# Deterministic Tag Lane Reproduction

**State:** VERIFIED PUBLIC-CANDIDATE PACKAGE / NOT PUBLISHED  
**Schema:** `tag_normalization_v1` / `1.0.0`  
**Runtime:** Python 3 standard library on a POSIX system

This package reproduces the correction described in **When AI Is the Wrong Tool**. For one exact synthetic/non-sensitive schema, the correct transformation is fully specified:

1. trim surrounding whitespace;
2. lowercase each value;
3. remove whole-value duplicates; and
4. sort the result.

The implementation uses deterministic local software rather than model inference. It preserves the bounded eligibility, kill switch, redacted owner-only audit, fail-closed behavior, and frozen representative cases.

## Run

```sh
python3 -m unittest -v test_tag_lane.py
python3 reproduce.py
```

The first command runs 22 named tests. The second runs the frozen 12-case representative replay from a clean temporary state and writes a deterministic result to `evidence/latest/result.json`.

## Expected result

- Unit suite: `22/22`.
- Representative replay: `12/12`.
- Eligible transformations: `7/7`.
- Mandatory denials: `5/5`.
- Accepted lengths: every length `1–64` in the unit suite.
- Concurrent audit: `32/32` parseable records.
- Prompt/completion tokens, retries, tools, model calls, and network calls: `0`.

## Scope

This is a public derivative, not the installed router. It does not attach to ordinary prompts, process sensitive or proprietary material, select cloud models, or establish production readiness. The installed-command and rollback records remain summarized in the related case because their original evidence contains machine-local configuration and paths.

## Contents

- `tag_lane.py` — deterministic lane and bounded audit.
- `pilot.config.v1.json` — frozen public configuration.
- `fixtures/approved.json` — minimal eligible fixture.
- `fixtures/cases.v1.json` — frozen 12-case representative set.
- `test_tag_lane.py` — 22 named tests.
- `reproduce.py` — clean-state deterministic replay.
- `MATRIX_DECISION.md` — preserved model-era decision summary.
- `CORRECTIONS.md` — public-safe correction lineage.
- `evidence/latest/` — reproduced result and manifest.
- `SHA256SUMS.txt` — integrity manifest generated after verification.