# NeuroShard frozen learning experiment — September 12, 2026

Phase 1 result: **FAIL**. The result uses the pre-committed sealed test and
retention thresholds. Fresh cannot rescue a failing test. All 20 generation
pairs are retained in result.json and generations.html.

Execution revision: `4f18f903b738d5f2e5ba6d31434b7e0ae5e06fd7`. Selection SHA-256:
`c583aba89d2c6c18f05efb6d3c63519b5c21b442a961727ddc76f331a98e58ac`. Baseline: `7ca58fefa5977056edfbca3241b174d1376f4be3042aafb3fd768cb2d21dbc02`.
Candidate: `ecb4262c2b7d63a0bfb0d17a9dc9023af33c351645d0e6ba12f087ba6d474677`.

The model has 134,515,008 trainable parameters. The recipe was 128 SGD steps,
256 actual training windows, learning rate 0.003 and global clipping 1.0.
One physical host ran three workers. No model growth or token issuance occurred.
The public 0.4.0 chain was not changed. Duration includes evaluation/generation:
7625.029 seconds. Peak experiment size:
71850233994 bytes. Transport retries: 0.

This is a paired, teacher-forced response-loss experiment with approximate
normal confidence bounds. It does not establish continual improvement,
independent operators, economical verification, capacity growth or production
chat quality. The seed already underwent Smol-SmolTalk fine-tuning; "sealed"
means committed before this run, not proven unseen during seed training.

Artifacts at https://neuroshard.com/experiments/learning-milestone-20260912/

- result.json: every document loss, generation, step and decision.
- selection.json, frozen-plan.json, plan.json: committed inputs and constants.
- window-measurements.json, target-counts.json: per-window float-hex losses and
  response counts used to aggregate one observation per document.
- document-loss-audit.json: every paired document change and descriptive
  summaries, including regressions. This post-scoring analysis changes no gate.
- generations.html: all 20 prompt/response pairs, including empty/poor outputs.
- candidate-checkpoint.tar.gz: exact native float32 checkpoint plus tokenizer;
  contains no credentials or intermediate training history. Its internal README
  explains the native object format and inspection command.
- reconstruction-result.json: byte-identical input reconstruction in a fresh
  local experiment directory. Both runs had the same operator and host.
- execution-environment.json: CPU/runtime and resource envelope, including
  other work sharing the host. Timings are not an isolated benchmark.
- evidence-verification.json: separate arithmetic and completeness checks.
- training-completion.json, checkpoint-package.json: graph/packaging checks.
- evidence-builder.py, checkpoint-builder.py: exact local publication scripts,
  with recorded paths for provenance, not protocol or client entry points.
- verification-note.md and integer_* files: separate synthetic arithmetic
  probes, including negative cost results; not the current FP32 verifier.
- checkpoint-delta-note.md, checkpoint-delta-result.json and
  checkpoint_delta_probe.py: a read-only lossless transport/storage probe on
  one real transition. Neither auxiliary probe changes this experiment's gate.
- token-loss-diagnostic.json and document_token_probe.py: 12 read-only forwards
  on two documents selected after the failed primary score, diagnosing content
  and end-token changes. This is not a new evaluation set or quality gate.
- plot_learning_result.py and figure-provenance.json: the source and environment
  for the standalone PDF, SVG and PNG comparison. Generate figures after the
  first evidence build and rebuild the final checksum manifest afterward.

From the execution revision, install the recorded dependencies and follow
docs/LEARNING_MILESTONE.md to reconstruct the selected source windows and train.
Large model tensors remain outside Git. Checksums detect byte changes; they do
not replace an independent replay of computation. Source model and corpus are
Apache-2.0; the checkpoint includes the upstream model card and attribution.