# One rule, nine presentations, three decision policies

The input corpus and protocol were published before the original run in wiki
commit `9717612`. See `PROTOCOL.md` for the frozen question, execution order and
limits. Its status describes the time of freezing; `observations.json` and
`audit-summary.json` now contain all 27 original outcomes. No original row was
repeated, no model was fitted, and no weights were downloaded for this study.

`raw/` contains byte-preserving gzip copies of all 27 original CLI JSON outputs,
per-command status (including original stdout/stderr hashes), and OS time output.
Source inputs use exact int64 JSON literals. The derived summary keeps counts and
model identities rather than re-serializing the large integer inputs as floats.
The source and its caller cases remain in `corpus/`.

The clean compiler source was `2e0e34f880baa2681ec48e22a5318e62739dcff9`,
Go1.27.2 with SDK0.2.39. At publication of this observation, the compiler adapter
is awaiting its public PR; the released0.6.27 binary has SDK0.2.37 and cannot run
the two new model formats. The weights themselves are already in the public
[SDK0.2.39 release](https://github.com/kimjooyoon/gooo-decision-runtime/releases/tag/v0.2.39-experimental).

## Read-only reproduction

With Go1.27.2, jq, gzip and shasum installed:

```sh
sh verify.sh
```

This verifies file checksums, reconstructs the original stdout bytes, checks
their hashes against each retained status, and reproduces both derived JSON
reports. It invokes neither a decision model nor a constructed program.
`collect.go` preserves every planned row, including an unreadable output if one
exists. `audit.jq` counts incomplete and unobserved first attempts separately.

## Interpretation

- All 27 invocations completed and each satisfied the same ten native caller
  cases after bounded construction. This is one rule family, not 270 tasks.
- Ordered-model first success was3/9 presentations, all three Korean forms;
  deterministic and interaction first success were0/9 each. The three local
  variable forms had identical model feature hashes and logits per wording.
- The ordered model needed4 candidates for the basic English description,1 for
  Korean and2 for the English paraphrase. Interaction needed2 and deterministic3
  for every presentation. Both models' scores changed with wording, even though
  interaction kept the same first mask.
- Basic English ordered first proposals satisfied5/5 outputs but only1/2 explicit
  intermediate Boolean conditions. Output agreement and declared-path agreement
  have separate denominators.
- Model inference medians were36.917µs (ordered) and96.459µs (interaction), nine
  observations each. Whole-command medians were0.757s and0.779s, versus0.758s for
  deterministic construction. Whole-command CPU medians were0.47s,0.46s and0.46s
  respectively; RSS medians were83.16,83.59 and83.22MiB.

`environment-note.txt` records concurrent full-repository regression work on the
host. CPU seconds count attributed user/system CPU time, not instantaneous host
utilization; RSS is the OS command measurement, not model-array size. Caches and
host load were uncontrolled. The observations do not establish speedup, general
Korean/English performance, calibrated probability or customer savings.

Inspect the public summary page for the next language/model-input question.