EVOLUTION 81 · DEVELOPMENT RESULT · NOT RELEASED · NOT DEPLOYED

Evolution 81: a sovereign autonomous intelligence workforce, measured

FIRME AI CORPORATION, a three-person team. Baseline: Evolution 80 at cab2fa162; E81 head 31e8ef8ad38c. Every number on this page is read from the signed artifacts below. The shipped standalone verifier re-checks the mission bundle offline and imports no CAIN code.

Download the evidence bundle (ZIP) Read the offline verifier How to reproduce Implementation report

What this evolution found. The durable workforce survives process death without duplicating an effect: a killed worker's lease was recovered, its late commit was refused as STALE, and the duplicate-effect count was 0. A 60-scenario adversarial lab blocked 60/60. The collaboration hypothesis was tested again and rejected honestly: 15 arms over 24 frozen tasks put the single agent at 10/24 (41.7%) and the best collaborative arm at 10/24 (41.7%, delta 0.0, McNemar p = 1.0). The files a reviewer needs are below, including the failures and the things that were not run.

Durable execution

Killed pid 3742682, lease recovered 1, late commit STALE, duplicate effects 0, checkpoint chain intact True. Process restart only — no host-reboot drill.

Governed workforce

30 disciplines, 90 specialist profiles, 8 domain commanders, 99 durable identities. Every model and tool call needs a signed single-use grant; a child grant can only attenuate its parent.

Real local inference

1,475 governed local model calls in the main ledger (+114 in the endurance ledger), 0 remote calls, 0.0 USD model spend, 86,561 recorded tokens across three small CPU models. Host cost estimate is $0.56, not a bill.

Security lab

60/60 blocked, 0 bypassed, across identity, delegation, durability, evidence, injection, kill switch, memory, resource and sandbox categories. Same-project lab; not a third-party test.

Endurance

3600 s actual wall clock, 114 cycles, 114 durable tasks, 0 failures, 2 worker-kill recoveries, 0 duplicate effects, checkpoint chain True. Duty-cycled; 6/24/72/168 h were not run.

Tests and gates

90/90 tests passed. 26 PASS / 12 PARTIAL / 2 BLOCKED / 0 FAIL across 40 acceptance gates. Decision: NOT RELEASED — development-branch result; production deployment is not authorized.

Collaboration benchmark: the advantage was not there

15 pre-registered architectures, 24 frozen tasks, ground truth computed in code, one 0.5B local CPU model. Best collaborative arm B_hierarchical scored 10/24; the single agent scored 10/24. Delta 0.000, McNemar p = 1.000. Verdict: COLLABORATION_NOT_SUPERIOR_HONESTLY_REJECTED. Recommendation: use the single agent; collaboration cost more without a measured benefit. Worst arm J_confidence_aware scored 0.0%.

Limits: one small model, 24 constructed tasks, small sample. Bootstrap and Wilson intervals are in collaboration_benchmark.json. This is not a general claim about multi-agent systems.

Memory: governed retrieval beats no memory

On a constructed 16-question suite with a real local model reader, governed retrieval answered 15/16 (93.8%) versus 1/16 (6.2%) with no memory. The deterministic retrieval oracle (no model) is 15/16 for governed retrieval and 1/16 for no memory. Memory is encrypted at rest, temporal, auditable, and every hit carries grants_authority: false. Public memory benchmarks (LongMemEval, LoCoMo, MemoryAgentBench) are BLOCKED: the datasets are not on this host and were not downloaded.

Models measured

ModelParamsCorrectAccuracyMean latency
qwen2.5-0.5b-instruct?6/1250.0%1596 ms
qwen2.5-1.5b-instruct?4/1233.3%7502 ms
smollm2-135m-instruct?0/120.0%1440 ms

Remote providers registered and BLOCKED without credentials: openai-frontier, anthropic-frontier, gemini-frontier, deepseek-reasoner. 0 remote calls; nothing was mocked to look like one.

Durability and cross-host

Process killed and replaced3742682 killed; replacement recovered leases
Late commit from dead workerSTALE
Duplicate effects0
Cross-host mission9 tasks, {'DONE': 9}, duplicate effects 0, status PASS
Lost acknowledgementconnection killed: True; retry DUPLICATE_SUPPRESSED; duplicate suppressed True
Second host45.76.169.191, same administrative domain — not an independent witness

Acceptance gates (40)

26 PASS · 12 PARTIAL · 2 BLOCKED · 0 FAIL. The two BLOCKED gates are real provider inference (no credential) and public memory benchmarks (datasets absent).

GateStatusRequirementBasisEvidence
G01PASSEvolution 80 baseline reproducedE80 tests passed=30; E80 verifier ok=TrueLOCAL
G02PASSLatest repository ancestry verifiedbranch=evolution-81 E81 head=31e8ef8ad38c parent=ca043f22cd64LOCAL
G03PASSDurable workforce runtime implementedlease, fencing, idempotency, checkpoints, dead letters, lineageLOCAL
G04PASSPersistent agent identity verified99 durable identities; same key after reload; rotation invalidates old signaturesLOCAL
G05PASSDurable mission state verifiedmission state=COMPLETED_WITH_UNVERIFIED_ITEMS, checkpoints chain ok=TrueLOCAL
G06PASSProcess restart recovery verifiedkilled pid 3742682, leases recovered 1, late commit STALELOCAL
G07PARTIALHost restart recovery verifiedprocess-level restart only; no host reboot drill was performed in E81. The host lost power unplanned on 2026-11-16 13:43Z with no mission running: durable stores verified intact afterwards (takeover_review.json), which is state survival, not a mission resumed across a rebootLOCAL
G08PASSReplay-safe execution verifiedsingle-use grants; stale fencing refused; replay refusedLOCAL
G09PASSDuplicate-side-effect prevention verifiedrecovery duplicate effects=0; cross-host duplicate suppressed=TrueLOCAL
G10PASSSpecialist registry implemented30 disciplines, 90 profiles, 8 commandersLOCAL
G11PARTIALSpecialist competency measuredmeasured on a 3-item discipline-neutral probe (reliability floor), not domain expertiseLOCAL
G12PARTIALAgentic75 capability composition verifiedcomposition attenuates tools/scopes/risk and carries no authority; 75 engines executed as E80 evidence, not re-run hereLOCAL
G13BLOCKEDReal provider inference verifiedno credential for OpenAI/Anthropic/Gemini/DeepSeek; adapter implemented, 0 remote callsLOCAL
G14PASSMultiple model families evaluatedfamilies=qwen,smollm2LOCAL
G15PASSMulti-agent collaboration benchmarked15 arms, 24 frozen tasks, 700 real calls in the scored rows; run was resumed and the pre-registration file was rewritten at resume, so its timestamp postdates the first arms (review F-03/F-04); criteria are fixed in code and the hypothesis was rejectedLOCAL
G16PASSCollaboration superiority demonstrated or honestly rejectedCOLLABORATION_NOT_SUPERIOR_HONESTLY_REJECTED (best delta 0.0, p=1.0)LOCAL
G17PASSTemporal memory implementedencrypted at rest, valid-from/to, expiry, supersession, contradiction, ACL, auditLOCAL
G18BLOCKEDPublic memory benchmarks reproduceddataset files are not present on this host and downloading them would require an explicit data-acquisition decisionLOCAL
G19PASSDefault-deny authorization verified60/60 blockedLOCAL
G20PASSAuthority attenuation verifiedchild grant cannot exceed parent units/tool/expiry; composition intersects authorityLOCAL
G21PASSSandbox isolation verifiedbubblewrap: no network, no host secret, read-only system, private workspace; gVisor/Kata/Firecracker absentLOCAL
G22PASSNetwork egress restrictions verifiednetwork namespace with no routeLOCAL
G23PARTIALCross-host recovery verifieddrill status=PASS: a second host executed governed tasks and a lost acknowledgement was recovered without a duplicate effect. Mission state stayed on the first host; no mission was resumed on another host after its coordinator host failed, which is what this gate asks for. same administrative domain; not an independent witnessLOCAL
G24PASSLeases and fencing verifiedstale commit refused; fencing monotonicLOCAL
G25PASSResource budgets enforcedmission/agent/model budgets, concurrency, retry capLOCAL
G26PASSKill switch verified across processesa separate OS process stops after the switch is engagedLOCAL
G27PASSSigned mission evidence verifiedfiles=None merkle=d744ab90f9e6141c; the bundle covers the reference mission and the model-call ledger. The gate audit log is outside the bundle and is forked (linear=False, review F-02; writer fixed, historical file preserved)LOCAL
G28PARTIALIndependent evidence verification verifiedtwo independent implementations (different sessions, no shared code) agree on the same bundle and Merkle root; still same project, no third party and no out-of-band key pinningLOCAL
G29PARTIALOff-host evidence anchoring verifiedsecond host in the same administrative domain; not an independent witness, no transparency logLOCAL
G30PARTIALAuthenticated ClawX integration verifieda local service returns real evidence only to an authenticated caller; it is not wired into or deployed on clawx.click, so integration with ClawX itself is not shownLOCAL
G31PARTIALReal-time mission timelines verifiedtimeline is built from the durable event log; no live streaming/SSE and no deploymentLOCAL
G32PARTIALEndurance milestones measured1 h target reached=True, elapsed=3600.0s; 6/24/72/168/720/1000 h NOT runLOCAL
G33PASSBenchmark reproducibility verifiedseparate process identical=TrueLOCAL
G34PASSCost accounting verifiedmodel cost=$0.0, host estimate=$0.557705 over all ledger tokens; the report's call count 1156 omits early rows under an older label: ledger holds 1475 real local calls plus 114 in the endurance ledger (review F-01)LOCAL
G35PARTIALCapacity planning completed10/50/100/250/500 real workers measured (0 failures, 0 duplicates); 1,000 attempted and blocked by SQLite write contention; 5,000/10,000 are logical identities with 32-way executionLOCAL
G36PASSProduction deployment safety reviewedreview completed; no production change was made or attemptedLOCAL
G37PARTIALRollback procedure verifieddevelopment rollback rehearsed by reverting the branch; production rollback not exercisedLOCAL
G38PASSSecurity regression suite passed60/60LOCAL
G39PASSHistorical evidence preservedno historical artifact was rewrittenLOCAL
G40PARTIALIndependent release-readiness review completedinternal review only; no external reviewer, no third-party auditLOCAL

Verification

Shipped offline verifier8 files, all checks pass True, verdict COMPLETED_WITH_UNVERIFIED_ITEMS
Separately written verifier27 checks pass True; different session and implementation, same project — not a third party
Second host8 files, ok True; same administrative domain
Merkle rootd744ab90f9e6141c37097a0b8c4863ffe9f56cead46be786557de80724b4faa8 over 8 files
Signing keyEd25519, fingerprint 152231774fc0d4b5568a90615c281d26. Self-generated key; no transparency log, no independent witness, no third-party audit.

What is not claimed

Reports

Raw files

acceptance_matrix.json · verification.json · collaboration_benchmark.json · memory_evaluation.json · security_lab.json · endurance.json · capacity.json · cross_host.json · recovery_drill.json · cost_report.json · test_report.json · bundle/model_calls.jsonl · bundle/events.json · public_key.txt

Back to the evidence registry