Standing record of package-level decisions: queued proposals, rejections with reasons,
and open questions. D-NNN ids are durable — never reused, never renumbered; a
reversed decision keeps its id and gains a superseding entry. Decision events are
still recorded as ADRs in adr/ (the ADR carries the full rationale; this ledger is the
"is this settled?" index). Adopted as package practice 2026-06-10 via
adr/0003-cross-application-review-dispositions.md (the review's P-012, part (a));
the matching audit evidence target shipped at the v3.5 cut (D-012).
Substance below is restated in this package's terms from its sources — the P-NNN /
R-NNN ids are the 2026-06-10 cross-application review's namespace, transcribed through
ADR 0003 (the report itself is untracked input evidence per MANIFEST.md, "Not
bundled"). No automated check polices that a ledger entry faithfully restates its
source; that restatement-policing check is not built. (Stating that here keeps this
file out of the over-claim class recorded in ADR 0001's Consequences, where MANIFEST.md
once promised automation the drift linter did not implement.) Entries also carrying
drafting constraints from the 2026-06-10 forward plan — an operator-session plan, not
a tracked artifact — name it in their Source line; the ledger entry itself is that
plan's durable record.
All shipped additively at the 2026-06-10 dual cut
(adr/0004-evidence-discipline-and-enforcement-honesty.md): no phase renumbering, no
new scored dimension, new fields optional; v3.4 audits and v1.5 profiles stay valid.
Before drafting, each premise was adversarially re-verified against the v3.4/v1.5
texts (two of the source review's confidence-1.0 claims died exactly that way — see
D-014/D-015); the reshapes that re-verification forced are recorded in ADR 0004's
Context.
- Status: Shipped (Audit Spec v3.5 cut, 2026-06-10; ADR 0004) · Source: ADR 0003 (P-001) · Priority: High
- The audit spec's Phase 10 gains a rule (the sources write "Audit Spec §10"; this
ledger normalizes to the Phase 10 rule cluster — its numbered 10.1/10.2 subsections —
since the §10 worked example carries no rules): per-cycle finding and fitness-function
ids are either monotonic and never reused across cycles (the strong form), or
mandatorily cycle-qualified (
YYYY-MM-DD/prefix) in any artifact that outlives its cycle — READMEs, configs, instruction files, commit messages. (The source proposal also listed ADRs; this package's own convention exempts ADRs as dated artifacts — the spec rule's final scope is decided at the cut.) The rule governs ids as cited in long-lived artifacts; the anchored in-run id patterns in the audit spec's §8.11/§8.12 schemas stay unchanged (a qualified id would violate them — at most an optional sibling cycle field is added). The package already practices and mechanically enforces the weak form (seeMANIFEST.md, "Identifier convention"). Strong-vs-weak was resolved at the cut (D-016): the shipped rule admits either discipline.
- Status: Shipped (Audit Spec v3.5 cut, 2026-06-10; ADR 0004) · Source: ADR 0003 (P-002) + the 2026-06-10 forward plan · Priority: Medium
- The citation schema (path / lines / evidence / method, plus the existing
snapshot_ref/current_refpair) gains optional observed-at, valid-until, and evidence-lane fields, promoting the three release-provenance lanes (repo-local / GitHub-hosted / unavailable, already inCONTRIBUTING.md; the spec enum will need tool-neutral lane naming per ADR 0001's drafting carve-out) from prose into the schema — so steady-state and focused-diff consumers of prior findings see a staleness window per citation instead of trusting one ship-time re-verification. Designed jointly with D-008; the new fields extendsnapshot_ref/current_ref, never duplicate them.
- Status: Shipped (Audit Spec v3.5 cut, 2026-06-10; ADR 0004) · Source: ADR 0003 (P-003) · Priority: High
- A named flag for governance or instruction text that claims automated enforcement which does not exist: every "CI-enforced", "blocked at merge", "automated by X" claim must map to an implemented gate, else it is flagged. The class is proven three times across two governed repos (this package's own over-claim recorded in ADR 0001's Consequences; two instances in the reviewed substrate repo); today it is only found incidentally.
- Status: Shipped (Audit Spec v3.5 cut, 2026-06-10; ADR 0004) · Source: ADR 0003 (P-004) · Priority: High
- The cross-agent instruction-contract checks gain a flag for instruction contracts citing artifacts that do not exist in the tree (e.g. a definition-of-done requiring a checklist from a template the repo does not contain). An unsatisfiable checklist item silently trains agents to skip checklist items.
- Status: Shipped (Audit Spec v3.5 cut, 2026-06-10, as §9.9 + the Phase 6 check; ADR 0004 — drafted last with drop authority, and the drop condition was not met) · Source: ADR 0003 (P-005) + the 2026-06-10 forward plan · Priority: High
- Map mandated review/approval friction against actual enforcement power per surface and flag inversions: hard-enforcing surfaces (e.g. scripts whose gates block merges) that carry no review requirement while weaker, warn-only surfaces carry heavy ones. The audit records authority facets and governance separately today; the ratio is where the reviewed repo's real gap hid. Caveat: sole-source evidence from one external repo, never independently re-verified here. If it cannot be worded operationally (blocks-merge / blocks-commit / warn-only / prose) and tool-neutrally at drafting time, it is moved to this ledger as Dropped with revival condition "corroborated by a second target repo" (D-017 — resolved as shipped; the drop condition was not met).
- Status: Shipped (Audit Spec v3.5 cut, 2026-06-10; ADR 0004) · Source: ADR 0003 (P-006) · Priority: Medium
- The Phase 9 suite classification gains a regression-trap class: behavioral traps seeded from observed agent failures (distinct from code regression tests), with the existing fitness-function demotion rule applied honestly — a trap corpus with no runnable harness is recorded as manual-review-only, not as coverage. Drafting constraint: the suite-mode enum is closed and declared twice — in the audit spec's §8.10 EvalSuite schema and in the Phase 9 prose — so both are extended in the same commit.
- Status: Shipped (Audit Spec v3.5 + Profile Directive v1.6 cut, 2026-06-10; ADR 0004) · Source: ADR 0003 (P-007) · Priority: High
- Three parts, shipped together: (a) a Phase 5 note that the audit's store-class enum
values (
long-term-memory,operator-rules-memory, …) are audit-internal taxonomy, never recommended project vocabulary; (b) the profile directive's §5 schema may declare forbidden terms the project's governance bans; (c) audit artifacts must restate such terms in the project's canonical vocabulary, and the Phase 1 collision check includes collisions between the audit's taxonomy and the project's controlled vocabulary. A project that hard-bans a term (the reviewed repo bans calling shared state "memory") must never have it injected by the audit's own artifacts.
- Status: Shipped (Profile Directive v1.6 + Audit Spec v3.5 cut, 2026-06-10; ADR 0004) · Source: ADR 0003 (P-008) + the 2026-06-10 forward plan · Priority: High
- The profile's agent surface gains: observed agent CLI version(s), observed model posture, and any alias pins (e.g. a settings file pinning a model alias) with what the alias currently resolves to; the audit gains a drift check for alias pins that no longer resolve to the recorded baseline. Silent vendor-side repointing of an alias is a failure class for every agent-operated repo, and the profile currently has no field where it would surface. Designed jointly with D-002 (same freshness machinery).
- Status: Shipped (Audit Spec v3.5 cut, 2026-06-10; ADR 0004) · Source: ADR 0003 (P-009) · Priority: High
- Where operator-rules/agent-memory stores exist, the audit records whether the project has an invalidation discipline — how agent-held claims get corrected when repo facts change — and flags memory surfaces with no such path. The spec inventories memory stores' owners and retention, but not their truth-maintenance relationship to the repo.
- Status: Shipped (Audit Spec v3.5 cut, 2026-06-10; ADR 0004) · Source: ADR 0003 (P-010) · Priority: Medium
- An implemented fitness function with non-trivial matching logic should ship a negative self-test wired to run wherever the fitness function runs; one without it carries reduced confidence in scoring. This names the package's own proven in-house pattern (the always-run drift-linter self-test hook) as a rule for adopters.
- Status: Queued (package patch, authored after the v1.6 cut so the new fields are included) · Source: ADR 0003 (P-011) + the 2026-06-10 forward plan · Priority: Medium
- Ship a JSON Schema for
project_profile.yamlas a derived artifact and validate it in the quality gate, with known-good and poisoned fixtures in the linter's--self-test. The profile schema currently exists only as prose in the profile directive's §5; the directive's procedural validations (citation fidelity, lockfile-over-manifest) stay agent-side — structural validity becomes mechanical.
- Status: Practice implemented 2026-06-10 (this file); the audit-spec evidence target (does the project maintain a standing open-question queue distinct from its ADRs? — a §11.10 governance item) shipped at the same cut (Audit Spec v3.5, 2026-06-10; ADR 0004) · Source: ADR 0003 (P-012) · Priority: Medium
- Status: Implemented 2026-06-10 (gate hardening;
MANIFEST.md"Content-hash binding" + drift-linter check with negative self-test) · Source: ADR 0003 (P-013) · Priority: Medium - Each derived file is attested against the sha256 of the authority texts it was last
synced to; the linter fails on divergence; a cut re-attests after re-syncing. Absorbs
the residual of D-015: the inline
(per AGENTS.md)cites were added 2026-06-10, and what the binding deliberately does not prove (semantic agreement of restatements) stays human review, perMANIFEST.md.
- The cited absolute-path file is gitignored personal scaffolding that never ships; the tracked wiring already uses the portable form; the harm premise ("breaks on any other clone") fails because the file does not exist on any other clone. Mistaking local machine state for repo content is a recorded false-positive class. Full refutation in ADR 0003, Context.
D-015 — Review claim R-004 (rule restated five times, no canonical source) — Rejected as stated 2026-06-10
- The hook-bypass ban appears in at least eight tracked statements, not five; a
canonical source is named:
.gitmessagecitesAGENTS.mdby name, and the skill labels its list a reminder checklist with its own source-of-truth section (two separate mitigations, per ADR 0003). Residual accepted, not the claim: three restatements lacked inline cites (added 2026-06-10, per ADR 0003 through D-013's scope), and no automated check polices the restatement set — that check remains unbuilt (see the header note). Full refutation in ADR 0003, Context.
- Status: Resolved 2026-06-10 (ADR 0004, Decides 1) · Source: ADR 0003 ("Does not decide")
- Monotonic never-reused ids versus mandatory cycle-qualification. Resolved by admitting both: the shipped rule requires a project to record one of the two disciplines and keep it; the package's own convention (cycle-qualification on undated surfaces) satisfies it.
- Status: Resolved 2026-06-10 — shipped (ADR 0004, Decides 4) · Source: ADR 0003 + the 2026-06-10 forward plan
- Drop condition (cannot be worded operationally and tool-neutrally) was tested during drafting and not met: §9.9 ships two short ordered scales and a decidable pairwise inversion condition. The sole-source caveat is recorded in ADR 0004's Carve-outs; revival path if field use disproves the wording is a superseding ADR.
- Status: Closed 2026-06-10 — tripwire not reached · Source: 2026-06-10 forward plan
- The v3.5/v1.6 cut shipped before the 2026-06-24 tripwire; the landed patches ride
into the
v1.5.0minor as planned, and nov1.4.1tag is needed.