# Retained canonical learning records: input visibility and loss

Read-only audit, 2026-10-11 KST. No fitting, prediction, candidate execution,
native compilation, or new performance observation. This directory contains
small derived records only; the original training material stays in the SDK.

Original source commit:
[`8468dea5588a3f5f19167b17eec8e3fe91dd6a49`](https://github.com/kimjooyoon/gooo-decision-runtime/tree/8468dea5588a3f5f19167b17eec8e3fe91dd6a49/studies/canonical-contract-learning-20261011/result).
`source-hashes.txt` identifies the three original compressed files.

## What is counted

`visibility.json` groups the 64 retained training documents by exact JSON values
of their saved canonical source arrays, output cases, and condition cases.
The four singleton acceptable-candidate bitsets each occur 16 times.
For each input group, the largest label count is the most a constant choice for
that group could satisfy. Summing these counts gives 16/64 for source only,
32/64 for either one of the two requirement streams, and 64/64 with both.

These counts are finite distinguishability bounds for these constructed records.
They are not trained-model accuracy or a held-out estimate. The audit does not
check what information survives encoding, pooling, or the hidden layers. Exact
full-input uniqueness does not prove the current model can learn the labels.
Original canonical first-choice training success remains 18/64.

`canonical-loss.json` and `legacy-loss.json` extract six checkpoints from the
original 2,000-entry histories. The producer records each mean loss before that
epoch's update. The final entry is not a newly evaluated post-update loss.
`last_1000_improvement` subtracts the epoch-2,000 value from epoch 1,000.
Small late decreases narrow the next investigation; they do not identify a cause.

## Reproduce by reading existing Git objects

From an existing SDK checkout that already contains the source commit, set
`audit_dir` to this downloaded directory. These commands only read retained data.
They do not call the study producer, the compiler, or a model.

```sh
set -euo pipefail
source_commit=8468dea5588a3f5f19167b17eec8e3fe91dd6a49
source_dir=studies/canonical-contract-learning-20261011/result
audit_dir=/absolute/path/to/canonical-learning-audit-20261011

while read -r expected source_file; do
  actual=$(git show "$source_commit:$source_file" | shasum -a 256 | cut -d ' ' -f 1)
  test "$actual" = "$expected" || exit 1
done < "$audit_dir/source-hashes.txt"

git show "$source_commit:$source_dir/training.json.gz" |
  gzip -dc | jq -f "$audit_dir/visibility.jq" |
  diff - "$audit_dir/visibility.json"

for mode in canonical legacy; do
  git show "$source_commit:$source_dir/history-$mode.json.gz" |
    gzip -dc | jq -f "$audit_dir/loss.jq" |
    diff - "$audit_dir/$mode-loss.json" || exit 1
done
```

All three comparisons should be empty. The original training study includes
broader saved-record checks; this audit adds only input grouping and loss extraction.