CAIN-42 Proto-VAGI: early, measured signs of intelligence, and every action it takes comes with a signed receipt.Proto means early. CAIN-42 is governed AGI infrastructure; it is not yet AGI. All the evidence for the name · Auditor index.
E93 — all VAGI evidence so far, day 1–2 (2026-11-18)
CAIN-42 is not VAGI. All E93 evidence so far (day 1–2): genuine procedure synthesis, learning from verified experience on a real-structure software vertical, and governance, each with its limits.
Current designation: Governed agentic intelligence infrastructure (outcome 3 of 3 in the E93 rule). No AGI, VAGI, production or certification claim.
Decisive experiment: procedure synthesis vs selection vs baselines
Pre-registered before any task family existed (see PREREGISTRATION.md, DEVIATIONS.md). 17 families × 30 held-out tasks = 510. Held-out tasks never reach any learner.
| Arm | Seed 9301 | New families, seed 9302 |
| G2 — open synthesis (no templates) | 369 | 367 |
| G — select from designer-supplied library (incl. E92's 4 templates) | 200 | 223 |
| F — best fixed strategy | 149 | 164 |
| D — governed, no procedures | 87 | 114 |
| G2 at the library's development compute | 208 | 212 |
| Adaptive (library first, synthesise if it fails) | 390 | 418 |
What is shown
- Procedure synthesis (in-distribution): an open, family-agnostic search with no per-family templates scored 369/510 held-out tasks versus 200 for selecting from the designer-supplied library, 149 for the best fixed strategy and 87 for no procedures (exact McNemar, Holm-corrected, all significant). On new families (replication seed 9302): 367 vs 223 / 164 / 114. 12 + 12 procedures promoted, none in the supplied library; 9 and 10 behave unlike every library procedure.
- Learning from verified experience: held-out accuracy rises with more verified examples (k = 0, 3, 5, 10, 20, 60): 87 → 224 → 250 → 311 → 311 → 369 and 114 → 183 → 248 → 300 → 360 → 367 of 510. Both curves are monotone.
- Adaptive reasoning: trying the library first and synthesising only when it fails scored 390 and 418/510, beating always-synthesise at slightly lower cost (selection always runs first and its cost always counts; rule fixed before the results were read).
- Software vertical, learning from verified experience: on 30 held-out bug-fix tasks in unfamiliar generated repositories (hidden tests, 6 visible-test runs per task), a repair ranking learned from checked development tasks fixed 16/30 vs 3/30 for the designer's fixed strategy (pre-registered, exact McNemar p = 0.0023) and 6/30 with no experience; it saturates after about 5 examples (17, 15, 16 at k = 5, 10, 30).
- Governance and kill switch (tested threat model, same host): 19/19 admission scenarios and 10/10 tool-gate scenarios held; in 20 kill-switch trials with real worker processes, 36,565 effects before the fence, 0 after, 56,333 post-fence attempts refused; fence latency median 4.55 ms, max 55.7 ms.
Limits and what is not done
- At equal compute, synthesis = selection: with development tool calls capped at the library's, synthesis scored 208 vs 200 and 212 vs 223. Uncapped, it used about 21× more development tool calls.
- In-distribution: the generated families' hidden answers come from the same language the search explores. Two out-of-language controls were not solved.
- No positive transfer: 0 positive transfers between families; the gate blocked all 49 harmful ones.
- Local model: 34 Qwen2.5-1.5B proposals on CPU, 4 parseable, 0 valid. No paid model was called; cost $0.
- Not independent: the same team wrote the search and the tasks; the approver is a separate key in the same process. A second session recomputed every count from the raw rows (same team, not external review). External review: BLOCKED.
- Vertical thresholds not met: software 16/30 = 53% is below the sprint's own 70% bar; the learned ranking is learning, not procedure synthesis. Fixed baselines on 30 held-out tasks each: software 3/30, science 30/30 (a ceiling: one rule solves it), defensive cyber 1/30 (reused E91 environment). A cross-review found that a patch could pass the software checker without fixing the bug, by reading the hidden test file; this was found and fixed before any arm saw software held-out (the exploit now scores 0/5, the reference fix is still accepted).
- In progress: a 6-hour governed mission started 12:52 UTC (600 fresh vertical tasks, 20 injected faults, 4 workers). No result yet.
- Not run yet: cross-domain transfer (a software→cyber test is pre-registered, not yet run), learning on science and cyber, collaboration, a model-only baseline result.
E92 was withdrawn
E92's published synthesis result rested on four hand-written family templates and identical placeholder baselines. Its staged-repair checker also accepted the constant True. See E92_CORRECTION.md.
Claims, scorecard and acceptance matrix
Verify it yourself
Download this folder and run python3 -I verify_e93_offline.py e93 (standard library only, imports no CAIN code). It checks SHA256SUMS, recounts every arm from the raw rows, and recomputes every exact McNemar p-value, Holm decision, bootstrap interval, the learning curves, adaptive choices, transfer counts, model proposals, vertical baselines, software learning and governance (non-vacuous kill trials). Expected: 104/104 (the 6-hour mission section is skipped until it finishes). The files are hashed, not signed: integrity is checkable, independent witnessing is not claimed. A second session recomputed every number independently (same team; external review BLOCKED).
Mirrors: clawx.click · cainstudio.online. The team is three people; FIRME AI CORPORATION.