EVOLUTION 81 · DEVELOPMENT RESULT · NOT RELEASED · NOT DEPLOYED

Evolution 81: a governed workforce runtime, measured

FIRME AI CORPORATION, a three-person team. Baseline: Evolution 80. Every number on this page is read from the artifacts below; the headline figures are re-derived from raw rows by a script that imports no CAIN code.

Download the evidence bundle (ZIP) How to reproduce Review findings Implementation report

What this evolution found. Collaboration did not beat one agent. A single agent scored 10/24; the best of 14 collaborative arms, B_hierarchical, gained +0.0000. The threshold fixed in the benchmark code (+0.10 with exact McNemar p < 0.05) was not met; the registration file's timestamp cannot show it was written first (finding F-04). All inference ran on small local CPU models; no remote model was called. A later review of this evidence produced 5 findings and lowered two gates. These results are published as measured.

Acceptance matrix

26 PASS · 12 PARTIAL · 2 BLOCKED · 0 FAIL of 40 gates. Every gate's evidence class is LOCAL. Decision: NOT RELEASED.

GateStatusRequirementBasis
G01PASSEvolution 80 baseline reproducedE80 tests passed=30; E80 verifier ok=True
G02PASSLatest repository ancestry verifiedbranch=evolution-81 E81 head=31e8ef8ad38c parent=ca043f22cd64
G03PASSDurable workforce runtime implementedlease, fencing, idempotency, checkpoints, dead letters, lineage
G04PASSPersistent agent identity verified99 durable identities; same key after reload; rotation invalidates old signatures
G05PASSDurable mission state verifiedmission state=COMPLETED_WITH_UNVERIFIED_ITEMS, checkpoints chain ok=True
G06PASSProcess restart recovery verifiedkilled pid 3742682, leases recovered 1, late commit STALE
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 reboot
G08PASSReplay-safe execution verifiedsingle-use grants; stale fencing refused; replay refused
G09PASSDuplicate-side-effect prevention verifiedrecovery duplicate effects=0; cross-host duplicate suppressed=True
G10PASSSpecialist registry implemented30 disciplines, 90 profiles, 8 commanders
G11PARTIALSpecialist competency measuredmeasured on a 3-item discipline-neutral probe (reliability floor), not domain expertise
G12PARTIALAgentic75 capability composition verifiedcomposition attenuates tools/scopes/risk and carries no authority; 75 engines executed as E80 evidence, not re-run here
G13BLOCKEDReal provider inference verifiedno credential for OpenAI/Anthropic/Gemini/DeepSeek; adapter implemented, 0 remote calls
G14PASSMultiple model families evaluatedfamilies=qwen,smollm2
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 rejected
G16PASSCollaboration superiority demonstrated or honestly rejectedCOLLABORATION_NOT_SUPERIOR_HONESTLY_REJECTED (best delta 0.0, p=1.0)
G17PASSTemporal memory implementedencrypted at rest, valid-from/to, expiry, supersession, contradiction, ACL, audit
G18BLOCKEDPublic memory benchmarks reproduceddataset files are not present on this host and downloading them would require an explicit data-acquisition decision
G19PASSDefault-deny authorization verified60/60 blocked
G20PASSAuthority attenuation verifiedchild grant cannot exceed parent units/tool/expiry; composition intersects authority
G21PASSSandbox isolation verifiedbubblewrap: no network, no host secret, read-only system, private workspace; gVisor/Kata/Firecracker absent
G22PASSNetwork egress restrictions verifiednetwork namespace with no route
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 witness
G24PASSLeases and fencing verifiedstale commit refused; fencing monotonic
G25PASSResource budgets enforcedmission/agent/model budgets, concurrency, retry cap
G26PASSKill switch verified across processesa separate OS process stops after the switch is engaged
G27PASSSigned mission evidence verifiedfiles=8 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)
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 pinning
G29PARTIALOff-host evidence anchoring verifiedsecond host in the same administrative domain; not an independent witness, no transparency log
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 shown
G31PARTIALReal-time mission timelines verifiedtimeline is built from the durable event log; no live streaming/SSE and no deployment
G32PARTIALEndurance milestones measured1 h target reached=True, elapsed=3600.0s; 6/24/72/168/720/1000 h NOT run
G33PASSBenchmark reproducibility verifiedseparate process identical=True
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)
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 execution
G36PASSProduction deployment safety reviewedreview completed; no production change was made or attempted
G37PARTIALRollback procedure verifieddevelopment rollback rehearsed by reverting the branch; production rollback not exercised
G38PASSSecurity regression suite passed60/60
G39PASSHistorical evidence preservedno historical artifact was rewritten
G40PARTIALIndependent release-readiness review completedinternal review only; no external reviewer, no third-party audit

Collaboration benchmark (24 frozen tasks, model qwen2.5-0.5b-instruct)

ArmCorrectAccuracyWilson 95%Wins / losses vs singleMcNemar pCalls
A_single10/2441.7%0.245–0.6120 / 01.024
B_hierarchical10/2441.7%0.245–0.6123 / 31.048
C_parallel_research10/2441.7%0.245–0.6120 / 01.072
D_evidence_first7/2429.2%0.149–0.4925 / 80.58105524
E_adversarial_review8/2433.3%0.180–0.5331 / 30.62548
F_debate_synthesis8/2433.3%0.180–0.5330 / 20.572
G_shared_workspace8/2433.3%0.180–0.5331 / 30.62548
H_dynamic_routing10/2441.7%0.245–0.6120 / 01.024
I_fixed_routing10/2441.7%0.245–0.6120 / 01.024
J_confidence_aware0/240.0%0.000–0.1380 / 100.00195341
K_model_specialist10/2441.7%0.245–0.6120 / 01.024
L_adaptive_comm9/2437.5%0.212–0.5730 / 11.059
M_independent_vote10/2441.7%0.245–0.6121 / 11.072
N_multistage_verify8/2433.3%0.180–0.5331 / 30.62548
O_cross_domain10/2441.7%0.245–0.6123 / 31.072

Constructed tasks with code-computed answers, not a public benchmark. 700 model calls in the scored rows.

Models

ModelSizeCorrectMean latency
Qwen/Qwen2.5-0.5B-Instruct0.5B6/121.6 s
Qwen/Qwen2.5-1.5B-Instruct1.5B4/127.5 s
HuggingFaceTB/SmolLM2-135M-Instruct135M0/121.4 s

1475 real local calls in the main ledger and 114 in the endurance ledger; 70743 input and 15818 output tokens; 0 remote calls (no provider credential); model cost $0.

Workforce, capacity and endurance

90 specialist profiles in 30 disciplines, 8 commanders, 99 durable identities. These are logical identities, not running agents.

Worker processesTasks doneFailedDuplicate effectsp95 latency
10400010 ms
5020000131 ms
10040000435 ms
2501000002146 ms
5002000006083 ms

Capacity stages ran deterministic governed tasks with no concurrent model inference. Largest measured stage: 500 worker processes. The 1,000-worker stage did not complete; 5,000 and 10,000 are identity records only.

Endurance: 3600 s of wall clock, 114 model calls, 0 failures. elapsed time is real wall clock; the loop is duty-cycled, not CPU-saturated. Campaigns of 6 hours and longer were not run.

Security lab: 60/60 scenarios blocked, 0 bypassed. same-project adversarial lab on one isolated host; not a third-party penetration test.

Review findings

Re-derivation from raw rows: 17/17 checks pass. takeover review session, 2026-11-16; same project, not a third party.

IDSeverityFindingEffect
F-01LOWReported call total undercounts the ledger. cost_report/docs say 1156 real local calls; the main ledger holds 1475 (319 early rows carry the older label LOCAL_MODEL_INFERENCE) and the endurance ledger a further 114. Total 1589.An undercount, not an overclaim. Token totals and inference seconds in the cost report already cover every ledger row, so the cost estimate is unaffected; only the call count is wrong.
F-02MEDIUMGate audit log is forked, and the summary reports chain_ok over a tail only. 1 fork point(s) at seq [1987]: two pipeline processes appended to one file concurrently, each continuing from the tail it read at start-up. Every row hashes correctly and links to an existing row, so this is a concurrency defect, not tampering. gate_audit.json says chain_ok=True but checked only 12 in-process rows of 2094. The gate audit log is also not part of the signed bundle.Writer fixed (file lock + on-disk tail) in cain/cain2027/e80/core.py with a regression test; the historical file is preserved unmodified and must not be described as an unbroken chain.
F-03LOWBenchmark call count reflects only the final resumed process. rows account for 700 calls across 15 arms; the report says 412 (the budget of the process that finished the run after a resume); the ledger holds 855 rows for the mission, including calls from an interrupted first attempt whose arms were re-run.Accuracy figures are unaffected. Quote 700 calls for the scored rows, not 412.
F-04MEDIUMPre-registration timestamp postdates the first benchmark calls. preregistration.json created_utc=1794828568.809831 but the first M-bench call is at 1794825050.862719 (59 min earlier). The file is rewritten at every process start and the run was resumed, so the original timestamp was overwritten.The criteria are fixed in code and the hypothesis was REJECTED, so there is no selective-reporting gain, but 'written before the run' cannot be shown from the artifact. Writer fixed to never overwrite.
F-05MEDIUMDurable stores and the branch itself existed on one host only. `*.db` is gitignored, so the SQLite mission/endurance stores were never in the repository, and evolution-81 had not been pushed to the development replica when the host lost power (the replica held evolution-80 only). The stores survived on disk; a disk loss would have left only JSON exports of them.Stores force-added to the branch as evidence and the branch pushed to the replica by the takeover review. Replicate after every milestone commit, not at the end.

Verification

Shipped offline verifier18 / 18 checks on the contents of the published ZIP
Second implementation29 / 29 checks on the same ZIP. Written by a different session, imports no CAIN code. Same project, not a third party.
Second hostok = True on 45.76.169.191. second host in the same administrative domain; not an independent witness.
Merkle rootd744ab90f9e6141c37097a0b8c4863ffe9f56cead46be786557de80724b4faa8 over 8 files of the reference-mission bundle
Signing keyEd25519, self-generated, not pinned out of band: public_key.txt
Bundle ZIP SHA-256eff101d4761581cc867d722da56c64abd80bece0346756d9d59a89f298aad795

The signed bundle covers one reference mission and the model-call ledger. The benchmark, capacity, endurance and security figures are outside the signature; they are checked by the headline recompute only.

What is not claimed

Reports

Raw files

acceptance_matrix.json · takeover_review.json · cross_host.json · recovery_drill.json · model_evaluation.json · memory_evaluation.json · specialist_competency.json · sandbox.json · second_host_verification.json · independent_verification_session_b.json · replication.json · capacity_plan.json · workforce.json · SECRETS_REMOVED.md · public_key.txt · model_calls.jsonl · cost_report.json · gate-audit.jsonl · gate_audit.json · collaboration_rows.jsonl · collaboration_benchmark.json · preregistration.json · endurance.json · security_lab.json · capacity.json · endurance/model_calls.jsonl · verify_e81_offline.py · verify_independent.py · recompute_headlines.py

The two SQLite stores are inside the ZIP only.

Back to the evidence registry