The honest state of Oczy

Oczy asks whether an agent can be changed by experience without keeping the experience on the answer path. The intended loop is compact: experience causes fast neural change; replay compresses that change; slow state absorbs it; the raw trace is deleted; behavior survives.

That is stricter than “an LLM with memory.” Retrieval may remain useful in a product, but it does not establish the mechanism. The primary research condition therefore disables answer-time retrieval while every comparison table keeps vanilla, retrieval, and oracle columns visible.

Core thesis

experience → fast change → replay → compression → slow change → trace deletionBehavior must survive because state changed, not because text was stored.

Current verdict

Not demonstratedNo completed result yet proves a cortex learned an unseen rule and expressed it through a frozen model after trace deletion.

Active test

Research 20 / Experiment 09A meta-trained fast/slow cortex coupled to a frozen Qwen language organ.

The first seven experiments were not a failed prelude. They removed easy explanations and exposed the actual bottleneck. A residual control vector can move posture without reliably carrying a new fact. KV content can be prefix-equivalent without producing robust recall. A minimal hand-authored loop can run and consolidate while its measured behavioral effect remains zero.

Two bounded subproblems produced useful component results. Context-scoped attractors reached a scope-selectivity index of 1.0 in a single run. The later bounded-growth campaign measured a stable persistent ratio of 0.002079 across five seeds—but an earlier A0b version hit its byte target by regenerating a random matrix, which made learned updates unable to persist. The honest surviving claim is bounded footprint, not compression of learned behavior.

Exp01

NULLBehavior delta and discrimination both measured 0.0.

Exp02

REFUTEDKV-slot rank-1 recall was 0 while the logit-bias control reached 3.

Exp04

POSITIVE, single runScope-selectivity index reached 1.0; no cross-seed variance estimate.

Exp06

ENGINEERING POSITIVEBounded-growth ratio 0.002079 across five seeds; learned-content persistence is not established.

Exp07

POSITIVE + NULLCurrent campaign uptake +1.0 and critic delta 0.0; the older zero-by-construction +1.0 headline remains superseded.

Earlier designs assumed a hand-written Hebbian update and an untrained projection could turn an experience embedding into state that a frozen model would know how to use. That assumption hid two learning problems.

First, the cortex must learn how to learn: what to write, where to address it, how to retain it, and when to consolidate or forget. Second, it must learn a communication protocol: a query-conditioned readout that a frozen specialist model can interpret without smuggling the answer back in as text.

• Adding organs before closing one causal loop increased surface area without solving credit assignment.
• The scope-slot reranker was the only component that survived organ triage, and it survived explicitly as a retrieval baseline.
• Research 18–21 now form a dependency-ordered program instead of a collection of loosely coupled organs.

Research 18 tested consolidation into LoRA weights as a comparator. Its early 0.333 result was invalid or inconclusive because corrected responses entered training, holdout information overlapped the tuning path, the holdout was revisited, and only two seeds ran. The repaired five-seed diagnostic found four positive student deltas, but the teacher missed the unchanged gate: 0.1765 below 0.2. The later run is scientifically blocked, not accepted.

Research 19 reached valid evidence only on its fourth attempt: three earlier runs failed model resolution, provenance or feature scaling, and artifact collection. The clean run proved a nonzero text-oracle ceiling on DEV, then failed the learned latent articulation gate. Research 20 is now the core premise test. Four canonical INT8 checkpoints are verified and the fifth seed retry was still running at the live executor check; the meta-test remains sealed.

R18

BLOCKEDTeacher DEV delta 0.1765 < 0.2; holdout mean 0.2667 is diagnostic only.

R19

BLOCKEDText oracle ceiling 0.357143; learned latent articulation gate did not pass.

R20

4 / 5 checkpoints verifiedFinal d1q8 retry running at 23:33 UTC; scientific meta-test remains sealed and unsigned.

R21

SPECIFICATION ONLYDepends on a positive causal-state result from R20.

Oczy no longer asks whether enough handcrafted memory organs can make an LLM appear adaptive. It asks whether a small meta-trained cortex can acquire an unseen rule online, preserve it in fixed-shape neural state, delete the lesson, causally control a frozen language organ, and lose the behavior when that state is zeroed or swapped.