Head-only activity: selected local result

Head-only activity: selected local result

The research branch passed 39 local regressions, including the original frozen
90-example/350-update Task A replay and 13 new scope tests. Source verification
passed for all 34 recorded files. The numerical and original full-training
bodies remain byte-identical to the extraction baseline.

The separately predeclared quadrant activity used an engineered feature basis,
64 training points, 100 disjoint evaluation points and three fixed seeds. Both
fitted arms cloned the same parent, using 200 iterations, batch size 16 and
learning rate 0.01. No seeds, data, recipe or thresholds changed after results.

Each head arm passed its predeclared accuracy-gain and loss-reduction thresholds.
Head-only cross-entropy fell by 0.7294, 0.7268 and 0.7272 respectively. The full
arms achieved lower final loss than head-only. No speed comparison is qualified.

Head-only updated 20 policy-head parameters and preserved the 37 frozen
parameters, frozen moments, value predictions and active parent checkpoint.
The tests additionally preserve nonzero frozen momentum, signed zero and array
identity. Receipts join parent/result, binding, encoded data and sampler state.
Canonical JSON digests normalize signed zero, so those direct tests remain
necessary beside digest comparisons.

The selected profiles finished within their 30/180/120-second wall deadlines.
The activity retained 111,849 bytes in 11 files, below its predeclared 1 MiB
post-run file limit. This is an observed retained-output check, not an operating
system allocation quota or a whole-process memory/CPU bound. Execution occurred
on one local macOS arm64 runtime with Node 24.5.0 and Python 3.14.5.

The exact selection (../evidence/head-only-2026-10-05/VALIDATION-SELECTION.json),
validation receipt (../evidence/head-only-2026-10-05/VALIDATION.json),
activity receipt (../evidence/head-only-2026-10-05/ACTIVITY-RECEIPT.json) and
stdout/stderr logs are retained. These are coordinator-executed results;
independent source review preceded them. An
independent readback (../evidence/head-only-2026-10-05/INDEPENDENT-READBACK.json)
verified all nine persisted checkpoints and reconstructed held-out predictions,
class counts and loss from their weights using separate mathematical code.
This was result verification, not a training rerun. A genuine read-only
Claude source review (../evidence/head-only-2026-10-05/CLAUDE-SOURCE-REVIEW.md)
accepted the source with limits and no blocking defect.

Claude identified regression gaps for staged bound failures, non-writable head
state and tensor aliasing at optimizer creation. The initial 39-test run did not
exercise those specific cases. A separately selected follow-up added three test
declarations with eight cases; all 16 tests in train-scope.test.mjs passed.
These assert that staged bound failures and non-writable state refuse before
weights, moments or step change, and that valid-shaped aliases refuse at
optimizer creation. Production sources stayed unchanged; removing the additions
restores the original 13 tests and helpers byte for byte.

The follow-up selection (../evidence/head-refusal-2026-10-05/VALIDATION-SELECTION.json),
receipt (../evidence/head-refusal-2026-10-05/VALIDATION.json), raw logs and
independent readback (../evidence/head-refusal-2026-10-05/INDEPENDENT-READBACK.json)
are retained with a capture inventory (../evidence/head-refusal-2026-10-05/CAPTURE.json).
This closes Claude's specific F2 coverage gap. Only the selected 16-test file ran
in the follow-up; the current 42-test full suite was not rerun. Independent
readback verified artifacts and source, without executing the tests again.

The canonical JSON example restrictions and callback receiver behavior are
documented in the training contract. Prose updates after the initial execution
have a separate provenance record; production sources and the activity protocol
stayed fixed. The follow-up test change has its own selection and results above.

The activity capture (../evidence/head-only-2026-10-05/activity/CAPTURE.json)
also retains the three parents, six fitted inference checkpoints and exact
training/evaluation data. These weights serve the synthetic quadrant task;
they do not carry repair or security skill. To reproduce on a new output path:

npm run verify:source
npm test
mkdir -p artifacts
node scripts/verify-head-only-activity.mjs --output "$PWD/artifacts/head-only-new-run"

Preserve existing output directories. The original protocol and its thresholds
remain fixed; a fresh result on another runtime is a separate observation.

This demonstrates selective readout learning for the authored quadrant task.
It does not establish representation discovery, learned-module transfer,
retention of changed policy behavior, scientific confidence, useful network
repair or hostile-agent security. The exact rule remains a perfect, simpler
reference. Historical ANN06/ANN07 failures, full ANN08/ANN09 campaigns and
container containment Q0-Q5 remain separate and unfinished.