Reports

Bayesilisk emits JSON and Markdown. JSON is the stable integration format. Markdown is intended for humans reviewing findings.

Main Fields

A full report includes:

• seed and tool version;

• effective configuration for attention, embeddings, model proposals, and safe base URL class;

• production-access boundary;

• scenario id and title;

• fragments and generated sub-scenarios;

• access pattern and domains;

• invariant id, layer, and expectation;

• observed result and observation detail;

• fingerprint and dedupe key;

• issue readiness;

• attention score and reasons;

• posterior probability and risk score;

• suggested issue title and body.

• safe model proposal provenance for accepted model-proposed scenarios.

Contextual reports also expose four first-class ledgers:

• observedByPlaywright: concrete browser evidence supplied by Playwright;

• selectedByGrassmannAttention: invariant planes selected for exploration;

• proposedByModel: untrusted scenario proposal activity;

• verifiedByBayesilisk: deterministic invariant results that issue payloads may use.

Finding Status

observedResult is deterministic:

• fail: the invariant did not hold for the scenario facts;

• pass: the invariant held;

• probe: the scenario is not confirmed enough to treat as a failure.

issueReadiness controls automation:

• ready-for-issue: safe to create an issue if it is not a duplicate;

• probe-only: needs a local verifier or human confirmation first;

• regression-watch: previously fixed or watched finding;

• do-not-open-muted: intentionally muted;

• no-issue-control: passing or control scenario.

Issue automation should require both:

observedResult = fail
issueReadiness = ready-for-issue

Scores

riskScore is a Bayesian-style priority after deterministic verification. It answers how important a verified result appears.

attentionScore is an exploration score. It answers where Bayesilisk should look next.

Keeping these separate prevents embeddings or model output from becoming a hidden bug oracle.

Muted, fixed, or regression-watch invariant planes can apply decayForFixedOrMuted to attention. That lowers future exploration priority but does not hide deterministic failures or alter riskScore.

Fingerprints

Bayesilisk fingerprints are stable identifiers derived from verified scenario and invariant data. Use them to dedupe issues, mute noisy probes, and mark fixed regressions in observation history.

Issue Payloads

The --issue-payloads mode emits deduped issue payloads for ready failed findings:

python3 -m bayesilisk --seed 150 --context /tmp/context.json --issue-payloads

The CLI does not mutate issue trackers. It returns payloads that another trusted workflow can review and submit.

When a finding came from a model-proposed scenario, issue payloads include the same safe modelProvenance object as the report. This provenance is intended for auditability only; it does not make the model output trusted.

Proof Artifacts

Downloadable examples:

• Example JSON report

• Example GitHub issue payloads

The README includes a proof-loop diagram, and the same boundary is represented as a text flow here:

Playwright evidence
  -> Grassmann attention
  -> candidate scenario from catalog, rules, or model
  -> Bayesilisk deterministic verification
  -> ready issue payload or reject/watchlist

Why This Is Not a Black Box

Bayesilisk reports keep evidence, routing, proposal, and verification separate:

• observedByPlaywright is browser evidence;

• selectedByGrassmannAttention is routing telemetry;

• proposedByModel is untrusted candidate scenario activity;

• verifiedByBayesilisk is the deterministic verifier ledger.

The issue-worthy result must come from verifiedByBayesilisk.

Model Unavailable? Still Works

The default report path does not require Ollama, hosted model APIs, API keys, or browser services. A fixed seed with the same local inputs still produces the same report, schema validation, fingerprints, risk scores, and issue payloads.