msd-core/docs/adr/1606-prohibition-enforcement-verify-seam.md

ADR 1606: prohibition-enforcement verify-time seam [Proposed]

• Status: Proposed (consolidation ADR — see "Relationship to ADR-550")
• Date: 2026-06-22

Provenance. Drafted 2026-06-22 to promote a decision that accreted as four chronological addendum blocks on ADR-550 (the 2026-06-12 test-tier disposition note that folded in #1259, then #1279, #1346, and #1278) into a single, first-class architecture-of- record for the verify-time enforcement subsystem. Authored by the #644/#1259 implementer. The area (why prohibitions are first-class at all) traces to the author's prohibition-elicitation (N18) and verifier-abstention (N17) findings — the author's own research, cited as motivation, not claimed as a novel GSD contribution; the enforcement mechanism in this ADR is motivated specifically by closing the caller-attestation fake- green hole. Verified against next + src/prohibition-enforcement.cts (≈43KB) and gsd-core/references/prohibition-probe.md.

Relationship to ADR-550 (read this first)

ADR-550 is the spec-phase probe contract: what a prohibition is, how it is represented (SPEC.md acceptance criterion ↔ must_haves.prohibitions:), and how it is tiered (test vs judgment, Decisions 3–7). That ownership is unchanged.

This ADR carves out and consolidates the verify-time enforcement mechanism for the test tier — the check prohibition-enforcement producer in src/prohibition-enforcement.cts — which ADR-550 only documents as a chain of dated addenda. The boundary:

Concern	Owner
Prohibition representation, tiering, spec→plan projection	ADR-550 (D3, D7)
Judgment-tier soft-gate / "never a silent pass" policy	ADR-550 (D4)
Test-tier enforcement producer (locate → prove-fail-first → run → dispose)	this ADR
Capability/core-rail placement of the verifier↔predicate contract	ADR-857 "Verification substrate vs. plug-in tier (the predicate boundary)" (decision #6), referenced by both

Dedup proposal (decide at PR review): on accepting this ADR, replace ADR-550's 2026-06-12 / #1259 / #1279 / #1346 / #1278 enforcement addenda with a one-line pointer to this ADR, leaving 550 to own the contract and this ADR to own the mechanism. Until that is agreed, 550's addenda remain authoritative and this ADR is non-binding.

Why this is still Proposed (audited 2026-07-17)

What shipped. The audit confirmed the enforcement mechanism this ADR describes is real and in place, not aspirational. All seven Decision points are present in src/prohibition-enforcement.cts and src/probe-core.cts: runProhibitionEnforcement (src/prohibition-enforcement.cts:654-732), dispositionForProhibition (src/probe-core.cts:470-516), the vacuity guards isNonVacuousNodeTestRed (src/prohibition-enforcement.cts:351-355) and isNonVacuousNodeTestPass, and defaultProveFailFirst's node-test branch (src/prohibition-enforcement.cts:591-627) — which does implement the #1906 mandatory-cleanFixture causation control exactly as Decision 4 / the 2026-07-03 addendum describe, not merely as a documented intent. Test coverage is substantial (tests/prohibition-enforcement.test.cjs, 1336 lines), and all six contributing issues (#644, #1259, #1278, #1279, #1346, #1906) are closed as COMPLETED on GitHub.

The blocker. This ADR names its own precondition for becoming binding, in its own words: "on accepting this ADR, replace ADR-550's 2026-06-12 / #1259 / #1279 / #1346 / #1278 enforcement addenda with a one-line pointer to this ADR ... Until that is agreed, 550's addenda remain authoritative and this ADR is non-binding." That dedup has not happened. Direct read of docs/adr/550-spec-phase-probe-contract.md confirms all four named addenda — "Addendum (2026-06-12; updated 2026-06-15)", "Addendum (2026-06-15, #1279)", "Addendum (2026-06-21, #1346)", and "Addendum (2026-06-15): optional check descriptor ... (#1278)" — remain in ADR-550 in full, verbatim; none has been collapsed to a pointer. A later, separate addendum in ADR-550 (2026-06-22, from #1607) does cross-reference ADR-1606 for the recall/representation-side "Alternatives considered," but that is additive scaffolding, not the enforcement-addenda dedup this ADR names as its own condition — the four target addenda are untouched by it. No commit, PR, or tracked issue was found executing the dedup.

Unblock condition. Edit docs/adr/550-spec-phase-probe-contract.md to collapse the four named addenda (2026-06-12/2026-06-15 update, #1279, #1346, #1278) into the one-line pointer this ADR calls for, then flip both ADR-550's cross-reference and this ADR's Status in the same PR. To check in minutes: grep docs/adr/550-spec-phase-probe-contract.md for ## Addendum (2026-06-12, #1279, #1346, and #1278 — if those headings still carry the full addendum text rather than a one-line pointer, the precondition remains unmet.

Context

ADR-550 D4 originally specified the test tier as a "hard gate in both interactive and autonomous modes" but left the mechanism unspecified. As the prohibition probe shipped (#644) and grew an enforcement producer (#1259→#1346), a set of load-bearing decisions had to be made that are not derivable from the contract alone:

• A generic producer cannot trust that a wired check actually fails on a violation. Caller attestation (failFirst: true) is unfalsifiable at verify time and was a fake-green hole.
• A generic producer cannot synthesize a violation for an arbitrary check, so proving fail-first requires an author-supplied known-bad subject.
• "The test went red" is not proof the content caused the red — an env-var-triggered or load-crash red forges a green.
• The check descriptor must round-trip through the flat parseMustHavesBlock shared by truths/artifacts/key_links without a parser rewrite that would regress those readers.
• A missing, partial, or un-provable check must fail closed, never silently pass — the whole point of the tier.

Decision

1. No path greens on attestation alone. A test-tier prohibition reaches passed/green only when its wired check (a) genuinely, non-vacuously runs and passes AND (b) is independently machine-proven fail-first against a known violation. The enforcement producer (runProhibitionEnforcement in src/prohibition-enforcement.cts) computes this as passed = proof.provenFailFirst === true && run.passed === true and only then emits non-empty enforcementEvidence. The disposition step (dispositionForProhibition in src/probe-core.cts) then greens a test-tier item only on that non-empty evidence; every other outcome — missing check, partial/invalid descriptor, can't-prove, throws, times out, no violation source, passes-on-violation, located:false — yields {status:'unverified', flagged:true} (gaps_found, never green) in both interactive and autonomous modes. (The green rule and the fail-closed default are intentionally split across the producer and the disposition so the disposition can fail closed even when the producer never ran.) This is the standing form of ADR-550 D4's guarantee.

2. Two wired-check kinds, one producer. The producer accepts exactly two mechanisms:

   • node-test — a node --test negative test that reports a real, non-vacuous failing test. Two distinct vacuity guards apply: isNonVacuousNodeTestRed rejects a load-crash red by requiring a failing test named distinctly from the target file; isNonVacuousNodeTestPass rejects the empty-file "0 tests = pass" forgery on the pass side. (Both live in src/prohibition-enforcement.cts.)
   • lint-rule — a lint/AST rule run through the project flat config as eslint --format json and filtered by ruleId (so local/* plugin rules load; a bare --rule cannot). Invoked via process.execPath against the resolved eslint CLI, not a bare eslint binary. Dogfooded on the in-tree local/no-source-grep rule.

3. Machine-proven fail-first via an author-supplied violation fixture. Before a clean pass can green, the producer runs the wired check against a known-bad subject and requires RED. The subject is sourced from violationFixture:

   • lint-rule: a file whose content violates rule; the rule id must appear in the JSON report (the rule must have teeth).
   • node-test: injected into the child as GSD_PROHIB_SUBJECT=<violationFixture>; the negative test reads that env var to locate its subject and must go RED against it.
   • Absent fixture → fail closed (never attestation). Existence is checked (fs.existsSync(path.resolve(cwd, fixture))) before spawning, symmetric with the lint-rule path which fail-closes on a < 1-file result.

4. Causation control (opt-in). A node-test red proves nothing about why it is red. An optional cleanFixture runs the same negative test a second time with GSD_PROHIB_SUBJECT=<cleanFixture> and requires non-vacuous GREEN (isNonVacuousNodeTestPass) — so fail-first is proven only when the check is RED on the violation AND GREEN on the clean subject (content-dependent red). Opt-in, not mandatory: absent cleanFixture, behaviour matches the pre-#1346 zero-authoring compose path and the "reds because the env var is set" case stays a documented residual for that one author's check. MANDATORY for the node-test kind as of #1906 (2026-07-03) — absent cleanFixture the node-test is un-provable (fail-closed), never proven on the violation alone; see the #1906 addendum below. The lint-rule kind needs no analog (its subject is the linted file; no env-var indirection).

5. Deterministic locate via five flat scalars — never a nested object. The wired-check descriptor is authored at spec-phase and projected onto the must_haves.prohibitions item as five flat scalar keys: check_kind, check_target, check_rule (lint-rule only), check_violation_fixture, check_clean_fixture (optional). A nested check: {} object is rejected: parseMustHavesBlock is a flat parser and its reconstructFrontmatter serializer is lossy for nested object-lists, so a nested shape would mangle the round-trip and risk the shared truths/artifacts/key_links readers. projectProhibitions (in src/probe-core.cts) emits the scalars only for a well-formed descriptor; descriptorFromProjection (in src/prohibition-enforcement.cts) reads them back into the { kind, target, rule? } CheckDescriptor. A prohibition authored with all five scalars machine-proves fail-first and greens end-to-end through the projection with zero hand-authoring.

6. failFirst is demoted, not removed (FF-08). The CheckDescriptor.failFirst field is kept for route-JSON backward-compat but demoted to a non-authoritative hint — the machine prover supersedes it. Removal was rejected (breaks the route-JSON shape mid- migration); demote-and-ignore satisfies "no path greens on attestation."

7. The contract is the CI surface, not the LLM. Per ADR-550 D5, CI deterministically tests parse/validate, the projection round-trip (fast-check property + CHK-03), the fail-closed guards (CHK-06), and backward-compat (CHK-07) — never the model's judgment. This ADR adds no CI claim over LLM behaviour.

Consequences

• Positive: the verify-time enforcement subsystem has a single architecture-of-record instead of five addenda on a spec-phase ADR; the "never a silent pass" guarantee is stated once, completely, with its fail-closed defaults; the descriptor's flat-scalar shape and its rationale are findable without reading the frontmatter parser.
• Costs: one more ADR to keep in sync with src/prohibition-enforcement.cts; the dedup against ADR-550's addenda must actually be executed at PR time or the repo carries two homes for the same decision (the explicit risk this ADR is meant to remove).
• Open conventions (renamable at review, zero live consumers): GSD_PROHIB_SUBJECT and the check_violation_fixture/check_clean_fixture scalars have no in-tree node-test consumer yet (the only live dogfood is the lint-rule local/no-source-grep; node-test fail-first is exercised only by synthetic temp fixtures). A rename or an argv-for-env-var swap is a mechanical zero-migration find/replace — surfaced here for the maintainer to settle at review, exactly as #1278/#1279 were. (Hyrum's Law: this ADR deliberately marks them as not-yet-depended-on so they remain changeable; the scalar key names emitted by projectProhibitions in shipped code are, by contrast, already a contract.)
• Boundary held: canon security/compliance is referred to /gsd:secure-phase + eslint, not minted here (ADR-550 D6); this ADR governs only bespoke product/values test-tier enforcement.

Alternatives considered (rejected & deferred)

Enforcement-side alternatives, each with the standing reason and a re-open condition. (The recall/representation/packaging-side alternatives — the withdrawn LLM classifier #652, a deterministic recall engine, a polarity field on truths, and the deferred dispatcher CLI — belong to the spec-phase contract and are recorded in ADR-550's "Alternatives considered.")

• A nested check: {} descriptor object — REJECTED. parseMustHavesBlock is a flat parser and reconstructFrontmatter is lossy for nested object-lists, so a nested shape would mangle the round-trip and could regress the shared truths/artifacts/key_links readers. Hence the five flat scalar keys (Decision 5). Re-open only if parseMustHavesBlock is replaced with a structured parser (its own ADR, with the full shared-reader regression surface).
• An inline producer-written violation snippet — REJECTED. To machine-prove fail-first the violation is an author-supplied fixture path (check_violation_fixture), not source the producer writes inline, which would bake rule-specific source into a generic producer (Decision 3).
• failFirst caller attestation as authoritative — REJECTED → DEMOTED (FF-08). Trusting the caller's failFirst: true was an unfalsifiable fake-green hole; the machine prover supersedes it and the field is demoted to a non-authoritative hint kept only for route-JSON backward-compat (Decision 6). Outright removal was also rejected (breaks the route-JSON shape mid-migration).
• Mandatory causation control — REJECTED in favour of opt-in. Requiring every node-test prohibition to ship a cleanFixture would regress the zero-authoring compose path and hard-gate every existing descriptor without one; the control is opt-in (Decision 4), leaving one documented residual rather than breaking working checks. ↳ SUPERSEDED 2026-07-03 (#1906) — see the addendum below. The blast-radius premise no longer holds: there is no in-tree node-test consumer (Consequences, "Open conventions"), so making the control mandatory hard-gates zero existing checks. Decision 4 is now mandatory for the node-test kind.

Addendum (2026-07-03, #1906) — node-test causation control is now MANDATORY (supersedes Decision 4's opt-in)

Decision 4 made the cleanFixture causation control opt-in to avoid regressing the #1314 zero-authoring compose path. That trade-off left a Goodhart hole open by default: a node-test that omits cleanFixture is proven fail-first on the violation alone, so a deceptive content-independent negative test — one that reds merely because GSD_PROHIB_SUBJECT is set, ignoring the subject's CONTENT (assert.ok(!process.env.GSD_PROHIB_SUBJECT)) — passes the proof. The proof's observed signal (RED) thus diverges from its target (RED caused by content), and the divergence is the default, not an edge case an author opts into.

Decision (owner ruling on #1906, 2026-07-03): for the node-test kind the causation control is required. A node-test descriptor that omits cleanFixture is treated as un-provable (fail-closed) — never accepted under the weaker violation-only proof. When a clean fixture is present, fail-first is proven exactly as before (RED on the violation subject AND non-vacuous GREEN on the clean subject); a deceptive content-independent test reds on the clean subject too, so the control fails and the proof does not pass.

• Why the opt-in's rationale no longer applies. Decision 4 / the rejected-alternative above kept the control opt-in to avoid hard-gating "every existing descriptor without one." Per this ADR's own Consequences ("Open conventions … no in-tree node-test consumer"), there are none — the only live dogfood is the lint-rule local/no-source-grep; node-test fail-first is exercised only by synthetic temp fixtures. So the mandatory control hard-gates zero real checks today; its cost is paid only by future node-test authors, who now must supply a clean control subject to earn a green.
• Scope — node-test only. The lint-rule kind is unchanged (byte-identical): its subject is the linted file, with no GSD_PROHIB_SUBJECT indirection, so the "reds because the env var is set" gap cannot exist there. Net effect on D4's disposition is unchanged — every miss/fail/un-provable still hard-gates; this only tightens what counts as proven.
• Breaking change (Hyrum). A previously-green node-test prohibition with no clean fixture now hard-gates. Disclosed as breaking with the ~zero in-tree blast radius noted above. cleanFixture stays optional at the type level (it rides both kinds; the lint-rule kind never uses it) but is required by the node-test prover.
• Mechanism. defaultProveFailFirst's node-test branch in src/prohibition-enforcement.cts: const clean = check.cleanFixture; if (!clean) return { provenFailFirst: false, … } before the existence + non-vacuous-GREEN control. Compiled by build:lib. ADR-550's 2026-06-21 (#1346) addendum "Why opt-in, not required" is superseded to match.

Cross-references

• ADR-550 — spec-phase probe contract; this ADR consolidates its enforcement addenda, and ADR-550 holds the recall/representation/packaging-side rejected alternatives.
• ADR-857 — section "Verification substrate vs. plug-in tier (the predicate boundary)" (decision #6): the verifier↔predicate contract lands on the core verify rail (non-toggleable), never in capabilities/; this seam is its concrete enforcement instance.
• gsd-core/references/prohibition-probe.md — the portable runtime reference.
• docs/how-to/resolve-prohibition-findings.md — user-facing resolution guide.
• Code: src/prohibition-enforcement.cts, src/probe-core.cts (projectProhibitions), gsd-core/workflows/verify-phase.md. Issues: #644, #1259, #1278, #1279, #1346, #1906.