# Public replay shape map

| Replay outcome    | Fixtures | Mark          |
| ----------------- | -------: | ------------- |
| Uniquely resolved |       44 | Teal circle   |
| No match          |       11 | Orange square |
| Ambiguity refusal |        1 | Rose diamond  |
| Unreplayable      |        3 | Gray triangle |
| **Total**         |   **59** |               |

![Three flat PCA views of public ledger-eval fixture shape.
Each mark has the same size; color and shape indicate replay outcome after the projection.](./_static/error-cloud.svg)

This is a compact, reproducible view of the committed public [public `ledger_eval` fixture corpus](https://gitlab.com/doering-ai/libs/tact/-/blob/main/tests/fixtures/ledger_eval/README.md) corpus.
It replaces the former mixed-population error cloud.
The old figure combined live agent lifecycle data, private fixture-derived material, and expanded local telemetry, then made a token-savings claim.
That claim and that population are retired: this figure is only a public-fixture replay benchmark.

## What is projected

The generator replays all 59 committed JSON fixtures with the current reconciliation ladder, then derives a six-column numerical matrix from edit **shape**:

| Feature                                        | Meaning                                       |
| ---------------------------------------------- | --------------------------------------------- |
| `log1p_old_characters`, `log1p_new_characters` | Size of requested old and new text            |
| `log1p_old_lines`, `log1p_new_lines`           | Old and new text line counts                  |
| `log1p_edit_character_delta`                   | Absolute character-size change                |
| `new_to_old_character_ratio`                   | Relative expansion or contraction of the edit |

The columns are centered and scaled before PCA.
Outcome, family, rung, category, case ID, all label-like codes, and snapshot measurements are excluded from that matrix.
A missing snapshot defines `unreplayable`, so zero-imputing its size would leak the display outcome into PCA.
Only after PCA does the graphic attach replay outcome as a color-and-shape label.
Axis labels give the explained-variance ratio for each principal component.

The generated JSON includes `loadings` for PC1–PC3: these are the standardized feature weights that make a component interpretable.
Compare absolute loading sizes, not signs; PCA component orientation is arbitrary.
The `explained_variance_ratios` list records how much standardized edit-shape variation each displayed component accounts for.

**Proximity means similar edit shape, not causality or correctness.** The triptych avoids a synthetic camera or depth ordering: it shows PC1 x PC2, PC1 x PC3, and PC2 x PC3 directly.

## Replay and safety boundary

"Uniquely resolved" means the public replay found one match under the current ladder.
"No match" and "ambiguity refusal" are distinct structured outcomes; the latter is an intentional refusal to guess.
"Unreplayable" means the committed fixture lacks a captured file snapshot, not that the ladder crashed.

The constrained same-intent comparison finds **0 candidates among 9 checkable same-intent pairs**.
It compares only fixtures whose later follow-up is both resolvable and highly similar enough to plausibly represent the same logical edit.
This is not a claim of zero production false heals, nor does it cover fixtures without checkable follow-up ground truth.

## Provenance, privacy, and reproduction

[`error-cloud.json`](./_data/error-cloud.json) is schema `tact.ledger-eval-shape-pca` version 3.
It binds the exact public fixture population to a deterministic SHA-256 manifest, records feature names, PCA variance/loadings, aggregate replay counts, and anonymous outcome/rung/PC coordinates.
It deliberately carries no case IDs, file paths, file extensions, old/new strings, snapshots, follow-ups, notes, families, or categories.

Regenerate the tracked JSON and SVGs from repository inputs:

```bash
uv run python scripts/build_error_cloud.py
```

Confirm that all generated files are current without changing them:

```bash
uv run python scripts/build_error_cloud.py --check
```

The generator depends only on the stdlib plus this repository's `tact.eval` replay module.
It does not inspect local telemetry, sessions, private corpora, network services, timestamps, or randomness, so repeated generation from unchanged repository inputs is byte-identical.