# Read the original condition representation

This post-hoc reader inspects the original 256 source records from the frozen
canonical role-learning study. It performs zero training, predictions, candidate
executions or independent correctness-oracle executions. It does not change the
three original models or construction results.

Pinned source payload:

- [SDK result revision](https://github.com/kimjooyoon/gooo-decision-runtime/tree/afe8effe60ab708f4d9c3d53642d1d531f9fe99a/studies/canonical-role-learning-20261011/result)
- Uncompressed `sources.jsonl` SHA256: `4c38d8ce1aeb101e4e464a2ff7c2e3eaec9a9c61272d12e21f63daeb497c3dfa`

The reader rejects a different payload. Run it with Go 1.27.2 and the retained
`sources.jsonl.gz` as its sole argument. It reads only standard Go libraries.
The original output is `report.json.gz`; `SHA256SUMS` covers the reader, this
explanation and the compressed report.

It pairs each condition-goal 0 document with goal 1 while retaining its other
specification fields. The 32 training and 64 related pairs must have equal
canonical source arrays and output-case arrays. The 32 excluded affine pairs also
change output goals and are reported separately. Pairs reuse templates and
representations and overlap prior paired analyses.

All 128 pairs change raw condition rows. All have equal unweighted means of the
32 raw condition fields. The 11 existing signed association fields preserve a
change: their largest coordinate difference is approximately 1/192 (0.00520833)
after arithmetic averaging, in this corpus. Every row's 32 original fields is
scaled by 8 inside the actual encoder; that common scaling preserves equality.

These arithmetic means are **before the learned row encoder**. The actual model
applies a learned affine transform and tanh to each row before pooling, and also
uses source interactions. Its retained logits already differ across these pairs.
Therefore this result establishes neither a learned-state collision nor a causal
explanation of the weak first-choice results. It locates a feature-representation
question that a new frozen comparison could examine.

For the positive threshold 167 example, both goals supply inputs 166/167/168 and
have one true and two false expectations. Their raw column averages agree. The
input-byte × expected-polarity association changes from `(166-167-168)/768` to
`(-166-167+168)/768`, a difference of `4/768`. This arithmetic describes the input
encoding; it is not a calibrated probability, gradient or measure of model quality.