# Read the two choices separately

This post-hoc aggregation reads the original canonical-pooling records published
at SDK source `3e579d197aa46f20186fd621c2b60d75f0d67b61`. It performs zero fits,
predictions and candidate executions. It does not independently rerun the oracle.

The standalone Go reader uses only the standard library. Supply the retained
`studies/canonical-pooling-20261011/result` directory from that SDK source:

```sh
GOWORK=off GOTOOLCHAIN=go1.27.2 go run main.go /path/to/result > inspected.json
```

It requires the exact compressed source/search SHA256 values from the original
manifest, 256 unique source records and 768 unique source/mode records. It checks
that bit0 means comparison order and bit1 means branch layout, that each first
actual attempt equals the model's saved proposal, and that recorded candidate
acceptability agrees with all authored output/condition flags.

`FirstOutputComplete` means all eight authored output rows pass for that first
candidate. `FirstConditionComplete` means all three authored condition rows pass.
The two bit-match counts compare each saved choice with the unique acceptable
mask in that document. All 256 retained documents have one acceptable mask;
the four masks are balanced within every reported cohort. Rows grouped by family
are subsets of the same documents, not additional observations.

All nine cohort/mode document counts and first-completion counts were compared
with the original retained report and matched. The saved report's authored and
probe first-completion counts also agree. The reader does not treat that agreement
as independent all-input correctness. Static analysis passed. File identity
checks and Go's typed integer parsing avoid silently grouping another dataset.

The result is descriptive: 36/128 joint first completions in the extreme model
coexist with 64/128 matches for each individual choice, also observed for mean
pooling and fixed order. This does not identify an optimization or representation
cause. Training-record branch matches improve from32/64 to48/64, while comparison
matches stay33/64. A subsequent experiment should distinguish those targets and
keep its new data/protocol separate from this inspected corpus.