CAIN-42 · RSI RESEARCH LAB
AGI: NOT ESTABLISHED · RSI: EXPERIMENTAL
CONSTRUCTED DOMAIN · BOUNDED CLAIMS
SIGNED EVIDENCE · OFFLINE VERIFIERS
Recursive Self-Improvement, Measured
AGI: NOT ESTABLISHED · RSI: EXPERIMENTAL
CAIN-42's RSI Research Lab studies one question under governance: can a system write
executable improvements into itself, and does the improved system solve held-out tasks the previous
one provably could not? Every claim below is a sealed evaluation with a committed answer key, a
matched control and an offline verifier that imports no CAIN code.
Best verified RSI result (E135)
35 / 48
sealed tasks solved; tier-6 6/6; tier-8 5/6; full control suite.
Matched-budget control
22 / 48
same library, same nominal budget, old frontier. The frontier adds +13.
Compute-only control
6 / 48
3.44M evaluations, no learned library: no capability beyond frozen.
Offline verification (E135)
1,551 / 1,551
signed checks recomputed from raw rows by a verifier with no CAIN imports; seed-131 replication repeats the direction.
What is established: in a constructed eight-concept domain, a governed loop
that writes its own executable macro library and selects a per-task search frontier raises sealed
solves from 26 to 30/48 and tier-6 from 2/6 to 6/6. The gain is not explained by
more library entries, more generations, more compute, or a smarter meta-selector.
What is not established: AGI, VAGI, unbounded RSI, human-level capability,
external replication or production readiness. Tiers 8, 12 and 16 remain unsolved. The domain is
constructed. One team wrote the engine, generator and verifier. There is no third-party audit.
The measured progression
| Evolution | What changed | Sealed result | What it proved / failed |
| E129 | First bounded RSI loop: self-written macro library, sealed tiers, frozen/random/flat/compute controls | 35/48 on the easier tier-1–8 set; tier-4 1.000 | Working bounded loop. Saturated at generation 2. |
| E130 | Scale curriculum (347 tasks), library (512), meta-strategy, 8 generations | main 26/48; fixed 28/48; flat 24; compute 17; frozen 6 | Vocabulary scaling hit a ceiling. The meta-improver hurt. One real rollback. |
| E131 | Per-task search portfolio hypothesis | main 25; fixed 27; plan 8 (invalid harness arm) | Failed intervention. Diagnosis: only 6 of 8 concepts were learned; the top-8 missed one. |
| E132 | Concept-complete curriculum + pinned alphabet + portfolio | main 30/48; plan 30; matched fixed 28; fixed 27; compute 6 | Tier-6 6/6 (was 2/6). Search claim positive. Meta selection neutral. Compute alone is nothing. |
| E133 | 4+4 meet-in-the-middle for tier 8 | status SOLVED, but 0/10 correct on every tier-8 task | Spurious fits. The MITM returned reversed halves; 16–48 rows did not identify an 8-concept chain. |
| E135 | Corrected half-order + active-identifiability task construction + full controls | 35/48; tier-6 6/6; tier-8 5/6; flat 22; random 15; compute 6 | RSI claim POSITIVE_BOUNDED (20/20; seeded replication repeats). One tier-8 task abstains; tiers 12/16 remain 0. |
| E136 | Deduplicated 8+4 tier-12 frontier (two 4-chunk compositions, fit/validation split) | Tier-12 unit: 10/10 on task 0000; campaign 0/6 (cap/state-set truncation) | 2,159 4-states → 1,632,736 8-states; ~8.3 min/task. The campaign hit the per-task cap and all six tier-12 tasks abstained; 2/6 also exceed int64. Tier 16 not attempted. Not a tier-12 campaign success. |
| E138 | Batched 8+4 and Z3 constraint-solver attempts for tier 12 | No campaign: batched join exceeded 900 s; Z3 encoding exceeded memory | The Python/NumPy join and the naive SMT encoding are not viable at 12 concepts. Next: compiled join, table-constrained CP, or learned depth-8 chunks. Not a result. |
Why this is a real RSI loop (narrow sense)
- The artifact is executable. The engine runs on a library of integer programs. The improver writes new entries from its own solved development programs; generation N+1 runs on generation N's macros.
- The evaluator is sealed. Answer keys are committed by SHA-256 before the run and never passed to the engine.
- Promotion is development-only. Sealed results never enter improvement or promotion.
- Controls are matched. Frozen, random, flat, compute, and same-library/matched-budget fixed-frontier arms.
- Evidence is recomputable. The verifier re-applies every recorded program to the sealed inputs, recomputes every accuracy, checks macro expansions, `library_used` digests, journal chains, the Merkle root and the Ed25519 signature.
Verify it yourself
The bundles are public. Each verifier imports no CAIN code.
python3 artifacts/e135/verify_e135_offline.py artifacts/e135 → 1,551/1,551 checks
python3 artifacts/e132/verify_e132_offline.py artifacts/e132 → 1,743/1,743 checks
python3 artifacts/e130/verify_e130_offline.py artifacts/e130 → 4,622/4,622 checks
python3 artifacts/e131/verify_e131_offline.py artifacts/e131 → 1,935/1,935 checks
- Tests:
python3 -m pytest tests/test_e132_rsi.py -q → 17 passed
Evidence bundles
The tier-12 blocker, measured
What works: the deduplicated 8+4 join solved one real tier-12 task 10/10 (2,159 4-states → 1,632,736 8-states, ~8.3 min). What does not: the campaign returned 0/6 because the per-task cap truncates the join; two of six tasks produce values around 10⁸⁰; the batched NumPy rewrite exceeded 900 s; the Z3 encoding exceeded memory. This is a computational wall, not a missing concept. Tiers 12 and 16 remain unsolved in a campaign.
Honest limits
- Constructed eight-concept integer domain; not professional work, not an external benchmark.
- Macros are data programs, not model weights. No model was trained or called.
- No external reviewer, no third-party audit, no SOC 2, no certification, no hardware attestation.
- The evidence key is self-generated: it proves integrity, not identity.
- Tier 8 is one composition step beyond the best verified plan; tiers 12 and 16 remain unsolved.
- Not deployed. No customer or revenue claim. AGI: NOT ACHIEVED.