Field note 06 · published 22 July 2026

The core bet: learn how to learn

Research 20 and Experiment 09 turn the missing protocols into a trainable system: a fixed-shape fast/slow cortex, a learned writer and consolidator, and a learned latent coupler into a frozen language organ.

Reading time: 16 min
Evidence cutoff: 22 July 2026 · R20 calibration v5 invalidated, width-3 v6 rerun active
Covers: Research 20 · Experiment 09 · meta_cortex/v2 · Campaigns 959e114 → v6

Verdict

All five developmental checkpoints are verified and published. The first calibration fanout ran but its scientific aggregation was invalidated by a partition-dependent donor selection. A corrected width-3 v6 rerun with partition-independent donor selection is active. The scientific claim remains deliberately untested.

Architecture: state that cannot grow with experience count

The cortex projects frozen LM features into 64 dimensions and maintains two 64×64 matrices. Fast state resets at task start; slow state persists after consolidation. A learned writer emits key, value, learning rate, and decay for a fixed outer-product update. A learned consolidation gate absorbs fast state into slow state, then clears fast state.

A query-conditioned reader maps slow state and the current request into a fixed-width soft bank consumed by the frozen Qwen language organ. During meta-test there is no gradient step and no optimizer update: only the frozen learned update equations may change cortex state.

Cortex state: 64×64 fast + 64×64 slowFixed shape; no per-episode slot list.
Feature bank: 3 × 896Fixed-width articulation surface in the DEV smoke.
Developmental θ: 207,364 parameters829,456 bytes in the v1 smoke artifact.
Frozen organ: Qwen2.5-0.5B-InstructHashed before and after; only cortex parameters train.

The task families prevent simple answer storage

Meta-training spans contextual remapping, rule transformations over unseen operands, and finite-state behavior across multiple turns. Complete rules, mappings, graphs, and output assignments are split across train, validation, and test; paraphrases of a rule cannot cross the boundary.

The one-shot meta-test includes no-update, random-writer, shuffled-feedback, zeroed-state, swapped-state, label-prefix, retrieval, and oracle conditions. Same-rule retention is insufficient: transfer and composition must improve, feedback meaning must matter, and the right cortex state must causally control the answer.

Why the instrument moved from v1 to INT8 v2

The human-approved cutover replaced the FP32 organ with TorchAO W8A32 weight-only INT8. Dynamic A8W8 was rejected because its forward path did not preserve the gradient needed from the frozen organ back to the soft bank. W8A32 kept FP32 activations and finite nonzero soft-bank gradients.

The cutover did not preserve old checkpoints as if they were comparable. It created a new instrument namespace, organ hash, runtime manifest, developmental checkpoints, and calibration requirement. Meta-test remained blocked.

The July 15–16 training campaign

Campaign 959e114 launched five canonical developmental seeds on private, internet-off Kaggle CPU kernels at a clean commit. The first wave used 30 tasks per family and two outer steps. Two checkpoints completed—one collected normally and one manually recovered after the scheduler timed out. Three jobs hit Kaggle’s 12-hour CPU limit before producing checkpoints.

The 15-task replacement wave recovered d0 and d3; d1 again exceeded the provider window. A uniquely named d1q8 retry changed only developmental task count to eight and completed. All five canonical checkpoints are now runtime-manifest verified and published as an immutable private dataset. The 90-shard DEV calibration fanout was then dispatched.

First wave: 2 / 5 checkpointsd2 complete; d4 complete and manually recovered.
Infrastructure null: 3 / 5 hit 12 hA compute-window constraint, not a scientific verdict.
Replacement waves: All 5 verifiedd0h, d3h, and d1q8 completed; checkpoints published as opaque-archive dataset.
Calibration v3: INFRASTRUCTURE-INVALIDAll 90 jobs failed before experiment code: Kaggle auto-extracted the registered archive.
Meta-test: BLOCKEDNo sealed seed inspected; no signoff requested or granted.

DEV v1INFRASTRUCTURE FAILURE
Tried
Load Qwen in the offline kernel.
Observed
Local model resolution failed.
Why
No local snapshot resolver under HF_HUB_OFFLINE.
Response
Added fail-closed local model resolution.
What it means
Loader repair only.
DEV v2INFRASTRUCTURE FAILURE
Tried
Train cortex parameters through frozen-organ features.
Observed
Inference tensors could not participate in backward.
Why
torch.inference_mode created incompatible feature tensors.
Response
Kept the organ frozen under no_grad while preserving usable feature tensors.
What it means
Gradient-plumbing repair only.
DEV v3INFRASTRUCTURE SUCCESS
Tried
Run one outer step, causal controls, deletion, and audit.
Observed
Exit 0; frozen hash stable; trace count 0; DEV score 0.0.
Why
The minimum smoke was intentionally underpowered for a claim.
Response
Recorded the zeros as mechanism diagnostics, not a refutation.
What it means
The DEV path runs; the hypothesis remains untested.
INT8 wave 1PARTIAL EXECUTOR SUCCESS
Tried
Five seeds, 30 tasks per family, two outer steps.
Observed
Two completed; three exceeded 12 hours.
Why
Workload exceeded Kaggle’s CPU execution window.
Response
Preserved failures and reduced only developmental task count.
What it means
Provider constraint, not scientific null.
INT8 wave 2PARTIAL EXECUTOR SUCCESS
Tried
Retry three seeds at 15 tasks per family.
Observed
Two completed; d1 exceeded 12 hours again.
Why
One seed still ran outside the execution envelope.
Response
Created the unique d1q8 eight-task retry.
What it means
Four of five prerequisites are verified.
INT8 wave 3EXECUTOR SUCCESS
Tried
Run the remaining seed under durable systemd scheduling.
Observed
d1q8 completed; checkpoint runtime-manifest verified.
Why
Eight tasks per family fit the 12-hour window.
Response
Published all five checkpoints as an immutable opaque-archive dataset.
What it means
All training prerequisites are satisfied; calibration may proceed.

Calibration: transport repair, bounded fanout, and a partition invalidation

The first calibration dispatch (v3) was infrastructure-invalid before any experiment code ran: Kaggle auto-extracted the registered instrument archive, but the bootstrap required the whole archive as a regular file for SHA-256 verification. The repair introduced opaque .tar.gz.bin transport names with whole-archive verification and explicit r:gz extraction, and republished source, instrument, and checkpoint datasets under new immutable identities.

The v5 bounded fanout then ran 116 jobs covering every registered developmental-seed and task cell. A one-task canary succeeded in 2h50m; four-task shards succeeded in 8.6–11.8 hours, but two crossed the 12-hour limit. The failed ranges were split into immutable two-task replacements. At its peak the fanout reached 29 successful and 6 provider-timeout four-task attempts.

A subsequent partition audit invalidated the v5 collection for scientific aggregation. The C6 state-swapped donor was selected from the next task inside the current shard, so changing shard boundaries changed the donor state and the scored state-addressing delta. One-task shards omitted C6 entirely, while endpoint aggregation requires all six C1–C6 conditions. Every v5 shard remains valid infrastructure and runtime evidence, but no v5 metric or scientific verdict may be derived from it.

The human-authorized repair was a version fix: make donor selection partition-independent by using the canonical next task within each full frozen validation family, then perform a full width-3 DEV rerun. The corrected v6 campaign contains exactly 150 three-task jobs covering all 5 × 90 registered cells. Its width-3 canary is pending behind the five legacy kernels allowed to drain. Meta-test remains blocked and was not accessed.

Transport v3: INFRASTRUCTURE-INVALIDAll 90 jobs failed before experiment code; archive auto-extraction conflict.
Transport v4: REPAIREDOpaque .tar.gz.bin transport; whole-archive SHA-256 verification; republished datasets.
Fanout v5: 29 succeeded, 6 timed outFour-task width has no deterministic margin under the 12-hour limit.
v5 aggregation: INVALID FOR SCIENCEPartition-dependent C6 donor selection; shard width changed the scored delta.
Repair: Partition-independent donorCommit a8c98d6; canonical next task within full validation family; 89 + 4 focused tests passed.
v6 rerun: 150 three-task jobsWidth-3 canary pending; covers all 5 × 90 cells exactly once.

Cal v3INFRASTRUCTURE-INVALID
Tried
Dispatch 90 calibration shards with the registered archive.
Observed
All 90 jobs failed before experiment code.
Why
Kaggle auto-extracted the archive; bootstrap required a regular file for SHA-256 verification.
Response
Introduced opaque .tar.gz.bin transport and republished datasets.
What it means
No scientific evidence; a transport defect, not a hypothesis result.
Cal v4 canaryINFRASTRUCTURE NULL
Tried
Run a five-task canary under the repaired transport.
Observed
Archives extracted correctly; runtime manifest matched; killed at 12-hour limit before writing a shard.
Why
Five calibration tasks exceeded the provider window.
Response
Reduced to one-task canary, then four-task bounded fanout.
What it means
Transport repaired; shard width still needed calibration.
Cal v5 fanoutPARTIAL EXECUTOR SUCCESS
Tried
Run 116 jobs covering every registered cell at four-task width.
Observed
29 succeeded, 6 timed out; two failed ranges split into two-task replacements.
Why
Four-task width has no deterministic margin under 12 hours.
Response
Preserved failed attempts and created immutable narrower replacements.
What it means
Executor evidence only; no scientific aggregation yet.
v5 partition auditINVALID FOR SCIENTIFIC AGGREGATION
Tried
Aggregate v5 shard metrics into DEV calibration.
Observed
Shard width changed the C6 state-swapped donor and the scored state-addressing delta.
Why
Donor selection was partition-dependent: the next task inside the current shard.
Response
Human-authorized version fix; partition-independent donor selection; full width-3 rerun.
What it means
Every v5 shard is valid runtime evidence, but no v5 metric or verdict may be derived.
Cal v6 width-3CANARY PENDING
Tried
Run 150 three-task jobs with partition-independent donor selection.
Observed
Width-3 canary pending behind five draining legacy kernels.
Why
Corrected campaign awaits canary promotion before fanout.
Response
Durable systemd chain verifies provenance, runtime, CPU-only hardware, and C1–C6 presence before promotion.
What it means
No calibration or scientific verdict yet.

What would count as a result

An ACCEPT requires positive adaptation, transfer, and composition with confidence intervals excluding zero; wins over random and shuffled-feedback controls; causal dependence on nonzero, correctly addressed cortex state; trace survival and specificity inside frozen margins; and passing fixed-width, trace-deletion, no-text, and frozen-organ audits.

Until calibration and sign-off are complete, a stable hash, a falling training loss, a checkpoint, or a nonzero DEV intervention is an engineering observation. The project’s central claim remains open.

Source trail

These are the primary repository artifacts used for this note. Status labels follow the current ledger and campaign records. The complete evidence ledger publishes every dated classification.

oczy/research/20-meta-trained-cortex-frozen-language-organ.md Read raw →
oczy/experiments/09-meta-trained-cortex-frozen-language-organ/README.md
oczy/experiments_logs/2026-07-16_campaign_959e114.md Read raw →
oczy/src/oczy/experiments/meta_cortex/
oczy/infrastructure/kaggle/RESEARCH_GUIDE.md