Checkpoint decision: portable-core-0.1.0

Checkpoint decision: portable-core-0.1.0

Why this checkpoint

The reusable numerical core, task harness, two adapters and typed workflow have a
small dependency closure. They can be copied without importing the larger game,
platform workspace or evolving campaign controllers. Their regression checks
exercise concrete numerical and operational behavior rather than relying on
research status documents.

This is an extraction checkpoint, not a declaration that the ANN research has
finished. Validation can proceed here while research continues separately.

Exact import

SOURCE-CHECKPOINT.json records every imported file's origin, source digest,
destination digest and size. The original research source was not a Git checkout;
there is no source commit to invent. The new Git repository provides commit
identity from this checkpoint onward.

Production core, adapters, composition, runtime and contracts are copied
byte for byte. Two verifier defaults change: the frozen reference path now
resolves inside this repository. One invalid-mask test now uses a single owned
harness and checks the mask-specific error, rather than rejecting a foreign
model before reaching the mask. Other tests are byte-identical. The original
baseline manifest remains intact.
The verbatim semantic extraction contains a final blank line required by its
source-equivalence check. A path-specific Git whitespace attribute preserves
that byte identity; other whitespace checks remain enabled.
fixtures/ann01/FIXTURE-SUBSET.json names the ten included files and the omitted
historical leaves. Verification claims cover consumed leaves, not the entire
historical capture. The included weights are synthetic Task A reference weights,
not qualified operational repair models.

Evidence boundaries

Fresh checkpoint results are stored under evidence/. Historical pass counts are
not imported as current results. Detailed generated traces stay under ignored
artifacts/ locally; compact tracked receipts bind their source and outcomes.

Retained research limitations

• The previous ANN06 260-mission cohort failed quality evaluation. All five

ANN07 A/B exposed-anchor retention evaluations failed. The selected ANN06 B
selection panel had 0/9 positive restarts; that denominator does not describe
every seed's primary trajectory or ANN07's exposed-anchor panel. Two ANN06
primary trajectories had four correct alpha restarts, while beta/gamma were
always missed. See the retained historical audit context (../evidence/historical-retention-context-2026-10-05/CAPTURE.json).

• Historical Task B fresh neural routing achieved 0/24 correct versus the rules

baseline's 24/24 and round-robin's 8/24. A mechanical verification PASS can still
faithfully report a poorly performing policy.

• The E3 native fit completed 5,250 updates but this does not qualify learning.

A later cold readback exceeded its 256 MiB RSS limit: 280,870,912 bytes versus
268,435,456. That attempt is failed/spent, not silently reset by this repo.

• The larger TRAIN set has 4,662 WAIT rows among 4,680 rows. Overall accuracy

alone is misleading; the 18 positive rows are not established as independent
incidents.

• Full ANN08 40-arm and ANN09 two-cycle campaigns remain unrun. Real container

containment Q0-Q5 remains unqualified.

• The corrected codec114 repair has a source-only review acceptance with limits;

no runtime qualification is transferred here. Those experimental sources are
excluded from this baseline.

These are historical research results reported by the research workspace, not
rerun claims. Their retained source record is the 2026-10-05 coordinator feedback
and the original ANN status/campaign artifacts. Nothing in this checkpoint turns
these negative results into passes.

Known component constraints

These constraints describe the immutable extraction tag. The later head-only
research branch has its own contract and
selected local results.

The shared trainer updates every bound role's hidden layers and policy head.
The value head's parameters receive no supervised updates, but its outputs can
change as shared hidden features change. Checkpoints retain inference weights,
not optimizer moments, RNG or data-stream position. Resumable training and
within-model selective training need new contracts and implementations.

Model ownership is local object ownership. Canonicalization assumes trusted
local objects or parsed JSON; it has no hostile-object accessor, cycle, depth or
byte-size isolation. Current numerical array validation can accept sparse array
holes before downstream inference rejects nonfinite outputs. Imported fixtures
are dense. Strict hostile-input parsing and dense-tensor validation are separate
hardening tasks; do not advertise these APIs as a secure external model service.

Advancement criteria

1. Keep the original extraction tag immutable and record every changed source.
2. Add selective training with frozen parameter/buffer/moment checks.
3. Measure a small held-out skill against rules and fresh/head-only controls.
4. Require old-skill retention and complete-workflow results before promotion.
5. Import larger research modules only with their complete dependency closure,
rights review, resource profile, tests and explicit unresolved limitations.

The first four criteria establish a useful next modular-learning checkpoint.
Production effect authority and hostile-agent containment require additional
work beyond that checkpoint.