# Codex LLM Defluffer A/B Results

Generated: 2026-06-08T01:08:07.080Z

Run dir: `llm-results\2026-06-08T01-06-48-318Z`

Variant: `standardGuardedDedupe` vs no script.

## Summary

| fixtures | avg input savings | avg baseline out toks | avg defluffed out toks | avg baseline score | avg defluffed score | critical losses | compression critical losses | unchanged-prompt losses | winners                |
| -------- | ----------------- | --------------------- | ---------------------- | ------------------ | ------------------- | --------------- | --------------------------- | ----------------------- | ---------------------- |
| 6        | 16.29%            | 194.3                 | 205.8                  | 4.83               | 4.5                 | 1               | 0                           | 1                       | {"tie":5,"baseline":1} |

## Cases

| id             | category        | input saved | base score | defluffed score | winner   | critical loss | compression-caused? | notes                                                                                                                                                            |
| -------------- | --------------- | ----------- | ---------- | --------------- | -------- | ------------- | ------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| demo_001       | code_refactor   | 33.65%      | 5          | 5               | tie      | false         | false               | Both answers preserve the original intent: Python DB retrieval from env config, retry_count >= 5, active true filtering, JS refactor without external libraries, |
| code_002       | code_generation | 4.23%       | 5          | 5               | tie      | false         | false               | Both answers implement parseUserId in TypeScript, handle null/undefined/empty string without throwing, avoid external libraries, return null for invalid input,  |
| json_004       | structured      | 2.94%       | 5          | 5               | tie      | false         | false               | Both answers return only valid JSON and preserve all keys and string values.                                                                                     |
| transcript_009 | transcript      | 29.82%      | 5          | 5               | tie      | false         | false               | Both answers preserve the original intent, summarize the invoice issue, and extract clear action items with owners.                                              |
| legal_013      | legal           | 0%          | 4          | 2               | baseline | true          | false               | Baseline correctly requests missing clause text. Compressed invents a summary and replaces required legal terms, losing mandated wording and defined terms.      |
| copy_017       | creative        | 27.08%      | 5          | 5               | tie      | false         | false               | Both responses correctly request the missing landing page copy and preserve the product name Defluffer Pro.                                                      |

## Notes

- Codex CLI agent overhead tokens are recorded separately in `report.json`, but not used as prompt savings metric.
- Input/output token counts use same local estimator as earlier viability run.
- Read-only Codex processes used; prompts told agent not to modify files or run shell commands.
- Critical loss with 0% input savings means model variance, not compression-caused damage.