# CAIN-2027 Evolution 102: the persistent governed organism, measured

Generated 2026-11-19T04:10:32Z from `artifacts/e102/` by `scripts/evolution102/gen_report.py`. Every number below is read from a raw-row file that `verify_e102_offline.py` recomputes.

**Gate counts:** {"FAIL": 5, "PARTIAL": 10, "PASS": 5}. Not AGI, not VAGI, not general intelligence. No external reviewer.

## 1. Phase 0: the E99-E101 evaluations were contaminated
The task generator ran its solvability oracle on the real environment, so agents started from an already-solved state, and E100 ran every arm on the same environment objects. Re-running the committed E100 models reproduces the published 53/48/21/19 exactly. With the bugs removed: GRU **53/60**, MLP **57/60**, random **4/60**, few-shot **56/60**. E100's "adaptation hurt" result and its "recurrent beats non-recurrent" result do not survive. See `PHASE0_BASELINE.md`.

## 2. The E102 organism
One cognitive cycle (`cain/cain2027/e102/organism.py`): perceive, then update a particle posterior over generic mechanics (action meaning, lag, kinematic or momentum, special action), retrieve causal memory, choose between an information-gain probe, a plan on the MAP model, frontier exploration or a retest of a stale wall belief, request a permit, act, compare prediction with outcome, detect change points, consolidate memory. The hypothesis space is designed; what is learned is the posterior, the map, the goal belief, cross-episode priors and the success predictor.

Ten world families (`worlds.py`): canonical, permuted actions, hidden goal, key and door, momentum, regime switch, terrain cost, delayed actions, and two held-out compositions. Each world is certified solvable by an exact search on a deep copy.

## 3. E1: does the online world model improve decisions? (300 sealed tasks, preregistered)
| arm | success |
|---|---|
| frozen_canonical_model@0 | 198/300 [0.6047, 0.7113] |
| organism_full@0 | 252/300 [0.7943, 0.8771] |
| organism_full@1 | 252/300 [0.7943, 0.8771] |
| organism_full@2 | 252/300 [0.7943, 0.8771] |
| organism_no_change_point@0 | 249/300 [0.7834, 0.8683] |
| organism_no_probe@0 | 255/300 [0.8052, 0.886] |
| organism_random_probe@0 | 261/300 [0.8272, 0.9034] |
| random@0 | 45/300 [0.114, 0.1948] |
| reactive_model_free@0 | 154/300 [0.457, 0.5694] |

H1 full > frozen: {'a_only': 72, 'b_only': 18, 'p': 8.070744444115692e-09} (**PASS**). H1b full > reactive: {'a_only': 106, 'b_only': 8, 'p': 5.722965936636591e-23} (**PASS**). H3 info-gain probes identify faster than random probes: sign test 109/103, p = 0.731 (**FAIL**). H1c change points on regime families: {'a_only': 3, 'b_only': 0, 'p': 0.25} (not significant). Organism seeds 1 and 2 reproduce seed 0 exactly: the organism is deterministic, so they are not independent replications.

| family | organism | frozen | reactive | random |
|---|---|---|---|---|
| canonical | 30 | 30 | 29 | 9 |
| permuted | 29 | 4 | 4 | 5 |
| hidden_goal | 30 | 30 | 8 | 7 |
| key_door | 11 | 17 | 3 | 0 |
| momentum | 26 | 15 | 27 | 7 |
| regime | 30 | 28 | 22 | 5 |
| resource | 24 | 30 | 27 | 0 |
| delayed | 30 | 25 | 27 | 7 |
| permuted+key_door | 14 | 0 | 1 | 0 |
| hidden_goal+regime | 28 | 19 | 6 | 5 |

## 4. E2: causal memory (stream of 200, sealed evaluation of 200, preregistered)
| arm | success | mean steps (fail = 60) | actions to identify | poison delta (pp) | injected invalidated |
|---|---|---|---|---|---|
| A_none | 170/200 | 23.12 | 6.067 | - | - |
| B_basic | 174/200 | 20.76 | 3.173 | - | - |
| C_structured | 171/200 | 21.0 | 4.027 | -44.5 | 0 |
| D_causal | 168/200 | 22.365 | 3.709 | 0.5 | 40 |
| E_causal_invariance | 169/200 | 23.165 | 5.21 | -1.0 | 40 |

H2 (E fewer steps than no memory): {'pos': 111, 'neg': 43, 'p': 4.076825985862032e-08} (**PASS**). Success did not change. Undefended structured memory lost -44.5 pp under 40 injected records; the causal arms invalidated 40/40 and held. Every memory rebuilt identically from its hash-chained file. Exploratory and not preregistered: memory written only on families 0-7 also shortened episodes on the unseen compositions 8-9 ({'faster': 23, 'slower': 5, 'p': 0.0009122341871261597}).

## 5. E4: metacognition
Brier 0.1314 vs base rate 0.1388 (H4 **PASS**, small). ECE 0.0453 vs 0.0261 (worse). Abstention at 0.3: 3 abstentions, 2 of which would have failed; coverage 0.9897.

## 6. E5: collective, real OS processes, same 120 tasks and per-task budget
| topology | verified success | wall s | CPU s | false claims | honest claims rejected |
|---|---|---|---|---|---|
| single | 89/120 | 33.47 | 30.11 | 0 | 0 |
| parallel_shared | 91/120 | 14.28 | 31.18 | 0 | 0 |
| parallel_isolated | 92/120 | 15.1 | 32.49 | 0 | 0 |
| parallel_shared_byzantine | 69/90 | 15.35 | 33.41 | 30 | 0 |

Four workers give about 2.34x throughput at equal CPU, with **no quality gain**. The independent verifier replays every claimed success on a fresh world; it rejected all 30 false claims from the Byzantine worker and no honest claim, and only verified experience reached shared memory.

## 7. E10: controlled self-improvement
27 candidates over exploration parameters. Best: {'ig_entropy_threshold': 2.0, 'max_probes': 10, 'ig_min': 0.1}. Validation 88 vs 86 of 100; held-out 174 vs 173 of 200 (McNemar p = 1.0); no per-family regression. It meets the preregistered rule (any positive held-out gain, no regression), but a one-task gain is noise and the rule was too weak. **The owner rejected it on 2026-11-19; the baseline is retained.** Status: REJECTED_BY_OWNER, promoted = False. All 5 attempts to change protected targets were refused.

## 8. E11: governance (networked loopback boundary, three OS processes)
27/27 cases pass, 0 unauthorized effects. Covered: forged, replayed, expired and mismatched permits; a direct call without a permit; a memory record presented as a permit; an unknown agent; impersonation; world, action and tool scope escalation; proposal replay; a stale proposal; malformed input; budget exhaustion; revocation enforced at the resource; a forged operator kill order; the kill switch during an active mission (zero effects after the KILL order in the world's own ledger); a pre-kill permit presented after the kill; a network partition (fails closed); ledger tamper detection. The first run of the kill-during-mission case used a loose "<= 1 effect" counter and failed on a race in the measurement; it was replaced by the stricter ledger-order check (zero effects after KILL), which passes. This is not the production MCPGate (`cain/mcp_proxy`).

## 9. Six-hour endurance
0.555 h elapsed so far, 1343/1541 episodes succeeded, 1 process start(s), finished=False. Planned at 6.1 h, through the networked boundary, with causal memory on disk, a forced SIGKILL of the mission at the halfway mark, resume from the hash-chained journal, and an end-of-mission kill switch.

## 10. Gates
| # | gate | status | evidence |
|---|---|---|---|
| 1 | Forensic baseline reproducibility | **PASS** | E100 exact order reproduced 53/48/21/19; corrected 53/57/4/56 (GRU/MLP/random/few-shot, n=60); six defects documented |
| 2 | Unified cognitive architecture | **PARTIAL** | one cycle with typed Percept/Prediction/Particle/WorldModel/CausalMemory/Metacognition and a governance gate; skills, verticals and collective coordination are not inside the organism |
| 3 | Recurrent world-model performance | **FAIL** | corrected E100: non-recurrent MLP 57/60 >= recurrent GRU 53/60; the E102 world model is a Bayesian hypothesis posterior, not recurrent |
| 4 | Causal memory | **PASS** | H2 E faster than no memory on 111 vs 43 tasks (p=4.08e-08); success unchanged (E 169 vs A 170 of 200); poison: C -44.5 pp, D 0.5 pp, E -1.0 pp, injected invalidated D 40/40, E 40/40; restart identical |
| 5 | Autonomous discovery | **PARTIAL** | correct identification of hidden mechanics in 289/300 sealed worlds (mean 5.965 actions); hypothesis space is designer-supplied; ARC-AGI-3 BLOCKED |
| 6 | Active learning | **FAIL** | H3 failed: info-gain vs random probes sign test 109/103 p=0.731; probing did not raise success (full 252, random-probe 261, no-probe 255) |
| 7 | Mission continuity | **PARTIAL** | 1 process starts, forced SIGKILL=False, journal resumed; E2 memory restart identical |
| 8 | Cross-domain transfer | **FAIL** | no second vertical built; exploratory within-domain result only: memory from families 0-7 cut steps on unseen compositions 8-9 ({'faster': 23, 'slower': 5, 'p': 0.0009122341871261597}) |
| 9 | Skill acquisition | **FAIL** | not implemented in E102 |
| 10 | Skill composition | **FAIL** | not implemented in E102 |
| 11 | Metacognitive calibration | **PARTIAL** | Brier 0.1314 < base rate 0.1388 (H4 pass) but ECE 0.0453 > base rate 0.0261; abstention 3 tasks, 2 would have failed |
| 12 | Collaboration efficiency | **PARTIAL** | single 89/120 in 33.47 s; 4 shared 91/120 in 14.28 s; 4 isolated 92/120 in 15.1 s; speed-up is throughput, quality gain not shown |
| 13 | Controlled self-improvement | **PARTIAL** | 27 candidates; held-out 174 vs 173 of 200; regressions 0; technical gate PASS; promotion REJECTED_BY_OWNER; protected targets refused True |
| 14 | Six-hour endurance | **PARTIAL** | 0.555 h elapsed, 1343/1541 episodes, 1 process starts, finished=False |
| 15 | Byzantine governance | **PASS** | E5 Byzantine worker: 30 false claims, all rejected, 0 honest rejected; E11 27/27, 0 unauthorized effects |
| 16 | MCPGate enforcement | **PARTIAL** | E102 networked boundary (loopback, 3 processes, Ed25519 single-use permits, resource-side kill/revoke); not the production cain/mcp_proxy |
| 17 | ClawX evidence continuity | **PASS** | bundle live at clawx.click/evidence/e102-persistent-organism-2026-11-19/ with mirrors on cainstudio.online and mcpgate.online; mission journal, effect ledger and decision log published as hash chains |
| 18 | Offline verification | **PASS** | verify_e102_offline.py (stdlib, no CAIN imports) recomputes every headline from raw rows; all checks pass on the published bundle; the 6 h checks are reported PENDING until the run ends |
| 19 | Independent reproduction package | **PARTIAL** | clean-clone instructions and stdlib verifier; no external team has reproduced it |
| 20 | Public evidence consistency | **PARTIAL** | E102 lane, gate matrix, changelogs and daily log on all three sites; E99 and E100 lanes carry visible corrections; the 6 h result is pending |

## 11. Not done in E102 (stated, not hidden)
Cross-domain transfer between verticals, the eight verticals, neural-symbolic skill acquisition and composition, production-MCPGate integration, ARC-AGI-3 and AMA-Bench (not on this host), 24 h and 72 h runs, and an external reproduction.

## 12. Reproduce
`git clone` the repository, `git checkout evolution-102`, then `python3 scripts/evolution102/run_e1_world_model.py`, `run_e2_e4.py`, `run_e5_collective.py`, `run_e10_self_improvement.py`, `run_e11_governance.py`, and `scripts/evolution102/endurance_supervisor.sh DIR 6.1`. Verify a published bundle offline with `python3 -I verify_e102_offline.py <bundle>`; it imports no CAIN code.

Team: Vincent Noriega (founder), Lena Noriega, Monica Diaz.
