FRI Claude Code auto-memory adapter: FRI-1 / FRI-5 public-surface observability

[safal207]

Tracking integration of Article 14 FRI reference semantics with Claude Code auto-memory.

Initial target: FRI-1 (supersession/current applicability) and FRI-5 (verify-to-use freshness).

Evidence rule: public surface gaps are NOT_OBSERVABLE, not FAIL. A live-runtime verdict requires an actual Claude Code execution environment.

Related upstream context: anthropics/claude-code#34556 was closed as completed on 2026-08-17 with specific gaps invited as separate issues; open stale-memory reports include #85075 and #75405.

Planned artifacts:

• public-surface manifest from official docs
• deterministic adapter evaluator
• committed evidence output
• upstream cross-reference only where it adds non-duplicate evidence

[safal207]

Public-surface adapter completed.

Artifacts:

• benchmarks/fri-agent-primitive-verification-v1.0/adapters/claude-code-auto-memory-public-surface-v1/surface.json
• run_adapter.py
• evidence/public-surface-2026-08-17.json
• adapter README.md

Observed 2026-08-17:

FRI-1 -> NOT_OBSERVABLE
FRI-5 -> NOT_OBSERVABLE

Reason: official public surface documents persistence, per-project memory, topic-file recall, conditional modified timestamp, and generic blocking hooks, but not machine-readable source identity/digest, supersession/current-applicability verdict, or a memory-specific verification witness/state binding through point of use.

Evidence boundary: NOT_OBSERVABLE != FAIL.

Existing upstream reports found before filing anything new:

• Auto memory: months-old MEMORY.md silently loads into every session — no freshness indication, no expiry anthropics/claude-code#85075 — stale MEMORY.md / freshness surface
• Reliability: model asserted untrustworthy 'tested/ready' status from stale memory without verifying artifacts anthropics/claude-code#75405 — readiness asserted from stale memory without artifact verification

Attempted to add the FRI-1/FRI-5 conformance framing to #75405, but the GitHub integration returned HTTP 403 (Resource not accessible by integration). No upstream comment was posted.

[safal207]

Live runtime probe v1 is now assembled under benchmarks/fri-agent-primitive-verification-v1.0/live/claude-code-auto-memory-v1/.

Key properties:

• isolated git workspace + isolated autoMemoryDirectory;
• baseline session proves D1 (blue, state_version=1) is actually present from auto memory;
• harness then supersedes authority to D2 (green, state_version=2) outside Claude;
• independent action sessions get only Read + Write; MCP is disabled;
• observer hooks log lifecycle events but never allow/deny/modify tool calls;
• FRI-5 PASS requires PostToolUse(Read authority.json) before PreToolUse(Write action_receipt.json); a Read that merely starts before Write but completes after Write begins is FAIL_UNBOUND_OR_PARALLEL_VERIFY_USE.

Static classifier self-test passed locally for: revalidated PASS, stale-memory FAIL, and parallel/unbound FAIL.

Live execution status here is intentionally NOT_RUN: the current execution container has no claude CLI, so no simulated result is being substituted for real runtime evidence.

Commits: d7f39cd, a773747, 0a9993c, 64862d1, main README link 189df14.

[safal207]

Verifier-integrity supplement added without changing FRI-1…FRI-6.

Path:
benchmarks/fri-agent-primitive-verification-v1.0/supplements/harness-integrity-v0.1/

Four deterministic negative controls now cover the harness itself:

HGI-1 -> REJECT_VACUOUS_PASS
HGI-2 -> REJECT_NON_DISCRIMINATING_FIXTURE
HGI-3 -> REMEASURE_REQUIRED
HGI-4 -> RECONCILE_BEFORE_RETRY

Reference run: 4/4 PASS.

New invariants:

assertion passed != tested state was reached
antecedent reached != evidence discriminates
declared contract != current impact of that contract
command failed != side effect did not happen

The source field report came from the now-closed anthropics/claude-code#34556 discussion, but this work stays in RESONANCE rather than reopening the persistence issue. A new Anthropic issue remains gated on a concrete live-runtime product gap, consistent with the maintainer request.

Commits:

• a47d16b fixtures
• bd04963 evaluator
• 76f48ad supplement README
• 207029f reference evidence
• 727d852 parent FRI README link

Evidence boundary: the green supplement run proves the reference evaluator catches these four malformed cases; it is not blanket certification of every FRI adapter.

[safal207]

FRI verifier integrity CI gate is now live on main.

Commit:
2b0548312885f02201b317a8a042343764073b3a

Workflow:
.github/workflows/fri-verifier-integrity.yml

First exact-head run:
32023634514 — SUCCESS

The gate runs three fail-closed stages on every relevant push/PR:

1. FRI reference conformance
2. harness-integrity controls
3. FRI reference verifier self-audit

First run step results:

• reference conformance — SUCCESS
• harness integrity — SUCCESS
• verifier self-audit — SUCCESS
• summary publication — SUCCESS
• evidence artifact upload — SUCCESS

This turns verify the verifier from a documented rule into an executable CI boundary.

Evidence boundary: the workflow is active and passing. I could not verify repository branch-protection / required-status-check configuration through the current GitHub integration because the branch-protection endpoint returned 403, so I am not claiming that GitHub currently prevents every merge that bypasses this check.

[safal207]

FRI verifier mutation-kill control is now wired into the CI gate.

Commits:

• c307245e4fcd61904a7b1336eeee2deb89cd3e71 — mutation-control runner
• b8019954460f27f4a02f0453ce3a6ba408b013cd — CI gate integration

Exact run:

• Actions run 32024523390
• job FRI verifier integrity
• reference conformance: SUCCESS
• harness integrity: SUCCESS
• verifier self-audit: SUCCESS
• mutation-kill control: SUCCESS

The mutation control deliberately reintroduces two known bad semantics:

1. FRI-2: remove the no-activity observability guard so 0 reads / 0 observations can be treated as healthy.
2. FRI-5: remove the missing-verification guard so None == None can authorize use.

Acceptance rule: the baseline audit must be green, and the corresponding audit controls (FRI-2C, FRI-5C) must fail under the mutated verifier. A surviving mutant makes the CI step fail closed.

Evidence artifact:

• artifact id 9286570928
• digest sha256:0d3e4d7d543a7ee06929ae37dfdcd3c9207cbf8a052d08e666ee2ee0f848c401

Scope boundary: this proves the self-audit kills these known semantic regressions; it is not a claim of complete mutation adequacy.

[safal207]

FRI verifier mutation matrix v0.2 is now green on exact head 749e1ca94f7d4584ac9a3ce2f5b59926e974c024.

GitHub Actions run: 32024746480

Observed:

reference conformance     6/6 PASS
harness integrity         4/4 PASS
verifier self-audit      22/22 PASS
mutation matrix            6/6 KILLED
mutation_score             1.0
required_score             1.0
coverage_complete          true
rule_coverage              FRI-1..FRI-6

The two new discriminating controls were necessary to make FRI-3 and FRI-4 mutation-sensitive:

• FRI-3E: same current owner identity, stale ownership epoch -> BLOCK_STALE_OWNER
• FRI-4D: completion receipt present while label is still running -> BLOCK_DEPENDENCY_NOT_COMPLETE

This matters because reachability alone was insufficient; the fixture set now contains cases where competing semantics produce different observable results.

Mutation artifact:

• artifact id 9286653311
• digest sha256:a6f79e6d56b30725efe0dfab3ef958a3c4b53e56aa40f2f7fabc7a4159778ddc

Boundary: this is six mandatory semantic mutants, one per FRI rule. It proves the current audit kills these known regression classes; it is not a claim of exhaustive mutation adequacy.

[safal207]

Automatic boundary mutation campaign is now part of the FRI verifier-integrity gate on current main.

Current exact head:
42ab261be4a4706debcd44f99a52154724c0c096

Exact run:
32025011777 — SUCCESS

New executable:
benchmarks/fri-agent-primitive-verification-v1.0/supplements/fri-reference-verifier-audit-v0.1/run_boundary_mutation_campaign.py

The campaign deliberately excludes top-level rule == ... dispatch comparisons so trivial routing breakage cannot inflate the mutation score.

Generated families and exact-head results:

boundary_compare  12/12 KILLED  score=1.0
guard_bypass      14/14 KILLED  score=1.0
guard_negate      14/14 KILLED  score=1.0
-------------------------------------------
aggregate          40/40 KILLED  score=1.0

Discovered semantic sites:

• 12 comparison sites
• 14 non-dispatch guard sites
• 40 generated single-site mutants total
• 0 survivors

The workflow requires --required-score 1.0, so any survivor now fails the FRI verifier-integrity job.

The same run also preserved:

• FRI reference conformance: 6/6 PASS
• harness integrity: 4/4 PASS
• verifier self-audit: 22/22 PASS
• manual semantic mutation matrix: 6/6 KILLED

Evidence artifact:

• artifact id 9286747179
• digest sha256:4e070597748ebe01dba2df6b61fa45dbc0aefd43f0c149f8255bcaa6b5fe46b6

Scope boundary: v0.1 covers comparison-boundary flips plus guard bypass/negation. Order-inversion mutants are intentionally not gated yet because adjacent guards can be semantically equivalent and need an equivalence-aware policy before a score is meaningful.

[safal207]

Order-inversion / equivalence-aware verifier gate is now live on exact head.

Head:
a2ba1d53a44150bd9483908a4eaca75072d7b297

Run:
32025451105 — SUCCESS

Reference layers:

• canonical FRI: 6/6 PASS
• harness integrity: 4/4 PASS
• verifier self-audit: 24/24 PASS
• manual semantic mutation matrix: 6/6 KILLED, score 1.0
• automatic boundary campaign: 40/40 KILLED, score 1.0

Order-inversion campaign:

• causal guard inversions: 6/6 KILLED, covering FRI-1…FRI-6
• declared pure reorderings: 3 EQUIVALENT
• SURVIVED: 0
• INVALID: 0
• scored mutation_score: 1.0

Two new intersection controls were required to make precedence observable:

• FRI-3F: no delivery + incomplete authority evidence -> BLOCK_NO_DELIVERY
• FRI-6E: no recovered state + incomplete lane evidence -> BLOCK_NO_RECOVERED_STATE

Equivalence policy:
no observed difference != equivalence.
A no-difference reorder is only excluded from score when it has an explicit pure-commutative equivalence basis; otherwise it is SURVIVED and fails CI.

Artifact:
9286903676

digest:
sha256:a4027ca951fc23a88939275fdec8031e914fa3b075fe3c9ee6350c7b76355270

Evidence boundary: this establishes verifier/harness adequacy for the declared reference semantics and mutation campaigns. It is not product-runtime certification.

[safal207]

FRI frontier moved one layer up into a separate compositional benchmark: #67 (PACC v0.7).

Reason for the split:

• FRI FRI Claude Code auto-memory adapter: FRI-1 / FRI-5 public-surface observability #63 verifies primitive truthfulness at a boundary;
• PACC PACC v0.7 — causal composition from authorization proof to observed outcome #67 verifies causal ordering across boundaries: proof -> current authority -> atomic consumption -> dispatch -> outcome.

Initial PACC exact-head evidence is already green on 505640b4607cb68e8c4f0a966ecb63fc47d02077: 6/6 composition fixtures PASS; 4/4 scored order mutants KILLED; 0 survived; mutation_score=1.0; artifact 9286996756, digest sha256:2c86599a7bcad55e615286c5e15c91a4a9a439e75b3d50b002d263d71c11f704.

This keeps the trust frontier explicit instead of silently expanding FRI's scope.