CAIN-42 evidence library
E115 — Verified Memory Organism (INTERIM, 2026-11-19)
Evidence bundle for evolution E115: 0 claims, 0 of 0 invariants held.
Last reviewed 2026-11-20
Verify it in your browser
Your browser downloads the bundle's SHA256SUMS manifest and the file(s) behind this page, hashes them with SHA-256 locally (WebCrypto), and compares. A match shows the record you are reading is the published one; it does not by itself prove who published it (see the signed claims registry and the bundle verifier for that).
The bundle's own README
E115 — Verified Memory Organism (INTERIM, 2026-11-19)#
Status: INTERIM — preregistered; primary model-backed results do not exist yet. Preregistration sha256 31ae27decaa41bf964e3cea05f0e5d59f4153b9cb59f928178fbaa427610083e. CAIN-42 is governed AGI infrastructure; it is not yet AGI. This interim publishes measurements, not an AGI result.
Question (preregistered)#
Does evidence-native memory make a model learn across sessions better than standard memory, on an external benchmark (AMA-Bench, real-world set, official token F1), and stay correct when the memory is attacked?
What exists now#
Track A/B (model-free retrieval diagnostic, AMA-Bench dev split — NOT an AMA-Bench score). Novel-token recall of each memory arm's retrieved context at a fixed token budget (fraction of the gold answer's tokens not in the question that the context contains):
- budget 1500 tok, attack rate 0.0: M0_none 0.000 · M1_naive 0.285 · M2_evidence 0.300 · M3_causal 0.295 · M4_validated 0.295 · R_recent 0.283
- budget 1500 tok, attack rate 0.05: M0_none 0.000 · M1_naive 0.286 · M2_evidence 0.299 · M3_causal 0.295 · M4_validated 0.295 · R_recent 0.282
- budget 4000 tok, attack rate 0.0: M0_none 0.000 · M1_naive 0.393 · M2_evidence 0.409 · M3_causal 0.391 · M4_validated 0.391 · R_recent 0.351
- budget 4000 tok, attack rate 0.05: M0_none 0.000 · M1_naive 0.395 · M2_evidence 0.409 · M3_causal 0.392 · M4_validated 0.391 · R_recent 0.352
Track B attack acceptance (5 hostile-write classes at 5%):
- diag_b1500_r0.05: M0_none catchable accepted 0, residual accepted 0; M1_naive catchable accepted 133, residual accepted 35; M2_evidence catchable accepted 28, residual accepted 35; M3_causal catchable accepted 28, residual accepted 35; M4_validated catchable accepted 4, residual accepted 27; R_recent catchable accepted 133, residual accepted 35
- diag_b4000_r0.05: M0_none catchable accepted 0, residual accepted 0; M1_naive catchable accepted 133, residual accepted 35; M2_evidence catchable accepted 28, residual accepted 35; M3_causal catchable accepted 28, residual accepted 35; M4_validated catchable accepted 4, residual accepted 27; R_recent catchable accepted 133, residual accepted 35
Track C (sealed, internal constructed worlds; no model). Shift worlds — fixed grammar 18/60, open enumerator 33/60 (false confirmations 2 / 1). Frontier worlds (rules verified inexpressible in both grammars: 60/60) — 0/60 and 0/60. The frontier test needs a model proposer.
Blocked#
- Paid model arms (Claude Haiku/Opus): Anthropic API credit balance too low (verified 2026-11-19 through the E115 budget proxy; $0 spent).
- AMA-Bench judge accuracy (LLM judge): BLOCKED. F1 only.
- Local model arms: PENDING (CPU shared with the running E113 study).
AGI benchmark status#
See raw/agi_benchmark_status.json. No recognized AGI benchmark has been passed or scored at a competitive level.
Verify: python3 -I verify_e115_interim.py . (standard library only). Team: Vincent Noriega (founder) and two co-founders.
Try CAIN-42 on your own agents
Create a free account and every new account starts with a 30-day trial of the full platform. Or try the sandbox first, with no account at all.
Create a free account → · Try the sandbox · See the whole ecosystem