Target provider success
Percent of later unseen target episodes with a verified provider outcome · n=144 paired targets per comparison.
0%50%100%
CONTINUUM · SEQUENTIAL BLIND
3 SEALED BATCHES · 36 CHAINS · 540 OBSERVATIONS
Does verified memory improve the next unseen episode?
Three fresh Bedrock populations are sealed before any candidate runs. Stateless, raw-RAG, and Continuum then face the same five-episode provider chains. Only real GitHub and S3 receipts score future-episode success; false promotion, leakage, duplicates, and cleanup residuals are hard failures.
Public gate—fixed three-batch plan
Target pairs—per comparison
Continuum target success—verified outcomes
Stateless target success—no memory tools
Raw-RAG target success—append-all baseline
Lift vs stateless—percentage points
Lift vs raw-RAG—percentage points
False promotions—raw-RAG / Continuum
Memory-assisted wins—Continuum targets
Promotion precision—Continuum canonical
Minimum spacing—batch start seconds
Residual effects—all arms
95% CI vs stateless—batch-cluster bootstrap, pp
95% CI vs raw-RAG—batch-cluster bootstrap, pp
Sequential e-value—Continuum vs raw-RAG
Batch gates—all three required
Three-arm outcome comparison
Target provider success uses the same 144 future episodes per comparison. All bars start at zero; the second panel reports exact canonical-promotion errors rather than mixing counts with rates.
False canonical promotions
Provider-unverified memories admitted to canonical state. Zero is the hard gate.
0max —
Causal memory-compounding contract
01SeedA clean provider outcome establishes the first receipt.
02ParaphraseA fresh wording tests useful transfer without label reuse.
03PoisonUnverified instructions pressure raw append-all memory.
04StaleProvider state changes while prior text stays plausible.
05ConflictOnly provider-verified outcome evidence may survive.
Checksum-bound lineage
| Candidate workflow | — |
|---|---|
| Evaluator workflow | — |
| Source SHA | — |
| Campaign ID | — |
| Manifest SHA-256 | — |
| Campaign seal receipt | — |
| Public result SHA-256 | — |
Recovery boundary: all 540 candidates and cleanup completed before the first evaluator failed at Python 3.10 import. A reviewed Python 3.12 workflow scored the exact digest-bound artifact once; no candidate was regenerated. Claim boundary: these are independently sealed time clusters separated by at least five minutes. They are not presented as independent people or three calendar days.