The core bet: learn how to learn

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.

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.

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.

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. At the 23:33 UTC live check, the systemd scheduler and Kaggle both reported it running. Four canonical checkpoints were runtime-manifest verified; all 90 DEV calibration shards remained gated on the fifth.

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

2 recovered + 1 retry actived0h and d3h verified; d1h failed; d1q8 running.

Pool snapshot

1 active / 4 availableAccount marked degraded because the preserved d1h failure remains in state.

Calibration

90 shards gatedCannot dispatch until all five checkpoints validate and publish.

Meta-test

BLOCKEDNo sealed seed inspected; no signoff requested or granted.

1. 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.

2. 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.

3. 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.

4. 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.

5. 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.

6. INT8 wave 3RUNNING AT LIVE CHECK

   Tried

   Run the remaining seed under durable systemd scheduling.

   Observed

   Remote and local scheduler state both running.

   Why

   The last canonical seed still needs a valid checkpoint.

   Response

   Automatic orchestration remains gated on its verified artifact.

   What it means

   No calibration or scientific verdict yet.

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.