approval-engine APPROVAL-WP-0002-T01 is still progress and its namespace has no
pods, so the live end-to-end proof cannot run. Built everything that does not
depend on it, with the engine behind a Protocol plus a fake carrying its real
refusal semantics, so its arrival is a wiring change rather than a build.
- memo.py: the Decision Memo, versions, binding document. Principal, Scope,
Awareness and Hat are dataclasses rather than dicts because the canonicalizer
requires a shape and a missing key should fail at construction rather than
deep inside hashing — which is exactly how it failed twice while building
this. Field names follow the governed canonicalizer (item_id, severity,
locator): the published vectors are the contract, so the object was aligned to
them rather than the reverse.
- presentation.py: the sole writer of view_hash. One writer, one canonicalizer,
one place to audit. Acknowledgment is an explicit method call and nothing
infers it from scroll, dwell or focus.
- disposition.py: verbs and guards G_NOAGENT, G_STEP, G_PRES, G_ACK, G_REASONS,
G_SEALED. accept is ABSENT from weak steps rather than present-and-disabled,
because a greyed-out accept still teaches the wrong model. Only accept reaches
the engine; a memo return is not represented there at all.
- provenance.py: claim routes per A-16. assert_human_control_dischargeable
refuses a registration-supplied human, so PR-11's limitation fires at the
point of use instead of sitting in a document.
- evidence.py: local outbox, commitment-only records carrying the
GH-DEC-2026-014 §4 existence assertion, per-class reconciliation counts, and a
custody-locator guard that rejects credentialed URLs (PR-12).
- approval_client.py: 409 duplicate_approver is success, 409 conflict terminal,
503 fail-closed, approval:consume refused before a token is requested.
87 tests pass, including every negative case in the Use Case Catalog and that a
fail-closed outcome is recorded as a stance application with no verb field —
never as a decline, because the human did not make one.
T08 stays progress: the live proof is the remainder.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3W1dQG7GFFM9d94jFx7iR
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 1565372@bnt-lap001
Assistant-Session: 16bb2f25-b34c-49ef-8e94-5fec3567a568
Gate House ruled all three questions within a day, attributing the speed to the
request being filed before the architecture with candidate answers and their
costs.
R1 PEP-shaped, confirmed as proposed. The ruling settles the shape; the layer
stays ours to declare, so layer.yaml is written in this repository's voice
rather than transcribed from the reply.
R2 yes to a presentation claim, no second catalog row, under three limits now
declared in layer.yaml and tested. Limit 2 — the claim must never be an input to
the decision it presents for — is load-bearing: our self-dealing argument was
accepted because it holds, not despite it. Limit 3 drives architecture, since
here the actor being audited and the evidence source are the same component.
R3 (b) with the authority rule: binding digest authoritative for what the
request is, view_hash only for what was shown, neither substitutable, and a
disagreement between them is a finding against the presenting surface rather
than a fact about the request. Linkage is co-reference; nesting was refused
because it reproduces the GH-DEC-2026-008 hash cycle.
Built to v0.8 obligation 3 rather than migrating later: axis enumerated, unknown
resolves to fail_closed, absent distinguishable from unknown in the record, and
published-equals-shipped asserted by test rather than claimed. Every stance is
fail_closed, which is a conclusion not a shortcut — ops-warden can justify
fail_open on a continuity argument that does not exist here.
GH-DEC-2026-010 inherited as a declared gap in four documents: a decision cannot
today be proven to have come from access-engine. The decision path must not be
described as validated while FLEX-WP-0024 is open.
46 tests pass. T05 and T07 unblocked.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3W1dQG7GFFM9d94jFx7iR
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 1565372@bnt-lap001
Assistant-Session: 16bb2f25-b34c-49ef-8e94-5fec3567a568
Verified the three published hashes reproduce byte for byte before promoting
anything, then moved the schema, canonicalizer and vectors into governed assets.
history/20260909-initial-exploration/ is untouched and stays the provenance
record.
- schemas/, informed_decision/, tests/vectors/ populated; the reference
canonicalizer's ad-hoc __main__ block replaced by a real
`python -m informed_decision` entry point.
- tests/test_canonicalize.py — 20 tests, all green. Published vectors, all four
isolation properties, canonical-form round-trip, key sorting, and a provenance
test asserting the governed fixtures have not drifted from history/.
- docs/specs/EvidenceModel.md — the two hashes, the split and why it exists, the
four isolation properties, the presentation record, the bundle, and the
relationship to audit-core.
- pyproject.toml, Makefile.
One test of mine was wrong on first run: it scanned for ", " to assert no
insignificant whitespace, which fires on prose inside a brief. Replaced with a
canonical round-trip comparison, which is the property actually meant. The
canonicalizer was correct.
EvidenceModel leads with what the model does NOT claim — no proof of
comprehension, no proof of reading (deliberately, since the alternative is
surveillance), no survival of a compromised surface, and audit-core's inherited
bound that a hash chain cannot prove a record was never sent.
T06 stays progress: the SCOPE.md rewrite is gated on the T02 ruling.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3W1dQG7GFFM9d94jFx7iR
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 1565372@bnt-lap001
Assistant-Session: 16bb2f25-b34c-49ef-8e94-5fec3567a568
