@gobing-ai/spur 0.4.4 → 0.4.5

package/config/rules/docs/verdict-confidence.yaml

@@ -1,68 +0,0 @@
-$schema: "@gobing-ai/spur/schemas/rule-file.schema.json"
-# Verdict-confidence rules (report corpus) — a committed verification record that states
-# a machine `Verdict:` line must also state the verifier's confidence, so a reader can
-# tell a PASS the verifier stands behind from one it would not.
-#
-# Contract owners: `Confidence: HIGH|MEDIUM|LOW` is specified for verify answers by
-# plugins/sp/skills/code-verification/SKILL.md (one `Verdict: PASS|PARTIAL|FAIL` line, one
-# `Confidence:` line, then the traceability tables) and for review reports by
-# plugins/sp/agents/super-reviewer.md. Both are already enforced for pipeline answers —
-# packages/app/src/services/verify-answer-lint.ts (`confidence-missing`,
-# `confidence-value`, and the coherence pair `confidence-unwarranted` /
-# `confidence-understated`) plus the verify stamp in config/workflows/task-pipeline.yaml —
-# but no rule read the standalone report documents agents commit under `docs/`. This file
-# is that backstop: the same three-level vocabulary, checked at the document level.
-#
-# Shape: one file-spanning forbidden match (`multiline: true` hands the evaluator the whole
-# document and the negative lookahead spans the newline) instead of line counting. The
-# lookahead demands a valid level, so a hedge (`Confidence: mostly`) reads as missing too.
-# The verdict prefix requires the closed `PASS|PARTIAL|FAIL` value on its own line, which is
-# what separates a record from prose: docs/reports/i31/0912-workflow-baseline.md:84 and
-# i31/1107-runall-p1-pilot.md:9 both open a bolded *conclusion* sentence
-# ("**Verdict: `partial — INSUFFICIENT_EVIDENCE retained`.**") and are not this contract.
-#
-# Severity `info`, and deliberately not in a preset — the same calibration choice the `docs`
-# category makes for sourced claims: findings surface for review without reddening gates
-# (`recommended-pre-check` runs `--fail-on warning`). To promote it, add `- docs` to that
-# preset's `extends` once the residual is triaged.
-#
-# Calibration (2026-10-13, `spur rule run --file config/rules/docs/verdict-confidence.yaml
-# --json`, unsuppressed): 3 findings over the 91 documents the include globs matched at that
-# timestamp (29 under docs/reports/, 62 under docs/plans/), all three the hole this closes —
-# docs/reports/2026-09-12-i21-verifyall.md:3,
-# docs/reports/2026-10-01-E71-force-verification.md:3 and
-# docs/reports/g61-verifyall-2026-09-12.md:3 each open `Verdict: PASS` and state no
-# confidence level anywhere in the document. They stay visible rather than suppressed:
-# retro-fitting a level into a historical record would invent the verifier's own claim.
-# Both directions were controlled in the same run: a scratch report holding
-# `Verdict: PASS` + `Confidence: HIGH` is not flagged, and the same file with
-# `Confidence: mostly` is.
-#
-# Include is deliberately the report corpus. `docs/plans/*brainstorm*.md` states confidence
-# as prose with its rationale (`**Confidence: HIGH.** All primitives … are verified present`,
-# 61 such lines across 18 plans) — a different, human-facing contract that a
-# level-only vocabulary check would drown in false positives, and those plans carry no
-# machine verdict line to anchor on.
-include:
-  - "docs/reports/**/*.md"
-  - "docs/plans/**/*.md"
-exclude:
-  - "docs/tasks*/**"
-rules:
-  - id: verdict-line-confidence-missing
-    description: >
-      A committed verification record that asserts a machine verdict line
-      (`Verdict: PASS|PARTIAL|FAIL`) must follow it with the verifier's confidence level —
-      `Confidence: HIGH|MEDIUM|LOW` — so a reader can tell a verdict the verifier stands
-      behind from one it would not. Contract in sp-code-verification (verify answers) and
-      sp-super-reviewer (review reports); a missing level is a lint failure for pipeline
-      answers (`confidence-missing`), and this rule is the backstop for the reports those
-      surfaces do not read.
-    severity: info
-    evaluator:
-      type: regex
-      config:
-        mode: forbid
-        multiline: true
-        flags: "i"
-        pattern: "^[ \\t]*\\**Verdict\\**:\\**[ \\t]*(?:PASS|PARTIAL|FAIL)\\**[ \\t]*\\n(?:[ \\t]*\\n)*(?![ \\t]*\\**Confidence\\**:\\**[ \\t]*\\**(?:HIGH|MEDIUM|LOW)\\**[ \\t]*\\.?[ \\t]*$)"