. The script checks structure against the schemas, that every referenced corpus entry exists, and that tier_entry_counts matches the entries actually scored.

8. Detect invalid runs

Assign a run-level status before computing recall:

• scored: the worker completed the authorized audit without external interference.
• build-blocked: source review completed, but executable evidence was blocked; score structural recall only and exclude reproducer metrics.
• tooling-blocked: source review and dossier extraction completed, but QEDGen, runtime metadata, or one or more verification lanes were unavailable. Score source-derived candidate recall, require a valid run manifest with resume commands, and exclude unavailable-lane metrics.
• policy-interfered: a venue safety or cybersecurity policy refused, truncated, redirected, or suppressed the authorized analysis. Exclude the run from recall, precision, severity, and model comparisons.
• tooling-error: preflight, source access, or required venue capability failed before meaningful analysis. Exclude the run from security metrics.
• contaminated: the worker could access labels, reports, prior results, or coordinator context. Exclude the run entirely.

Policy interference is not a finding miss. Preserve the refusal or moderation signal in the run metadata so benchmark dashboards distinguish model behavior from venue-policy behavior. When real third-party targets trigger policy interference, move capability evaluation to intentionally vulnerable local fixtures or another explicitly authorized test target rather than weakening the venue's safeguards.

Comparison modes

• Skill regression: same model, corpus, commit, prompts, and run count; change only the skill version.
• Model regression: same skill and corpus; change only the model identifier and reasoning setting.
• End-to-end release score: current default skill and model; compare against the prior release but label both changed dimensions.

Avoid changing model and skill simultaneously when attributing improvement. Skill-regression and model-regression comparisons must assert identical per-tier entry counts. If their tier composition differs, set comparison.composition_valid to false and label the comparison invalid; never interpret the score delta as a model or skill regression.

Output

Write under the explicit benchmark directory:

<output>/<entry>/<run-id>/
  baseline.json
  sanitize.log
  run-1/domain-dossier.json
  run-1/run-manifest.json
  run-1/report.md
  run-1/repro.log
  run-N/...
  matches.json
  score.json
  summary.md

Never publish or commit third-party vulnerability details automatically. Maintain responsible-disclosure handling for unmatched HIGH/CRITICAL findings.