# Reading recorded responses to changed goals

This post-hoc reader consumes the original canonical-pooling study's 256 sources
and 768 search records. It performs zero fits, predictions or candidate executions.
Its two input files are pinned by SHA256; it checks unique source/record IDs,
source choice identities, the single acceptable mask and each actual first attempt.

The same 256 documents are paired twice: change the output goal while retaining
the condition goal, and change the condition goal while retaining the output-goal
selector. Each pairing has 128 pairs, each read in fixed/mean/extreme mode.
These are overlapping observations, not 768 new independent experiments.

Every pair must have identical canonical source arrays. Output-goal pairs also
have identical condition arrays and different output arrays. Condition-goal pairs
have different condition arrays. Their output arrays are explicitly compared:
32 affine-family pairs change both requirement streams and are named
`condition_goal_with_output_change`. They cannot isolate the condition channel.
The initial reader incorrectly required all condition-goal pairs to have equal
outputs and stopped at the first affine pair. The corrected reader preserves and
separately labels all such pairs; no original experiment was repeated.

`RequiredBitChanges` comes from the retained unique acceptable masks.
`ProposedBitChanges` comes from the retained first actual choices. Bit0 is
comparison operand order; bit1 is branch layout. `RecordedLogitChanges` counts
changed recorded option-score pairs, with no fresh forward pass. Fixed ordering
has no learned logits. `BothFirstCorrect` and `EitherFirstCorrect` count pairs;
their sum equals the number of correct first choices across the paired documents.
The checked relation and complete pair IDs/masks/scores are retained in the report.

Related evaluation has 64 condition-only pairs: every pair requires both bits to
change, but both models retain both first-choice bits in every pair. Their
recorded score pairs nevertheless change for all64. For64 output-only pairs, the
branch bit must change and comparison bit must stay. Mean changes4 comparison
bits and4 branch bits; extreme changes36 and44. Both first choices are complete
in0 mean pairs and4 extreme pairs. These observations narrow the next learning
question; they cannot identify its optimization or representation cause.

Run the read-only inspector with Go1.27.2 and the public original records:

```sh
GOWORK=off GOTOOLCHAIN=go1.27.2 go run main.go \
  /path/to/gooo-decision-runtime/studies/canonical-pooling-20261011/result
```

The original producer and weights are at SDK source
`3e579d197aa46f20186fd621c2b60d75f0d67b61`. The reader itself uses only the Go standard
library. Its static check passed. A read-only rerun for reader validation is
separate from the original one-use training and construction observation.