EVOLUTION 80 · DEVELOPMENT RESULT · NOT RELEASED · NOT DEPLOYED

Evolution 80: adaptive routing and multi-agent coordination, measured

FIRME AI CORPORATION, a three-person team. Baseline: Evolution 79 at f4449aa7913623bc034fa4434ca96c2a2919a889. Every number on this page is read from the signed artifacts below, and the bundle's verifier recomputes them from raw per-task rows.

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

What this evolution found. The two hypotheses it was built to test were not supported. The adaptive router scored 22/48 against 19/48 for the best fixed path (exact McNemar p = 0.647606). No multi-agent architecture significantly beat one agent; the best, c08_evidence_first, gained +0.1042. The routing direction repeated on 24 replication tasks (10 vs 8), without significance. These results are published as measured.

Release gate

13 PASS · 10 PARTIAL · 1 FAIL · 5 BLOCKED · 3 NOT EVALUABLE (32 gates). NOT RELEASED. Evolution 80 is a development-branch result. Production deployment is not authorized.

GateStatusBasis
01 Baseline reproduction (E79 144/144, clean room, second host, 25 tests)PASSdocs/evolution-80/EVOLUTION79_REPRODUCTION.md
02 Public evidence reconciliationPARTIALprovenance established and branch files corrected; live sites unchanged (deployment not authorized)
03 Real local model inference, governed and signedPASS1154 calls, 153132 tokens, 0 cloud
04 Adaptive routing beats the best fixed path (H1)PARTIALrouter 22/48 vs m15_direct 19/48, p=0.647606
05 Collaboration advantage (H2)FAILbest multi-agent arm c08_evidence_first delta 0.1042; significant arms: []
06 System recognises when collaboration hurtsPASSarms measured worse than one agent: ['c07_specialist_committee']; the adaptive graph communicates only on disagreement
07 Replication in a separate process (H3)PASS{"a_correct": 8, "b_correct": 10, "n": 24, "diff": 0.0833}
08 Heterogeneous model registryPARTIAL3 local CPU models measured; larger and other-family models NOT EVALUATED
09 Remote model providersBLOCKEDno credential or funding; purchase prohibited; never mocked
10 Specialist capability profilesPARTIAL10 of 20 measured; 10 NOT EVALUATED
11 Skill acquisition (LoRA) with held-out, transfer and regression checksPASSheld-out 0.0 -> 1.0; transfer delta 0.125; NO_REGRESSION; promotion AWAITING_HUMAN_APPROVAL
12 Memory answer accuracy (constructed set)PASSanswer accuracy governed_retrieval 1.0, naive_recent 0.2143, no_memory 0.2143, ungoverned_retrieval 0.7143
13 Public memory benchmarks (LongMemEval, LoCoMo, MemoryAgentBench, AMA-Bench)NOT EVALUABLEnot run
14 Autonomous research project end to endPASSpre-registered, frozen policy, raw rows, separate-process replication; local fixture-free literature review of abstracts only
15 Self-improvement candidate evaluated, not self-promotedPASSbenefit demonstrated: False; state AWAITING_HUMAN_APPROVAL
16 Agentic 75 audit and compositionPARTIAL75/75 engines executed on bundled samples; 0 model-backed; 0 independently evaluated; 4 composed in a mission
17 Durable execution under injected failures (one host)PASSinjected 8/8; duplicates 0; stale commits refused 2
18 Cross-host scheduling, checkpoint replication, network partitionNOT EVALUABLEnot run in E80; E72/E78 evidence preserved
19 Security adversarial labPASS51/51 blocked; same-project lab, bubblewrap only
20 Kill switchPARTIALin-process gate refuses decisions and pre-issued grants; no distributed halt tested
21 Stronger isolation (gVisor, Kata, Firecracker, Kubernetes Agent Sandbox)NOT EVALUABLEnot installed
22 Mission controller with evidence-based completionPASSCOMPLETED_WITH_UNVERIFIED_ITEMS_EXCLUDED; nodes {'VERIFICATION_FAILED': 4, 'VERIFIED': 9}; verified-but-wrong 0
23 ClawX command-center backendPARTIAL20/20 views backed by artifacts; local token-gated endpoint; not deployed; no live process view
24 MCPGate enforcementPARTIALreference gate enforces single-use grants on every model and tool call and passes bypass tests; the production MCP proxy was not modified or re-tested
25 CAIN Studio unified public explanationPARTIALarchitecture and status documented; no site change deployed
26 Endurance 1 hPASS3600 s actual elapsed in one process; 80 cycles; 640 governed tasks (560 local model calls, 80 sandbox executions); 0 failures; 0 duplicate task ids; checkpoint chain intact; duty-cycled (44% busy); accuracy 252/640 is a small-model score, not a quality claim
27 Endurance 6 h / 24 h / 72 h / 168 hBLOCKEDnot run
28 Signed evidence with offline verifierPASSsee evidence.json; self-generated key
29 Second-environment verificationPARTIAL166/166 checks on 45.76.169.191 (development replica); same administrative domain, not an independent witness
30 Transparency log / independent witnessBLOCKEDpublication not authorized; no Rekor entry; no independent witness
31 Third-party auditBLOCKEDnone commissioned
32 Production deploymentBLOCKEDprohibited for Evolution 80

Test split (48 unseen tasks)

ArmCorrectAccuracyModel calls
c01_single11/4822.9%48
c02_parallel_vote11/4822.9%144
c03_pairwise_debate11/4822.9%139
c04_critic_reviser11/4822.9%96
c05_planner_executor13/4827.1%96
c06_hierarchical15/4831.2%181
c07_specialist_committee10/4820.8%144
c08_evidence_first16/4833.3%96
c09_parallel_synthesis13/4827.1%192
c10_adaptive_graph22/4845.8%60
det_exec6/612.5%0
fixed_pot05_else_direct0516/4833.3%48
fixed_pot15_else_direct1527/4856.2%48
m05_direct11/4822.9%48
m05_pot5/2910.4%29
m15_direct19/4839.6%48
m15_pot13/2927.1%29
oracle_any_path32/4866.7%42
router22/4845.8%42

1154 real local CPU model calls, 153132 tokens, 0 cloud calls, $0.00. Three small models (Qwen2.5 0.5B and 1.5B, SmolLM2-135M). Constructed closed-form tasks: unseen instances of known templates.

Endurance

PASS: 3600 s actual elapsed in one process; 80 cycles; 640 governed tasks (560 local model calls, 80 sandbox executions); 0 failures; 0 duplicate task ids; checkpoint chain intact; duty-cycled (44% busy); accuracy 252/640 is a small-model score, not a quality claim. ACTUAL elapsed wall clock; no virtual or accelerated time. Campaigns of 6 hours and longer were not run.

Verification

Shipped offline verifier166 / 166 checks on the build host
Second host166 / 166 on 45.76.169.191 (development replica). same administrative domain as the build host; not an independent witness.
Separately authored verifier52 / 52, verdict PASS. Written by a different session, imports no CAIN code; it was executed by the session that finished the evolution, on the same host. report
Merkle root1ae86e51063095240c42f436eda8787edbbe8a49539c4ed55da584a20093bc2a over 117 artifacts
Signing keyEd25519, fingerprint 4acfa1815616155b7f5ebc1b70c9501fa189d7a9e2e5ffc934f0422c5f1f1e33. Self-generated key. The fingerprint has not been published out of band. No transparency-log entry. No independent witness.
Bundle ZIP SHA-25681828ba427869e0299056c556cc9276b97f23f437ef3015ceda33455e32d4e3d

Run lineage

Three benchmark runs were made on 2026-11-15. Runs 1 and 2 are preserved in the repository and superseded; only run 3 is evidence. Run 3 was interrupted when its session ended and was resumed with the unchanged command; no stage ran twice. Record: run_lineage.json, explanation: EVIDENCE_AND_REPRODUCTION.md.

What is not claimed

Reports

Raw files

evidence.json · release_gate.json · collective_benchmark.json · preregistration.json · endurance.json · offline_verification.json · second_host_verification.json · canonical_evidence_index.json · public_key.txt · bundle/model_calls.jsonl · bundle/gate_audit.jsonl

Back to the evidence registry