enhance(verify-phase): node-test causation control — prove the RED is content-caused (#1346)

The #1279 node-test machine-proof confirmed a known-bad subject drives the
negative test RED, but could not distinguish a genuine content-violation from a
deceptive test that reds merely because GSD_PROHIB_SUBJECT is set. Add an
optional fifth flat scalar `check_clean_fixture` (-> CheckDescriptor.cleanFixture)
threading a KNOWN-CLEAN control subject through projectProhibitions +
descriptorFromProjection. When present, the prover also runs the check against
the clean subject and requires GREEN, so fail-first is proven only when the check
is RED on the violation AND GREEN on the clean subject (content-dependent).

Opt-in and additive: absent a clean fixture the prover behaves exactly as
post-#1314 (no control, documented residual), preserving the zero-authoring
compose path; the lint-rule kind needs no analog (its subject IS the linted
file, no env indirection). Coverage: RED-first deceptive case, positive,
missing-clean fail-closed, round-trip read-back/emit, fast-check property
extended to the 5th scalar, and an end-to-end COMPOSE capstone (honest vs
deceptive). Docs: ADR-550 dated addendum, prohibition-probe reference,
spec-phase + verify-phase workflows.

Closes #1346

Claude-Session: https://claude.ai/code/session_01GsPRb8zvpcT7Eat6vZw8PX
This commit is contained in:
Dave
2026-06-21 10:44:48 -04:00
parent eb81faaeab
commit c09c13f295
11 changed files with 323 additions and 50 deletions

View File

@@ -365,10 +365,15 @@ For each Requirement gathered so far, run the two-stage recall→precision pass:
- `check_target` — the negative-test file path (for `node-test`), or the path to lint
(for `lint-rule`).
- `check_rule` — the eslint rule id (e.g. `local/no-source-grep`); `lint-rule` only.
- `check_violation_fixture` (#1346) — path to a KNOWN-BAD subject the wired check is run
- `check_violation_fixture` (#1279) — path to a KNOWN-BAD subject the wired check is run
against to **machine-prove fail-first**; rides BOTH kinds. Capture it to let the item green
end-to-end with zero hand-authoring at verify time; for `node-test` the negative test should
read its subject from the `GSD_PROHIB_SUBJECT` env var so the prover can inject this fixture.
- `check_clean_fixture` (#1346) — **optional** path to a KNOWN-CLEAN control subject. When
captured, the `node-test` prover also runs the check against it and requires GREEN — proving
the violation's RED is caused by the subject's *content*, not by `GSD_PROHIB_SUBJECT` merely
being set. Capture it for a stronger guarantee; omit it and the check still proves fail-first
on the violation alone (the content-causation residual stays documented for that case).
This is a **SOFT capture (CHK-04): a `test`-tier prohibition WITHOUT a descriptor is still
allowed** — if the author cannot yet name the wired check, leave the descriptor empty and
proceed. It is NOT a hard authoring block; the item simply stays fail-closed/flagged
@@ -395,7 +400,7 @@ For each Requirement gathered so far, run the two-stage recall→precision pass:
written (test or judgment tier); otherwise leave `unresolved`. **`--auto` NEVER auto-dismisses
a prohibition** — a wrong dismissal is the exact silent failure this probe eliminates (PROB-06,
the load-bearing safety property). On a `test`-tier auto-resolution, capture the `check_kind` /
`check_target` / `check_rule` / `check_violation_fixture` descriptor **only when a wired check is unambiguous**; otherwise
`check_target` / `check_rule` / `check_violation_fixture` / `check_clean_fixture` descriptor **only when a wired check is unambiguous**; otherwise
leave it empty — `--auto` NEVER fabricates a check path or fixture (a wrong locate is re-validated and
fails closed at the producer, but a fabricated path is still noise to avoid). Log:
`[auto] prohibitions: R resolved, U unresolved`.
@@ -408,7 +413,7 @@ Populate the `## Prohibitions` section of SPEC.md from the resolved prohibitions
`resolved`/`test` row is a checkable negative acceptance criterion; `resolved`/`judgment`
rows route to judgment review; `⚠ UNRESOLVED` rows are flagged as assumptions). A
`resolved`/`test` row ALSO carries its captured `check_kind` / `check_target` / `check_rule` /
`check_violation_fixture` descriptor when present (so the projection feeds `verify-phase`'s deterministic locate + machine-proof, #1278 + #1346);
`check_violation_fixture` / `check_clean_fixture` descriptor when present (so the projection feeds `verify-phase`'s deterministic locate + machine-proof + causation control, #1278 + #1279 + #1346);
a `test` row with no captured descriptor is still valid — it stays fail-closed/flagged
downstream rather than blocking authoring.

View File

@@ -76,11 +76,11 @@ Aggregate all must_haves across plans for phase-level verification.
gsd_run check prohibition-enforcement <request.json>
```
where `<request.json>` carries `{ prohibition, check, mode }` — `check` being the wired mechanical-check descriptor `{ kind: 'node-test' | 'lint-rule', target, rule?, violationFixture, failFirst? }`, with `kind`/`target`/`rule`/`violationFixture` now sourced from the projected `check_*` scalars (not author/verifier invention — #1278 + #1346). For `node-test`, `target` (from `check_target`) is the negative-test file path; for `lint-rule`, `target` is the PATH to lint and `rule` (from `check_rule`) is the eslint rule id (e.g. `local/no-source-grep`) — both required (a lint-rule without `rule` is not a valid wired check). `violationFixture` (from `check_violation_fixture`) is the path to a KNOWN-BAD subject the producer runs the check against to **machine-prove fail-first** (for `node-test`, injected via the `GSD_PROHIB_SUBJECT` env convention — #1279); `failFirst` is a DEMOTED, non-authoritative hint kept only for backward route-JSON shape (no path greens on it alone — FF-08). The producer LOCATES the wired check from the projection, **machine-proves it is fail-first** by running it against the violation and confirming it goes RED, RUNS it for a genuine non-vacuous pass, builds `enforcementEvidence`, and emits the `dispositionForProhibition()` verdict (#1259 + #1278 + #1279, ADR-550 D5d). Fail-first is **machine-proven, not caller-attested** — absent a provable violation the producer fails closed, never falling back to attestation. Route the result by its typed fields:
where `<request.json>` carries `{ prohibition, check, mode }` — `check` being the wired mechanical-check descriptor `{ kind: 'node-test' | 'lint-rule', target, rule?, violationFixture, cleanFixture?, failFirst? }`, with `kind`/`target`/`rule`/`violationFixture`/`cleanFixture` now sourced from the projected `check_*` scalars (not author/verifier invention — #1278 + #1279 + #1346). For `node-test`, `target` (from `check_target`) is the negative-test file path; for `lint-rule`, `target` is the PATH to lint and `rule` (from `check_rule`) is the eslint rule id (e.g. `local/no-source-grep`) — both required (a lint-rule without `rule` is not a valid wired check). `violationFixture` (from `check_violation_fixture`) is the path to a KNOWN-BAD subject the producer runs the check against to **machine-prove fail-first** (for `node-test`, injected via the `GSD_PROHIB_SUBJECT` env convention — #1279); the optional `cleanFixture` (from `check_clean_fixture`) is a KNOWN-CLEAN control subject the `node-test` prover ALSO requires to stay GREEN, proving the RED is content-caused (#1346); `failFirst` is a DEMOTED, non-authoritative hint kept only for backward route-JSON shape (no path greens on it alone — FF-08). The producer LOCATES the wired check from the projection, **machine-proves it is fail-first** by running it against the violation and confirming it goes RED, RUNS it for a genuine non-vacuous pass, builds `enforcementEvidence`, and emits the `dispositionForProhibition()` verdict (#1259 + #1278 + #1279, ADR-550 D5d). Fail-first is **machine-proven, not caller-attested** — absent a provable violation the producer fails closed, never falling back to attestation. Route the result by its typed fields:
- **`status: 'green'`, `flagged: false`** (a genuinely-passing wired negative test / lint rule, `located: true`, non-empty `evidence`) → the item is satisfiable → it can reach **passed**.
- **missing, non-attested, or genuinely-non-passing check** (`located: false` OR `status: 'unverified'`, `flagged: true`) → **hard-gate**: disposes flagged-unverified, NEVER green, routing to `gaps_found` in BOTH interactive and autonomous modes (a failing mechanical check blocks even AFK; ADR-550 D4 / D3). The deterministic fail-closed default backing every miss/fail is `dispositionForProhibition()` in probe-core (`status: 'unverified'`, `flagged: true` on empty `enforcementEvidence`).
> **Descriptor source — deterministic locate + machine-proof compose (#1278 + #1346, DELIVERED).** The `check` descriptor's `{ kind, target, rule, violationFixture }` is now sourced **deterministically from the projected `check_kind` / `check_target` / `check_rule` / `check_violation_fixture` scalars** on the `must_haves.prohibitions` item (authored at `/gsd:spec-phase`, projected by `projectProhibitions`, read back via the `descriptorFromProjection` adapter). So both halves close with **zero manual descriptor authoring** — the verifier neither invents the locate (#1278) nor hand-supplies the violation fixture (#1346): a prohibition authored with all four scalars machine-proves fail-first and greens end-to-end through the projection alone (removing the spoofable invent-at-verify-time surface; ADR-857 §147 exogenous grading). **Fail-closed is preserved:** an item with NO projected descriptor, a PARTIAL one (e.g. a `lint-rule` missing `check_rule`), OR a descriptor with **no `check_violation_fixture`** makes `descriptorFromProjection` return `null` / an under-specified or fixture-less descriptor, which falls through to the producer's fail-closed paths (`located: false`, or located-but-unprovable) → flagged-unverified, NEVER green, in BOTH modes. `failFirst` is demoted and greens nothing on its own (#1279, FF-08). Residual (tracked **#1346**): the node-test proof confirms the fixture exists and the check goes RED, but cannot generically prove the red was *caused by* the subject's content vs the env merely being set.
> **Descriptor source — deterministic locate + machine-proof compose (#1278 + #1346, DELIVERED).** The `check` descriptor's `{ kind, target, rule, violationFixture }` is now sourced **deterministically from the projected `check_kind` / `check_target` / `check_rule` / `check_violation_fixture` scalars** on the `must_haves.prohibitions` item (authored at `/gsd:spec-phase`, projected by `projectProhibitions`, read back via the `descriptorFromProjection` adapter). So both halves close with **zero manual descriptor authoring** — the verifier neither invents the locate (#1278) nor hand-supplies the violation fixture (#1346): a prohibition authored with all four scalars machine-proves fail-first and greens end-to-end through the projection alone (removing the spoofable invent-at-verify-time surface; ADR-857 §147 exogenous grading). **Fail-closed is preserved:** an item with NO projected descriptor, a PARTIAL one (e.g. a `lint-rule` missing `check_rule`), OR a descriptor with **no `check_violation_fixture`** makes `descriptorFromProjection` return `null` / an under-specified or fixture-less descriptor, which falls through to the producer's fail-closed paths (`located: false`, or located-but-unprovable) → flagged-unverified, NEVER green, in BOTH modes. `failFirst` is demoted and greens nothing on its own (#1279, FF-08). Causation (**#1346**): supplying `check_clean_fixture` adds an opt-in control — the `node-test` prover also requires GREEN on a known-clean subject, proving the RED is content-caused; with no clean fixture that one residual case (a deceptive test reding merely because the env var is set) stays a documented constraint, an author opting into the stronger proof by wiring a clean control.
**Option B: Use Success Criteria from ROADMAP.md**