20 — Meta-trained plastic cortex over a frozen language organ

Pre-registered 2026-07-09 (human-authorized before implementation). Agents running this experiment MUST NOT edit this spec. Deviations are reported as deviations. The concrete build-and-run specification is experiments/09-meta-trained-cortex-frozen-language-organ/.

Problem

Oczy's cortex had mutable state but no learned learning algorithm. A correction embedding was accumulated through a hand-authored Hebbian rule and projected back into the LM as if "a representation of the experience" were automatically "the direction that makes the model use the experience." S1.3, S1.4, S2.1, and S2.4 jointly refuted that shortcut.

Research/19 adds direct gradient training and distinguishes a label-text store from latent control, but it still optimizes a fresh cortex directly on the task being evaluated. It does not test whether the cortex has learned a reusable algorithm for acquiring new memories and behavior from experience.

The missing developmental layer is an outer loop. Across many task episodes, it must train:

• how experience and feedback write fast cortex state;
• how current context addresses that state;
• how fast state consolidates into slow persistent cortex weights;
• how a fixed-width latent control signal drives a frozen language organ; and
• how learning remains specific instead of overwriting unrelated behavior.

Retrieval remains a mandatory external baseline but is disabled in the primary cortex condition. The experiment isolates the raw capability of cortex state to carry post-experience information.

Hypothesis

H-META-CORTEX: after developmental meta-training across a distribution of learning episodes, a cortex with learned write, read, and consolidation rules can adapt to a previously unseen task from correction, consolidate the change, delete all raw traces, and cause a frozen language organ to execute the learned rule on held-out and compositional probes through a fixed-width latent channel.

The hypothesis requires all of the following:

1. the task family and rule are unseen during developmental training;
2. the language-organ parameters remain bit-identical;
3. no correction, label, exemplar, or retrieved content enters the answer path;
4. no optimizer/backpropagation runs during the evaluation episode;
5. the learned cortex state causally controls the behavior; and
6. semantically wrong or shuffled feedback does not produce the same gain.

Architecture

Frozen language organ

• HFDriver with Qwen/Qwen2.5-0.5B-Instruct is the initial substrate.
• Its tokenizer, embedding, transformer, and LM-head parameters are frozen.
• Hashes are recorded before developmental training, after developmental training, and after every evaluation seed.
• It provides perception features and articulation only. A stronger frozen model may be introduced only in a new pre-registered spec.

Cortex

The cortex owns every trainable or mutable component outside the frozen organ:

• developmental parameters theta, learned in the outer loop;
• fast state F_t, updated after each experience;
• slow state S_t, changed only by consolidation;
• learned write rule U_theta(F_t, observation, feedback, outcome);
• learned consolidation rule G_theta(S_t, F_t);
• query-conditioned read rule R_theta(S_t, F_t, query_feature); and
• latent articulation coupler P_theta, which emits a fixed-width soft embedding/KV bank for the frozen language organ.

At evaluation time theta and P_theta are frozen. The learned rules may change F_t and S_t; standard gradient descent, target-token optimization, and LM updates are forbidden.

Information boundary

The primary condition may persist only serialized cortex state. It may not persist or consult:

• raw experience or correction text;
• tokenized traces;
• target labels or expected answers;
• exemplar embeddings or nearest-neighbor indexes;
• episode IDs or task IDs;
• variable-length memory proportional to episode count; or
• an LM KV cache produced directly from the original correction.

The latent bank is recomputed from the current query and persistent cortex state. Fixed-width latent control is a readout of neural state, not a stored trace.

Developmental task distribution

The instrument is separate from eval/v2 and must be frozen as meta_cortex/v1 before the main run. It contains deterministic generators for three learning families:

1. Contextual remapping: an arbitrary symbol or word maps to different outputs under different contexts.
2. Rule transformation: a correction defines a small input-output rule that must be applied to unseen operands, not repeated verbatim.
3. Finite-state behavior: feedback changes a latent policy that must retain a goal across multiple turns.

Meta-train, meta-validation, and meta-test split by complete rules and task instances, not surface paraphrases. Meta-test rules, output assignments, and compositions must be unreachable from meta-training seeds.

Each task family supplies:

• a pre-learning probe set;
• one to five experience/feedback events;
• held-out same-rule probes;
• compositional probes combining learned operations;
• unrelated specificity probes; and
• a deterministic oracle-context form used only for the validity gate.

Instrument freeze and threshold distribution check

Before any meta-test run:

1. materialize generator version, seeds, family split, scorers, and probe counts;
2. compute SHA-256 hashes into a MANIFEST.json;
3. run no-update and repeated-run distributions on meta-validation;
4. derive the specificity equivalence margin from the observed no-update repeatability distribution;
5. freeze sample size using a power analysis on meta-validation effect sizes;
6. obtain human sign-off on the manifest, margin, and sample size; and
7. never change them without a version bump.

No threshold is selected from meta-test data.

Protocol

Phase A — oracle controllability gate

Give the frozen language organ the complete rule and worked demonstrations in text on meta-validation. It must beat organ-only zero-information performance with a 95% CI excluding zero. If it cannot express the rule when fully informed, the corresponding task family is BLOCKED, not a cortex refutation.

Phase B — developmental meta-training

Unroll full learn/consolidate/probe episodes on meta-training tasks. Optimize theta and the latent coupler through the post-learning behavioral loss plus specificity and state-size regularizers. Meta-validation selects optimizer, architecture, and stopping point. Meta-test is never observed.

Phase C — frozen-rule evaluation

Freeze developmental parameters. For each meta-test task and seed:

1. initialize empty F_0 and S_0;
2. record pre-learning behavior;
3. present experience and correction events;
4. update fast state through U_theta only;
5. consolidate once through G_theta;
6. delete all raw traces and verify count zero;
7. score same-rule, transfer, composition, and specificity probes; and
8. repeat causal interventions with cortex state zeroed, swapped, and feedback shuffled without rerunning the learning episode.

Use at least 5 developmental seeds and at least 5 evaluation seeds. Exact task counts are frozen by the pre-run power analysis.

Matched conditions

C1/C3 isolates online state change. C2/C3 isolates developmental meta-training. C3/C4 isolates feedback. C3/C5 and C3/C6 isolate the causal state path.

Primary metrics

1. adaptation_delta — post-learning minus pre-learning accuracy on unseen meta-test tasks.
2. transfer_delta — C3 minus C1 on held-out inputs governed by the learned rule.
3. composition_delta — C3 minus C1 on novel compositions of learned rules.
4. feedback_semantics_delta — C3 minus C4.
5. causal_state_delta — C3 minus C5.
6. state_addressing_delta — correct C3 state minus C6 swapped state.
7. trace_free_survival — post-deletion score minus the score immediately before deletion, evaluated against the frozen equivalence margin.
8. specificity_delta — change on unrelated tasks relative to C1.
9. persistent_bytes, update latency, articulation latency, and behavior_delta_per_byte reported as a resource table, not a lone score.

All behavioral metrics report per-family estimates, pooled estimates, seed and task counts, and 95% CIs. Seeds do not substitute for independent task rules.

Acceptance

Accept H-META-CORTEX only if:

• adaptation_delta, transfer_delta, and composition_delta are positive with 95% CIs excluding zero;
• C3 beats C2, C4, and C5 on their matched primary comparisons with 95% CIs excluding zero;
• the correct state beats the swapped state;
• trace-free survival and specificity remain inside their DEV-frozen equivalence margins;
• fixed-width and deletion audits pass; and
• all frozen-language-organ hashes match.

Refute if validity gates pass but any acceptance condition fails. A C7 or C8 win does not rescue C3 and cannot be reported as cortex metabolism.

Blocked if the oracle controllability gate fails or if the instrument lacks human-approved distribution checks and manifest freeze.

Kill and interpretation rules

• If C1 matches C3, the online update is unnecessary.
• If C2 matches C3, the learned update rule did not earn its role.
• If C4 matches C3, the system is reacting to generic update magnitude rather than learning from feedback content.
• If C5 matches C3, the behavior does not causally depend on persistent cortex state; inspect leakage before any rerun.
• If only same-rule retention passes but composition fails, record the result as neural association storage, not learned behavioral dynamics.
• No task-specific hyperparameter rescue is allowed on meta-test. It belongs in a new spec.

Reporting

Log developmental curves separately from the one-shot meta-test report. The final report includes the frozen manifest, task-family split, oracle gate, condition tables, causal interventions, trace and hash audits, resource table, exact commands, and all nulls. Write to experiments_logs/<date>_s20_meta_trained_cortex.md.

Dependency

Research/20 is the core cortex experiment. Research/21 may begin only after H-META-CORTEX accepts on at least one rule family and the latent interface passes its causal-state audit.