Build log and development notes

The foundation was committed as a series: core primitives, the audit log, the substrate, the Permission Engine and the CLI, then three test commits and the M1/M2 document.

Sourcecommits c260deb, 87eb97d, a014713, dd77417, ce581d7, 3ecca7d, 2afb5d9, 1d662a6, 8b33536 · docs/M1-M2.md @ 9243204· 8 Oct 2026

Ten invariants are mapped to where they are enforced — for example “unknown never becomes ALLOW” to the decision-matrix tests, and “no probabilistic systems” to the banned-import checks.

Sourcedocs/M1-M2.md (“Invariants and where they are enforced”) @ 9243204 · tests/test_architecture.py::test_no_banned_imports_or_calls @ 9243204

After review, each component has one rules module that both the write path and replay call, so the engine cannot commit an event that replay rejects.

Sourcecommit 6bfa6b1 · docs/M3.md @ 9243204 · tests/test_world_model.py::test_unrecordable_input_never_leaves_an_orphan_allow @ 9243204· 9 Oct 2026

A four-lens adversarial review (invariants, Permission-Engine bypass, determinism and replay, test adequacy) reproduced 32 defects against M3, all fixed.

The review report is not in the repository; its count is quoted, not checkable.

Sourcecommit 6bfa6b1 (message) · docs/BUILD_STATUS.md @ 9243204

“Tests: 511 pass (72 new)” after the M3 hardening.

Sourcecommit 6bfa6b1 (message)· 9 Oct 2026

The nine defects the first M4 review confirmed — among them a stale bond restored from old evidence, a FAIL outrun or cherry-picked away, and a bond dropping its dependencies’ taint — are each pinned by a regression test.

Sourcedocs/M4.md (“Adversarial review”) @ 9243204 · tests/test_bond_hardening.py::test_a_stale_bond_cannot_be_restored_by_re_judging_old_evidence, ::test_admissible_evidence_cannot_be_cherry_picked_away, ::test_a_bond_inherits_the_taint_of_everything_it_relies_on @ 9243204

The second round found that caller-supplied subclasses could decide a verdict that replay then refused, and that unbounded evidence nesting could exhaust the stack; inputs are now canonicalized and nesting is capped at 16 levels.

Sourcecommit f5e7cc4 · tests/test_verification_hardening.py::test_caller_subclasses_cannot_decide_the_verdict, ::test_evidence_nesting_is_bounded @ 9243204· 9 Oct 2026

Every one of those regression tests fails on the original M4 commit 0effd02.

No retained run against 0effd02 shows it.

Sourcedocs/M4.md @ 9243204 · commit f5e7cc4 (message)

“Tests: 617 pass (94 new)” at the original M4 commit.

Sourcecommit 0effd02 (message)· 9 Oct 2026

Mutation-killing tests for the 16 places where a source mutation survived the M4 suite. The review’s verifier did not complete, so the gaps are listed as open rather than claimed.

Sourcedocs/M4.md (“Test adequacy”) · docs/BUILD_STATUS.md @ 9243204

M5 through M12 landed in six commits in under an hour (15:01–15:55 UTC), each with its own test file and its own document.

Sourcecommits 6210943, 9e813b2, f2b03ed, 7716181, 3b57101, 689eb3e · docs/M5.md … docs/M11-M12.md @ 9243204· 9 Oct 2026

Adversarial review of M5–M12, batched into a final validation pass at the owner’s direction.

Sourcedocs/BUILD_STATUS.md (“Final pass: per the user’s direction…”) @ 9243204

From M11–M12 on, the repository gate runs the six experiments on every gate, alongside the tests, lint, formatting, strict typing and the deterministic CLI run.

Sourcescripts/gate.sh @ 9243204 · commit 689eb3e (message)

Scope is chosen by the caller: it records where a belief applies and is not a trust boundary.

Sourcedocs/M3.md (“Behavior that is intentional”) @ 9243204

Inference from invalidated or superseded facts is allowed — reasoning from a negative result is legitimate — while inference from a hypothesis is refused.

Sourcedocs/M3.md @ 9243204 · tests/test_world_model.py::test_inference_from_a_hypothesis_is_refused @ 9243204

Each component assumes it is the only writer of its own events; anything else injecting them is detected at replay rather than prevented.

Sourcedocs/M3.md @ 9243204 · tests/test_domain_authorization.py::test_mutation_with_no_decision_is_detected @ 9243204

A STALE bond keeps its derived confidence: STALE is a lifecycle state, and whether a bond can be relied on is what is_composition_sound() reports.

Sourcedocs/M4.md (refuted findings) @ 9243204