# Evolution 85 — Acceptance Matrix

**FIRME AI CORPORATION · CAIN-42 · Evolution 85** — branch `evolution-85`, baseline Evolution 84 (`2c419a042`). Development result. Not deployed. Not released. Built by the three-person team.

Computed by `cain/cain2027/e85/acceptance.py` from the artifact files; a missing artifact makes a gate NOT EVALUABLE.
A design document is not counted as an implementation, a simulation is not counted as a deployment, a local witness
is not counted as an independent organization, and registered identities are not counted as workers.

**20 PASS / 15 PARTIAL / 2 FAIL / 1 BLOCKED / 1 NOT EVALUABLE** — decision: **NOT RELEASED**

| Gate | Name | Status | Class | Evidence | Finding |
|---|---|---|---|---|---|
| G01 | E84 baseline reproducible | **PASS** | LOCAL | baseline.json | 6 E84 tests pass; clean-copy receipt verifies (exit 0), tampered copy exit 40; same project, not a third party |
| G02 | Existing governance reused without duplication | **PASS** | LOCAL | baseline.json, governed/summary.json | 0 changed lines under cain/receipt, cain/action_commit, cain/identity_federation, cain/mcp_proxy.py, scripts/evolution84, cain/cain2027/e81, cain/cain2027/e80 since the E84 commit; E85 receipts pass the unmodified E84 verifier (62 checks) |
| G03 | Intelligence architecture documented | **PASS** | LOCAL | docs/evolution-85/VERTICAL_AGI_ARCHITECTURE.md | component map with a status for each capability; a document, not an implementation |
| G04 | Model-agnostic reasoning interface | **PARTIAL** | LOCAL | eval/summary.json (providers, arms, routing) | one interface drove 3 local models (qwen2.5-0.5b-instruct, qwen2.5-1.5b-instruct, smollm2-135m-instruct) and a deterministic tool; router selects on measured accuracy; no remote provider was exercised (no credential), so provider-neutrality is shown for local models only |
| G05 | Persistent memory provenance | **PASS** | LOCAL | eval/learning_events.json, tests.json | 34 E85 tests incl. ledger: every item has source, evidence reference, status, creator; {'VERIFIED': 27} usable; signed hash-chained events |
| G06 | Verified knowledge updates | **PASS** | LOCAL | eval/strategy_history.json | 18 strategy decisions, each with a validation measurement and an approver different from the proposer; 1 promoted, 17 rejected; untrusted/injected/held-out content quarantined in tests |
| G07 | Cross-domain evaluation harness | **PASS** | LOCAL | eval/rows.jsonl, eval/summary.json | 100 held-out tasks, 10 families, 9 arms, code-computed truth with a second implementation; constructed tasks, not a public benchmark |
| G08 | Held-out generalization tests | **PASS** | LOCAL | eval/tasks.json, eval/summary.json | held-out seeds disjoint from train/validation (checked at run time and by the verifier); unseen-domain family: single 2/10, with transferred strategy 2/10. The gate is that the test exists and ran, not that transfer succeeded |
| G09 | Learning improvement measurements | **PARTIAL** | LOCAL | eval/summary.json (comparisons.B_learned) | held-out: single 24/100, with promoted strategies 25/100 (delta +0.010, exact McNemar p=1.0); measured, but not a statistically significant improvement |
| G10 | Single-agent baseline | **PASS** | LOCAL | eval/summary.json | A_single 24/100 = 0.24 (Wilson 95% [0.1669, 0.3323]); floor 0.3 |
| G11 | Multi-agent collaboration comparison | **PASS** | LOCAL | eval/summary.json | comparison ran with cost and latency; result: C_specialist_verifier 0.24 (delta +0.000, p=1.0, 2.58x tokens); C_hierarchical 0.26 (delta +0.020, p=0.831812, 4.12x tokens); D_selfconsistency3 0.26 (delta +0.020, p=0.774414, 3.03x tokens) |
| G12 | Mission planning and decomposition | **PARTIAL** | LOCAL | verticals/summary.json, eval/summary.json | three fixed mission templates executed end to end; a model-written plan arm was measured (C_hierarchical 0.26); no dynamic decomposition engine was built or evaluated in this evolution |
| G13 | Persistent workforce identities | **PASS** | LOCAL | swarm/identities.json | 10000 Ed25519 identities registered and stable across re-open; these are records, not workers |
| G14 | Real worker execution measurements | **FAIL** | LOCAL | swarm/stages.json | real OS worker processes: [10, 50, 100, 250, 500]; 0 duplicate effects; 1,000/2,500/5,000 processes not run; concurrent model inference measured separately = 1 |
| G15 | Long-running mission recovery | **PARTIAL** | LOCAL | endurance/endurance.json | 1 h actual wall clock (3600.0 s), 11 worker crashes recovered, 14400 tasks, 0 duplicate effects; 6/24/72 h and host reboot not run |
| G16 | Delegation attenuation | **PASS** | LOCAL | governed/lab.json | scope escalation, undelegated capability, wrong agent, expired and revoked chains refused before any decision; chain re-checked offline by the E84 verifier |
| G17 | Tenant isolation | **PASS** | LOCAL | clawx/clawx.json, tests.json | tenant A's evidence refused by tenant B's resource and enforcement point; evidence store cross-tenant reads refused (24 refusals, 0 leaks); one host |
| G18 | Live MCP enforcement integration | **PARTIAL** | LOCAL | mcp/integration.json | the unmodified MCP interceptor ran in a sandboxed dev instance in front of E85 execution: 9/9 expectations met. The hosted production proxy was not changed and does not emit these receipts |
| G19 | Effect-bound E84 receipts | **PASS** | LOCAL | governed/package_local | receipt binds action, delegation, policy, authorization, enforcement, execution and the resource's signed read-back; altering the observation digest is rejected (exit 40) |
| G20 | External resource observation | **PARTIAL** | LOCAL | governed/summary.json (remote_host), governed/package_remote | resource ran on a second host (machine 1e54e16d7b5b), key generated there, observation verified, everything removed afterwards; same administrative domain, not a third-party system; the kill switch is not attached at the remote resource |
| G21 | Standalone receipt verification | **PASS** | LOCAL | governed/package_local | E84 verifier (imports no CAIN code) run with python -I from a copy: 62 checks, 0 failed |
| G22 | Independent witness readiness | **PARTIAL** | SIMULATED | governed/package_local/receipt.json | checkpoints are cosigned by an operator-run witness (SIMULATED_LOCAL) with consistency checks; independent cosignatures: 0. No independently administered witness exists |
| G23 | Global kill switch | **PASS** | LOCAL | swarm/stages.json, endurance/endurance.json | GLOBAL kill stopped every worker in 5 stages (up to 500 processes) and at the end of the 1 h run; a new action afterwards is DENIED; the kill is a signed chained event |
| G24 | Tenant and swarm kill switch | **PASS** | LOCAL | governed/lab.json, tests.json | all ten scopes block their own target and nothing else (tests); TENANT, SWARM, MISSION, TOOL, CAPABILITY, ORG, AGENT, RESOURCE, DELEGATION kills refused new actions in the lab |
| G25 | Worker fencing and revocation | **PASS** | LOCAL | swarm/stages.json, governed/lab.json, endurance/endurance.json | agents revoked before start completed 0 tasks; pre-kill grants stay refused after reinstate (STALE_GRANT_EPOCH); 5 mid-run revocations all stopped their process; 0 duplicate effects |
| G26 | Kill-switch latency benchmark | **PARTIAL** | LOCAL | swarm/stages.json, governed/summary.json | measured: in-process denial p95 0.982 ms after engage; across processes, slowest worker stop 5373.96 ms from the kill request at up to 500 workers; latest in-flight commit 5215.49 ms after the engage transaction; one host, shared file, no network propagation measured. The 1000 ms target is met at <=100 workers, not at 500. |
| G27 | Multi-agent endurance benchmark | **PARTIAL** | LOCAL | endurance/endurance.json | 1 h completed (3600.0 s actual, 24 worker processes over the run, 120 governed receipts); 6 h, 24 h and 72 h not run; deterministic tasks, no model calls in the endurance run |
| G28 | ClawX evidence persistence | **PARTIAL** | LOCAL | clawx/clawx.json | 21 records (15 receipts) persisted idempotently in a local hash-chained store; this is not the deployed ClawX service |
| G29 | ClawX tenant-isolated access | **PARTIAL** | LOCAL | clawx/clawx.json | 3 tenants; 24 cross-tenant attempts refused, 0 leaks; public view carries digests only; local store, no network interface or deployment |
| G30 | Three vertical prototypes | **PASS** | LOCAL | verticals/*/result.json | evidence_pack: precision 1.0, recall 1.0 on 12 labelled items; fabrication_audit: precision 1.0, recall 0.7 on 20 labelled items; mcp_tool_review: precision 0.8889, recall 0.8 on 20 labelled items. Deterministic engines, fixtures written for this evolution; not real-world precision and not model reasoning |
| G31 | Customer-facing demonstration | **PARTIAL** | LOCAL | scripts/evolution85/run_verticals.py, docs/evolution-85/COMMERCIALIZATION_PLAN.md | a runnable demonstration and a customer explanation exist; no customer has seen it |
| G32 | Commercial readiness | **PARTIAL** | - | docs/evolution-85/COMMERCIALIZATION_PLAN.md | offer, pricing hypotheses, pilot proposal and onboarding checklist are written; no paid assessment, pilot or payment resulted from this evolution, so there is no market validation |
| G33 | Security and threat-model review | **PARTIAL** | LOCAL | governed/lab.json, docs/evolution-85/THREAT_MODEL.md | 33/33 adversarial scenarios refused; threat model written; all testing by the same project, no independent review |
| G34 | Reproducible clean-environment verification | **NOT EVALUABLE** | - | replication.json | replication.json missing: no clean clone on a second host was recorded |
| G35 | Documentation accuracy | **PASS** | LOCAL | scripts/evolution85/docs_e85.py, verify_e85_offline.py | numeric documents are generated from the artifact files; the offline verifier checks the headline numbers in the final report against the raw rows |
| G36 | Regression testing | **PASS** | LOCAL | tests.json | E85 34 passed; E84 6; E81 64; E80 30; 0 failed |
| G37 | Independent evidence review | **BLOCKED** | - | - | the offline verifier is written by the same project; no independent reviewer, witness or auditor has examined this evidence |
| G38 | Production release readiness | **FAIL** | - | this matrix | not ready: no remote model result, no independent witness or review, 1 h endurance only, the hosted MCP proxy is not integrated, nothing is deployed. Decision: NOT RELEASED |
| E84-G38 | E84 gap: kill switch and bypass path | **PARTIAL** | LOCAL | governed/lab.json | kill switch: implemented and enforced at decision, enforcement and resource (was missing). Bypass: the resource refuses any call without pinned enforcement evidence; a direct file write is DETECTED by the seal audit; a sandboxed worker cannot open the store. Not prevented: a same-user process outside a sandbox can still write the file |

No percentage is given: it would hide the FAIL and the BLOCKED gate.
