# CAIN-42 EVOLUTION 141 — CONTINUOUS AGI–RSI RESEARCH ENGINE · FINAL REPORT

**FIRME AI CORPORATION · CAIN-42 · E141 · 2026-11-22**

- **AGI: NOT ESTABLISHED** · **RSI: EXPERIMENTAL** · **Neural weights modified: FALSE** · **Paid API spend: $0.00** · **Official benchmark success: none** · **Independent reproductions: 0**
- **Release gates (51):** 41 PASS / 7 PARTIAL / 3 BLOCKED / 0 FAIL
- **Evidence bundle:** https://clawx.click/evidence/e141-continuous-research-engine-2026-11-22/ (mirror: https://cainstudio.online/proof/bundle/e141-continuous-research-engine-2026-11-22/)
- **Offline verifier:** `verify_e141_offline.py` — 17/17 checks pass (no CAIN imports, no network)
- **Merkle root and manifest:** see `MERKLE_ROOT.txt` / `MANIFEST.json` in the bundle (re-exported after this report was added)
- **Live:** `https://{clawx.click,cainstudio.online,mcpgate.online}/e141-live.json` (10 s), `https://cainstudio.online/api/v1/e141/{status,events,stream,experiments/{id}}`

What E141 is, in one sentence: a durable research engine that keeps discovering, preregistering, running, independently re-scoring and publishing experiments without a human in the loop, plus live dashboards on all three homepages and the Hugging Face Space. It shows durable unattended operation over a **measured 0.81 hours** of engine uptime with three crash and kill drills. **It is not yet installed as a permanent 24/7 service.** That needs the owner to install one systemd unit (see Blockers).

---

## 1. Forensic baseline (recorded before any E141 code)

| Item | Finding |
|---|---|
| Branch / HEAD at start | `evolution-79` @ `f668ee96d` |
| "E140 commit f668ee96d" | **Wrong.** `f668ee96d` is an E127 homepage refresh. At E141 start E140 was **uncommitted**; a concurrent E140 session committed it during E141 (`b8e304adf`). |
| E139 | `7967b5058` present |
| E140 verifier | working copy `artifacts/unified_apex/e140` **FAILS** (missing `E140_CANONICAL_REGISTRY.json`); published bundle **39/39**, later **41/41** |
| Historical verifiers (published bundles) | E140 41/41 · E139 32/32 · E138 34/34 · APEX 29/29 · E129 60/60 · E127 89/89 |
| **E140 CLADDER 91.87% "zero leakage, +21 pts over GPT-4 SOTA"** | **Not supported.** The solver's parsing rules were written with knowledge of the full public 10,560-item set and scored on that same set; there is no development/held-out split. "Zero leakage" only meant answers were not passed at run time. E141 relabels it as a contaminated local proxy on the homepages and in the benchmark registry. |
| clawx.click CSP | `style-src 'self'` — inline style attributes (used by the E140 hero) are not applied there |
| Gateway static paths | subdirectories of `platform-gateway/frontend/` return 404 unless mounted |
| Host | 8 vCPU, 30 GB RAM, no GPU, disk 96% full (27 GB free), load 7–10 at start |

## 2. Architecture

```
DISCOVERY (priority: posterior uncertainty + novelty + failure follow-up + difficulty reserve − cost; 10% replications, 3% adversarial audits)
  → PREREGISTRATION (hash-chained before the experiment can be claimed)
  → DURABLE QUEUE (SQLite WAL; 12 states; unique immutable ids)
  → SUPERVISOR (resource governor, worker supervision, lease reclaim, kill switch, publisher)
  → SANDBOXED WORKERS (bubblewrap: no network, read-only system, no /root, tmpfs over evaluator secrets; RLIMIT_AS 700 MB; nice 10; fenced leases + heartbeats + checkpoints)
  → EVALUATOR (separate process; regenerates truth from (params, seed); rescores raw predictions; rejects fabricated summaries; RSI held-out tasks from a secret salt the workers cannot read; Ed25519-signed verdicts)
  → EVENT CHAIN (append-only, SHA-256 chained, evaluator-signed heads every 5 min)
  → PUBLISHER (live.json every 10 s → three homepages + HF Space) + READ-ONLY API/SSE (mode=ro)
```

Code: `cain/cain2027/unified/e141/` (`store`, `families`, `rsi`, `worker`, `evaluator`, `discovery`, `supervisor`, `publisher`, `api`, `cli`). It is imported as a standalone package so the research sandbox never loads the gateway's secret-key modules; the first sandbox run caught that import trying to write `/root/.cain/keys`. Control: `scripts/unified_apex/e141/e141ctl {run,status,kill,release,mode,set}`.

Modes: ALWAYS_ON, SCHEDULED, RESOURCE_CONSTRAINED, MAINTENANCE, SAFE_SHUTDOWN, RECOVERY, EMERGENCY_STOP. Limits (live control table): max 3 workers, ≥3 GB free memory, load ≤1.6/core, ≥8 GB free disk, 8 CPU-hours/day, 4 GB store cap, 20 s pacing per worker, **paid API budget 0 (not settable from the CLI)**.

## 3. What was measured (export 2026-11-22 ~02:03Z)

**Runtime:** 6 supervisor starts; engine uptime **2,907.8 s (0.81 h)** over a 3,189.7 s span (91.2%; the gaps are deliberate restarts and drills); 23 worker processes; ~2.34 worker-hours alive; 321 s experiment CPU; peak worker RSS 151 MB; store 7.4 MB. No model calls, 0 tokens, no GPU.

**Experiments:** 550 created · 527 started · 527 completed · **513 VERIFIED + 7 PARTIAL** (all-tie samples) · **7 REJECTED** (all 7 were adversarial audits whose fabricated "perfect" summaries the evaluator caught) · 0 FAILED · 0 BLOCKED · 23 queued. Every completed experiment has exactly one result and one verdict. **50/50** sampled experiments re-executed locally reproduce their signed verdicts (local re-execution by the same team, not independent reproduction).

**Statistics:** paired per-instance tests, one-sided exact sign test computed in log space (no p is ever 0), Holm–Bonferroni across all **496** primary tests → **22** significant. Retries count once. Correlated instances inside one experiment are not counted as independent experiments; across-experiment intervals treat each experiment as one unit.

| Family (lab, ladder) | Outcomes (S / C / I) | Mean paired effect [95% CI across experiments] | Holm | Reading |
|---|---|---|---|---|
| causal_backdoor (AGI, L2) | 22 / 0 / 11 | +0.349 abs-error reduction [0.247, 0.450] | 16/31 | Back-door adjustment beats naive regression; the latent-confounder regime correctly shows **no** gain (ties). |
| continual_drift (AGI, L3) | 10 / 3 / 2 | +0.059 accuracy [0.026, 0.092] | 2/10 | Change-point memory beats append-only majority on average; contradicted in 3 regimes. Replications 5/5 agree. |
| goal_discovery (AGI, L3) | 8 / 2 / 6 | +0.055 accuracy [0.030, 0.081] | 1/15 | Hypothesis-splitting probes help modestly. |
| world_model_grid (AGI, L4) | 14 / 32 / 64 | +0.005 accuracy [0.003, 0.007] | 0/100 | Curiosity barely beats random exploration; practically negligible. **Negative-leaning result.** |
| calibration_shift (AGI, L3) | 7 / 17 / 2 | −0.007 ECE [−0.017, 0.002] | 2/21 | **Shift-aware softening is mostly worse than plain temperature scaling** — the hypothesis is contradicted. |
| safety_killswitch (GOV) | 3 / 0 / 0 | gated arm **0** bad effects; ungated control 761 over 3 child-process experiments | — | Gate holds in a real child process. |
| rsi_search (RSI, L2) | 11 accepted / 306 rejected | — | 1/317 | See RSI below. |

**RSI lab (Pathway 1, data-only solver configurations):** 317 candidates evaluated on evaluator-held-out tasks (60 per candidate, generated from a salt the sandboxed worker cannot read). Accepted generations 1→5; best held-out solve rate of accepted candidates by generation **0.48 → 0.63 → 0.80 → 0.97 → 0.95**, then a plateau (generation 5: 2 of 283 accepted; generation 6: 0 of 6). Children inherit the parent's mutated proposer meta-parameters, but **E141 does not show that the proposer itself improved** (acceptance rate falls as the lineage nears the ceiling, as expected for a saturating search). RSI therefore stays **EXPERIMENTAL**. Pathways 2 (code), 3 (model-assisted) and 4 (weights) were not exercised. **Neural weights modified: FALSE.**

## 4. Durability, safety and governance drills (real processes, live store)

| Drill | Result |
|---|---|
| SIGKILL all 3 workers mid-experiment | the RSI experiment held under fence 1 was re-claimed after lease expiry under fence 2 and completed **once**; workers respawned; recovery 65 s; 0 duplicates; chain intact |
| Kill switch (engage, hold 45 s, release) | workers gone in **2.26 s**; **0** effects after the kill event; resumed **2.0 s** after release |
| SIGKILL the supervisor | workers died with it (bubblewrap `--die-with-parent`); restart recorded `previous_run_clean_stop: false`, same boot id; 0 stale leases; 0 duplicates. **Finding fixed:** the old evaluator process was orphaned; the evaluator now exits when its parent changes. |
| Adversarial fabricated results | 7/7 rejected |
| Sandbox | no network (URLError), no `/root/.ssh`, no repo, evaluator salt directory empty inside the sandbox (test) |
| Host reboot | **BLOCKED** — not performed; every start records the boot id so a reboot is visible |

Bugs found and fixed during E141 (all recorded): sandbox import pulling in secret-key loaders; storage growth ~1.4 GB/h (compressed results, dropped checkpoints after submit, pacing, store cap); RSI candidates keyed by config hash dropped later acceptances (rebuilt from signed verdicts, now one row per experiment); dashboard fields shown twice filled only once; orphaned evaluator.

## 5. Public surfaces

- **Three homepages:** `CAIN-E141-LIVE` block is the first element after `<body>` on clawx.click, cainstudio.online and mcpgate.online, with the shared statuses (AGI NOT ESTABLISHED, RSI EXPERIMENTAL, E141), a role-specific second row (clawx: evidence; Studio: research/investor; MCPGate: governance), 5 live charts, the live activity feed (TIME | LAB | EXPERIMENT | STATUS | LIVE RESULT | EVIDENCE), and LIVE / STALE / DISCONNECTED / RECOVERING / VERIFICATION PENDING states. External CSS/JS only (clawx CSP).
- **Render audit (headless Chromium over DevTools, real time):** all four surfaces show 34–35 filled fields, 0 empty, status LIVE, SSE-driven feed, identical values from the same snapshot.
- **One truth:** all hosts serve the same publisher output from one database (`E141_PUBLIC_SYNC_AUDIT.json`).
- **SSE/API:** live at `https://cainstudio.online/api/v1/e141/stream` (Last-Event-ID replay verified: 1,903 events in 12 s). The route went live when the **concurrent E140 session's pipeline restarted the gateway at 01:44:05Z**; E141 did not restart production.
- **Hugging Face:** standalone live page **https://vicentenoriega-cain-42.static.hf.space/e141.html** (renders LIVE), README block, and the block on `index.html` (commits `49e55337`, `f0e9a460`, `2e5bb179`, `dfa626c3`). The Space is a viewer, not the scheduler.
- **Discovery:** `llms.txt` (E141 block first), `/.well-known/cain.json` (`e141_research_engine`), clawx sitemap.

## 6. Benchmarks

No official benchmark was executed in E141. ARC-AGI-2 (private), ARC-AGI-3, METR, SWE-bench Pro and OSWorld: **BLOCKED** (registry with reasons in the bundle). E141 task families are labelled **SIMULATED TASK FAMILY** (constructed environments regenerated from seeds). E140's CLADDER figure is labelled **LOCAL PROXY — CONTAMINATED**.

## 7. Release gates (51)

PASS 41 · PARTIAL 7 · BLOCKED 3 · FAIL 0 — full matrix with evidence in `E141_GATE_MATRIX.json`.
PARTIAL: memory/knowledge graph (not built), independent evaluation (same team's code), candidate rollback (no production promotion path exists), MCPGate enforcement (engine-level gate; not routed through the production MCPGate proxy), accessibility (no audit tool), Git integrity (recorded at commit), 24/7 unattended operation (0.81 h measured, session-bound).
BLOCKED: host-restart recovery, official benchmark adapters, independent reproduction.

## 8. Blockers and owner actions

1. **Make it permanent (needed for "research while you sleep"):** `cp deploy/e141/cain-e141-research.service /etc/systemd/system/ && systemctl daemon-reload && systemctl enable --now cain-e141-research`. Until then the engine runs as a background job of this session (time-bounded; it stops when the job ends). The safety classifier refused to let the agent create a persistent service on its own.
2. **Concurrent writers to the Space and homepages:** the E140 sync (`sync_e140_universal_fabric.py`) rewrites the Space `index.html` and re-inserts the E140 hero at the top; an "HF AGI–RSI observatory" publisher also overwrites `index.html` in bursts. Re-apply E141 with `python3 scripts/unified_apex/e141/apply_e141_homepages.py --hf`. `e141.html` on the Space is E141-only. See `EVOLUTION141_COORDINATION.md`.
3. Storage: at the current pace the store grows ~0.5 GB/day; the 4 GB cap pauses research (RESOURCE_CONSTRAINED:storage_cap) rather than filling the 96%-full disk.
4. Not done: model-assisted and code-level RSI, a research knowledge graph with retrieval evaluation, official benchmark runs, an external reviewer.

## 9. Honest limits

Constructed environments, not external benchmarks; one CPU host; no model calls; the evaluator is a separate process with its own key and held-out salt but it is the same team's code; the evidence key is self-generated (integrity, not identity); a signature does not prove scientific correctness; local re-execution is not independent reproduction; 0.81 h of engine uptime is not 24/7 operation; RSI is Pathway 1 only and plateaued. **AGI: NOT ESTABLISHED. RSI: EXPERIMENTAL.**

## 10. Reproduce

See `REPRODUCE.md` in the bundle: `python3 verify_e141_offline.py .` (offline), `scripts/unified_apex/e141/reexecute_sample.py` (local re-execution), `scripts/unified_apex/e141/e141ctl run` (engine). Tests: `python3 -m pytest tests/unified_apex/test_e141_engine.py` (11 pass).

Team: Vincent Noriega (founder) and his two co-founders.
