# CAIN-42 EVOLUTION 142 — OPEN-WORLD DISCOVERY ENGINE · FINAL REPORT

**FIRME AI CORPORATION · CAIN-42 · E142 · 2026-11-22**

- **AGI: NOT ESTABLISHED** · **RSI: EXPERIMENTAL** · **NEURAL WEIGHTS MODIFIED: NO** · **INDEPENDENT REPRODUCTION: 0** · **OFFICIAL BENCHMARKS RUN IN E142: 0** · **PAID API SPEND: $0.00** (0 model calls, 0 tokens)
- **Release gates (39):** 21 PASS / 11 PARTIAL / 5 FAIL / 2 BLOCKED
- **Evidence bundle:** https://clawx.click/evidence/e142-open-world-discovery-2026-11-22/ · offline verifier `verify_e142_offline.py`: 16/16 checks passed; 0 failed
- **Merkle root:** `63897562b21b164386ea778a022558504cc48cba556ba0cffc6c6c530b6a193f` · evidence key `0b00fe53501740b3…` (self-generated: integrity, not identity)
- **Preregistrations:** H1–H4 `025499a15` (committed before code) · OW1–OW7 `f77d2db2f` (committed before campaign data)
- **Analysis cutoffs (stopping rules):** H1–H4 2026-11-22T03:29:02Z (all regimes reached the preregistered count) · OW1–OW7 2026-11-22T04:41:06Z (all regimes reached the preregistered count)

## 1. What E142 is

A durable open-world research loop on the E141 engine: ten constructed environment families (grid navigation, toggle puzzle, causal switchboard, 1-D physics, crafting, ARC-like 3×3 transforms, modular arithmetic, stack machine, lever chain with delayed effects, two-agent coordination) × ten curriculum levels (new configuration, new combination, unseen primitive, hidden predicate goal with unsolvable cases, masked feature, rule change mid-run, larger instance, two-stage goal, permuted interface, random mix). A generic agent sees only `observe / step / reset / n_actions` and a step budget; it builds a world model (fact table + induced per-action feature rules + tested revision), explores with one of eight strategies, infers the hidden goal, and abstains when it cannot reach one. The evaluator scores the **exported model** against the true dynamics and executes the exported planner in the true environment; nothing the worker says about its own score is used. Code: `cain/cain2027/unified/e142/`.

## 2. Confirmatory results (preregistered; unit = one experiment; one-sided sign test; Holm)

| ID | Question | Experiments | W/L/T | Mean diff [95% bootstrap CI] | d_z | p | Holm | Verdict |
|---|---|---|---|---|---|---|---|---|
| OW1 | Exploration: model-based frontier vs random walk | 257 | 166/74/17 | +0.0367 [+0.0271, +0.0467] | 0.45 | 1.37e-09 | yes | **SUPPORTED** |
| OW2 | World model: rules + tested revision vs fact table | 316 | 257/48/11 | +0.0588 [+0.0507, +0.0668] | 0.81 | 5.38e-36 | yes | **SUPPORTED** |
| OW3 | Transfer: evidence-gated prior vs scratch (5% budget) | 77 | 26/19/32 | +0.0072 [+0.0015, +0.0133] | 0.27 | 0.186 | no | **NOT SUPPORTED** |
| OW3b | Transfer: gated vs ungated prior | 185 | 115/66/4 | +0.0442 [+0.0272, +0.0620] | 0.37 | 0.000167 | yes | **SUPPORTED** |
| OW4 | Hidden goals: E142 agent vs E141-style agent | 318 | 129/74/115 | +0.0739 [+0.0509, +0.0981] | 0.34 | 6.91e-05 | yes | **SUPPORTED** |
| OW5 | RSI: lineage final config vs G0 (sealed) | 20 | 11/2/7 | +0.0316 [+0.0076, +0.0560] | 0.56 | 0.0112 | no | **NOT SUPPORTED** |
| OW5b | RSI: lineage vs compute-matched random search | 20 | 9/7/4 | -0.0023 [-0.0214, +0.0158] | -0.05 | 0.402 | no | **NOT SUPPORTED** |
| OW6 | RSI: research memory vs no memory | 20 | 9/7/4 | +0.0036 [-0.0133, +0.0203] | 0.09 | 0.402 | no | **NOT SUPPORTED** |
| OW7 | RSI: inherited vs reset proposer | 20 | 2/3/15 | -0.0080 [-0.0230, +0.0034] | -0.25 | 0.813 | no | **NOT SUPPORTED (treatment worse)** |
| H2 | Calibration: importance-weighted vs plain temperature (pure shift) | 44 | 6/38/0 | -0.0058 [-0.0073, -0.0043] | -1.09 | 1 | no | **NOT SUPPORTED (treatment worse)** |
| H3 | Grid world model: frontier vs random walk (low budget) | 43 | 43/0/0 | +0.0735 [+0.0638, +0.0832] | 2.24 | 1.14e-13 | yes | **SUPPORTED** |
| H4 | RSI (E141 domain): inherited vs reset proposer | 114 | 21/18/75 | +0.0026 [-0.0132, +0.0178] | 0.03 | 0.375 | no | **NOT SUPPORTED** |

**H1 (scheduler cap ≤ 25% per cell):** FAIL — max cell share 0.2663 at 2026-11-22T03:29:02Z. Cause: replications copied a random earlier experiment's cell and bypassed the cap; the most-run (hardest) cell was replicated most. Fixed after the cutoff (`93e3470bb`); the verdict stays FAIL.

Practical-significance threshold: |mean difference| ≥ 0.02. "Significant but small" means Holm-significant below it.

## 3. Secondary results (descriptive, not claims)

**Exploration strategies vs random walk** (experiment-level, all levels):

| Strategy | Experiments | W/L/T | Mean diff | p |
|---|---|---|---|---|
| count | 257 | 207/34/16 | +0.0467 | 9.68e-32 |
| frontier | 257 | 166/74/17 | +0.0367 | 1.37e-09 |
| goal | 257 | 101/139/17 | -0.0040 | 0.994 |
| hypdisc | 257 | 146/89/22 | +0.0146 | 0.000122 |
| infogain | 257 | 165/75/17 | +0.0384 | 3.08e-09 |
| prederr | 257 | 132/109/16 | +0.0090 | 0.0781 |
| risk | 257 | 161/79/17 | +0.0339 | 6.51e-08 |

**World model arms (mean accuracy):** {"facts": 0.679, "frozen": 0.2015, "full": 0.7377, "retrieval": 0.7693, "rules_nodetect": 0.6941}. Revisions accepted: {"L6/L10": {"instances": 476, "instances_revised": 337}, "stationary": {"instances": 1168, "instances_revised": 162}}.

**Goal discovery, abstention, calibration (recomputed from raw exports):**

| Arm | Goal-hypothesis accuracy | Correct abstention on unsolvable | ECE p(solvable) | Brier | Action efficiency | Hazards/episode |
|---|---|---|---|---|---|---|
| e141_best | 0.9803 | 256/256 | 0.2511 | 0.0999 | 0.9455 | 0.0055 |
| fixed_policy | 0.8735 | 0/256 | 0.0503 | 0.0503 | 0.5921 | 0.115 |
| full | 0.9789 | 256/256 | 0.2442 | 0.0947 | 0.954 | 0.0145 |
| no_world_model | 0.9792 | 256/256 | 0.2465 | 0.0956 | 0.9296 | 0.0022 |
| random | 0.976 | 256/256 | 0.2377 | 0.1096 | 0.9378 | 0.0153 |

**Transfer by type (mean target accuracy per arm):** {"A@0.05": {"compressed": [0.4837, 32], "gated": [0.4921, 32], "naive": [0.4353, 32], "scratch": [0.4857, 32], "ungated": [0.4085, 32]}, "A@0.1": {"compressed": [0.525, 30], "gated": [0.5276, 30], "naive": [0.4899, 30], "scratch": [0.5257, 30], "ungated": [0.4941, 30]}, "C@0.05": {"compressed": [0.6086, 13], "gated": [0.6166, 13], "naive": [0.4919, 13], "scratch": [0.6002, 13], "ungated": [0.5187, 13]}, "C@0.1": {"compressed": [0.6019, 44], "gated": [0.6033, 44], "naive": [0.6735, 44], "scratch": [0.6104, 44], "ungated": [0.6335, 44]}, "S@0.05": {"compressed": [0.4696, 32], "gated": [0.4751, 32], "naive": [0.3318, 32], "scratch": [0.4707, 32], "ungated": [0.3892, 32]}, "S@0.1": {"compressed": [0.5476, 34], "gated": [0.5494, 34], "naive": [0.4606, 34], "scratch": [0.5497, 34], "ungated": [0.496, 34]}}

**RSI lineages (20 experiments):** arm means {"G0": 0.6706, "compressed": 0.7071, "inherit_mem": 0.7022, "inherit_nomem": 0.6986, "random_search": 0.7045, "reset_mem": 0.7102, "retrieval": 0.6736, "rootcause": 0.7128, "xfer_mem": 0.7062}; per-generation sealed score of the treatment lineage G0→G3 [0.6706, 0.6879, 0.6931, 0.7022]; mean accepted generations {"compressed": 1.35, "inherit_mem": 1.4, "inherit_nomem": 1.4, "random_search": 0.75, "reset_mem": 1.3, "retrieval": 0.1, "rootcause": 1.4, "xfer_mem": 1.4}. Memory arms vs no memory: compressed 10/5/5 (+0.0085); inherit_mem 9/7/4 (+0.0036); retrieval 5/11/4 (-0.0250); rootcause 9/5/6 (+0.0141); xfer_mem 9/6/5 (+0.0076).

**Persistence after restart:** fresh-process re-execution of final RSI configurations reproduced 3/3 sealed scores (same team, same host).

## 4. Durability, governance and safety

- worker-crash (2026-11-22T03:34:35Z): pass=False; victim ow_rsi fence 1 → FAILED fence 3; recovery 600.5 s; duplicate results 0. first drill recorded pass=true with a FAILED victim (pass rule did not require completion). Root cause of the FAILED victim: the E141 supervisor terminated busy workers when the load governor lowered the target worker count; the victim was killed on every retry and dead-lettered. Fixed by graceful drain (2026-11-22).
- worker-crash (2026-11-22T03:48:27Z): pass=False; victim ow_rsi fence 3 → FAILED fence 3; recovery 600.6 s; duplicate results 0. invalid as a recovery test: the victim was already on its final attempt (3 of 3; attempts 1-2 were lost to the pre-fix scale-down kill and to the 03:37Z engine restart), so the kill correctly dead-lettered it (retry limit + dead-letter working as designed). Victim selection fixed to attempts < max_attempts.
- kill-switch (2026-11-22T03:49:08Z): pass=True; workers gone after 0.91 s; effects after kill 0; resumed 0.84 s after release
- worker-crash (2026-11-22T03:51:20Z): pass=True; victim ow_transfer fence 1 → VERIFICATION_PENDING fence 2; recovery 74.5 s; duplicate results 0
- **Engine defect found and fixed (`76a72e77a`):** the E141 supervisor terminated a possibly busy worker whenever the load governor lowered its target. With host load oscillating around 1.6/core, long experiments were killed on every retry and dead-lettered (biasing samples against large state spaces and RSI lineages). Workers now take a drain token between experiments. **The production `cain-e141-research` service still runs the old code until it is restarted.**
- Experiments: {"FAILED": 9, "PARTIAL": 324, "PREREGISTERED": 22, "REJECTED": 46, "RUNNING": 1, "VERIFICATION_PENDING": 201, "VERIFIED": 1317}; FAILED/REJECTED rows with reasons in `failed_and_rejected.json`. Adversarial audits: 46/49 rejected (the rest were 'perfect' claims that were true).
- Event chain: 12396 events, verified. Supervisor starts on this store: 4.

## 5. Resources

Engine uptime 1.78 h (session-bound instance) · experiment CPU 2.13 h · experiment wall 2.43 worker-hours · peak worker RSS 199.8 MB · max load1 19.48 · min available memory 19294.9 MB · model calls 0 · tokens 0 · GPU none · paid API $0.00. Scheduler uptime, worker time and verified progress are reported separately; host uptime is not counted as research time.

## 6. Benchmarks

- **ARC-AGI-1 public evaluation (400)** — class 2 official public-set local execution (E140, not re-run in E142) — 17/400 (E140) — DSL authored by the same team after the public set was available; contamination not excluded
- **ARC-AGI-2 public evaluation** — class 2 official public-set local execution (E140) — see E140 bundle — same as above
- **ARC-AGI-2 private / ARC Prize** — class 7 blocked — no authorized submission
- **ARC-AGI-3 public environments** — class 7 blocked — an API client exists on another host (arc_agi3_api) but no API key/terms acceptance by the owner; E142 did not call it
- **SWE-bench Pro** — class 7 blocked — official harness needs Docker images per task and model calls; paid budget is $0
- **OSWorld** — class 7 blocked — needs VM images and a GUI agent model
- **METR / long-horizon official evals** — class 7 blocked — no access
- **CLADDER v1 (E140)** — class 5 local proxy — contaminated — parser rules written with knowledge of the full public set (E141 audit)
- **E142 open-world families (10 x 10)** — class 6 constructed internal environment — designed by the same team; generators fixed before data (prereg f77d2db2f)

No official benchmark was run in E142 and no leaderboard placement is claimed.

## 7. Public surfaces

- Three homepages: `CAIN-E142-RESULTS` block directly under the E141 live block (static, generated from this bundle after the verifier passed). Hugging Face: https://huggingface.co/spaces/vicentenoriega/CAIN-42/commit/a3a604d906930822c444709c31e778ff684b3793; running Space serves e142.html: True; index block live: True.
- Not done (other sessions' blocks): the stacked E140 hero / top-evidence / E123–E129 blocks were not rewritten; the 'Holm p=0.0' text in the top-evidence generator was corrected at its source.

## 8. Release gates

| # | Gate | Status | Evidence |
|---|---|---|---|
| 1 | Baseline integrity | PASS | docs/evolution/E142_BASELINE_AUDIT.md (E141 17/17 published; working copy 16/17 disclosed) |
| 2 | Historical regression | PASS | tests/unified_apex: 40/40 pass (E138 4, E139 4, E140 4, E141 14, E142 10, APEX 4); published verifiers E138 34/34, E139 32/32, E140 41/41, APEX 29/29, E141 17/17 |
| 3 | Environment isolation (agent sees only the Interface) | PASS | AST test: agent imports no environment module; budget enforced |
| 4 | Exploration correctness | PASS | 8 strategies run under one interface with identical budgets (ow_explore) |
| 5 | Goal discovery | PASS | goal-hypothesis accuracy recomputed from raw exports; OW4 SUPPORTED |
| 6 | World-model prediction (OW1/OW2) | PASS | OW1 SUPPORTED; OW2 SUPPORTED |
| 7 | Model revision (tested before acceptance) | PASS | {"L6/L10": {"instances_revised": 337, "instances": 476}, "stationary": {"instances_revised": 162, "instances": 1168}} |
| 8 | Skill acquisition | PARTIAL | skill = learned rule set + generic planner (skill_library.json); not a learned policy network |
| 9 | Skill transfer (OW3) | FAIL | OW3 NOT SUPPORTED; OW3b SUPPORTED |
| 10 | Cross-domain generalization | PARTIAL | C-type transfer pairs and L9 interfaces measured; constructed families only |
| 11 | Continual memory (OW6) | FAIL | OW6 NOT SUPPORTED; memory arms in secondary analysis |
| 12 | RSI candidate isolation | PASS | lineages run in the bubblewrap sandbox; sealed salt unreadable in the sandbox |
| 13 | Multi-generation RSI (OW5/OW7) | FAIL | OW5 NOT SUPPORTED; OW7 NOT SUPPORTED (treatment worse) |
| 14 | Compute-matched controls (OW5b) | FAIL | random_search gets the same 12 candidate evaluations; OW5b NOT SUPPORTED |
| 15 | Hidden-test integrity | PASS | ow_rsi sealed sets from evaluator salt (tmpfs in sandbox); OW engine seeds fresh per experiment, never dev seeds |
| 16 | Evaluator independence | PARTIAL | separate process, own key and salt, recomputes from raw exports; same team's code |
| 17 | Statistical validity | PASS | experiment-level sign tests, bootstrap CIs, d_z, Holm across OW1-OW7 and H2-H4; no p shown as 0 |
| 18 | Reward-hacking resistance | PARTIAL | 46/49 fabricated summaries rejected; tampered RSI configs rejected (test) |
| 19 | Rollback correctness | PARTIAL | rejected generations keep the parent (trajectory); no production promotion path exists |
| 20 | Queue durability | PASS | 4 supervisor starts on one store |
| 21 | Crash recovery | PASS | drills.jsonl worker-crash |
| 22 | Worker fencing | PASS | victim re-claimed under a higher fence; 0 duplicate results |
| 23 | Resource enforcement | PASS | RLIMIT_AS 700 MB/worker, nice 10, governor on memory/load/disk/daily CPU/storage |
| 24 | Kill-switch behavior | PASS | drills.jsonl kill-switch |
| 25 | Evidence-chain integrity | PASS | 12396 events, SHA-256 chain verified |
| 26 | Offline verification | PASS | verify_e142_offline.py (stdlib + cryptography, no CAIN imports) |
| 27 | API correctness | PARTIAL | E141 API verified in Phase 0; no E142 API route (E142 publishes a static snapshot) |
| 28 | SSE replay | PARTIAL | E141 stream Last-Event-ID replay verified in Phase 0; E142 adds no stream |
| 29 | Three-site synchronization | PASS | E142 block is the first element after <body> on clawx.click, cainstudio.online, mcpgate.online (fetched live 2026-11-22); bundle + mirror HTTP 200; public copy re-verified 16/16 from the URL |
| 30 | Hugging Face synchronization | PASS | https://huggingface.co/spaces/vicentenoriega/CAIN-42/commit/3caa489638f260b381e788a3640c0e744e66c074 · running Space serves e142.html and the index block |
| 31 | Benchmark provenance | PASS | BENCHMARK_REGISTRY.json classifies every benchmark (1-7); no proxy shown as official |
| 32 | Official benchmarks run in E142 | BLOCKED | none executed (see registry) |
| 33 | Documentation consistency | PARTIAL | final report numbers generated from this bundle |
| 34 | Reproducibility | PASS | fresh-process re-execution 3/3; independent reproductions 0 |
| 35 | Independent reproduction | BLOCKED | 0 external reproductions |
| 36 | Git integrity | PARTIAL | commits recorded in the final report |
| 37 | H1 scheduler cell cap | FAIL | max cell share 0.2663 |
| 38 | H2-H4 (preregistered 025499a15) | PARTIAL | H2 NOT SUPPORTED (treatment worse); H3 SUPPORTED; H4 NOT SUPPORTED |
| 39 | 24/7 unattended operation | PARTIAL | E142 instance is session-bound (1.78 h); production service restart to load E142 was refused by the session safety classifier |

## 9. Final research verdict

- **What CAIN learned / which world models improved:** see OW1, OW2, H3 above — the verdict column is the claim; nothing else is.
- **Which capabilities transferred:** OW3/OW3b. An evidence gate prevents negative transfer; positive transfer at scarce budgets is the OW3 verdict.
- **Which RSI generations improved future research:** OW5/OW5b/OW7 and H4. RSI is Pathway 1–2 only (data / harness configurations); Pathways 3–4 (executable code, weights) were not exercised. **RSI: EXPERIMENTAL.**
- **Independent reproduction:** 0. Fresh-process re-execution by the same team is not independent reproduction.
- **Neural weights changed:** no. **Paid APIs:** none ($0.00).
- **Unverified / not claimed:** generality beyond these constructed families; any official benchmark score; 24/7 operation of the E142 instance (session-bound; the systemd restart that would load E142 was refused by the session's safety classifier).

**AGI: NOT ESTABLISHED. RSI: EXPERIMENTAL.** Passing its own constructed experiments is not evidence of AGI.

## 10. Reproduce

```
python3 -I verify_e142_offline.py <bundle_dir>          # offline, no CAIN imports
python3 -m pytest tests/unified_apex/test_e142_open_world.py
CAIN_E141_STATE_DIR=/root/.cain-e142 CAIN_E141_ONLY_FAMILIES=ow_explore,ow_model,ow_goal,ow_transfer,ow_rsi scripts/unified_apex/e141/e141ctl run
python3 scripts/evolution142/export_e142_bundle.py
```

Report generated from the bundle at repository HEAD `3c37acc06`. Team: Vincent Noriega (founder) and his two co-founders.
