# MSE Validation

Curated tabular validation subset for the distributional synthetic replay generator used in the AgentPerfBench submission.

This directory intentionally does **not** store raw runner JSON payloads. The Hugging Face dataset keeps the loadable tables and notes; raw aggregate and per-turn JSON artifacts remain in the R2 archive referenced by `raw_json_r2_uri` and `per_turn_json_r2_uri` in `summary.parquet` / `manifest.csv`.

R2 archive root: `s3://agent-bench/bench_mse/`

## Files

- `summary.parquet`: one row per aggregate validation run, including R2 URIs for raw JSON provenance.
- `manifest.csv`: human-readable copy of the summary table.
- `fidelity_deltas.csv`: paired MSE-vs-real median deltas by turn-depth bin and metric.
- `notes/`: session notes documenting the debugging process and morphology/prefix-cache findings.

## Main Validation Stages

| Stage | Role |
| --- | --- |
| `prefix_aware` | Final SWE-bench proof: code-morph filler plus APC-aligned 1024-token shared prefix. |
| `apc_off_v2` | Ablation with prefix caching disabled. |
| `morphology` | Code-morph filler without shared prefix. |
| `source_locked_baseline` | Source-locked English-filler baseline. |
| `paper_terminalbench` | Clean TerminalBench validation. |
| `paper_swe_noise_*` / `paper_swe_c20` | Initial SWE-bench paper validation and repeat/noise-floor runs. |

## Supplementary Stages

| Stage | Role |
| --- | --- |
| `option2_no_replacement` | No-replacement sampling progression runs. |
| `high_concurrency` | C=100 saturation/debug runs; not used as the main fidelity claim. |

## Headline SWE-Bench Cascade

The key diagnostic comparison is turn 10-19 E2EL median error for source-locked SWE-bench C=5:

| Condition | Turn 10-19 E2EL delta |
| --- | ---: |
| APC on, English filler, source-locked | +45.5% |
| APC on, code-morph filler, no shared prefix | +31.1% |
| APC off, code-morph filler, no shared prefix | +11.7% |
| APC on, code-morph filler, 1024-token shared prefix | -3.2% |

This supports the paper claim that distributional replay for agentic serving must match both token statistics and serving-visible prefix structure. Token-count matching alone is insufficient for deep SWE-bench turns; code morphology and APC-aware shared prefix structure close the residual gap.

## Filter Rule

Main rows satisfy:

- paired MSE + real trace replay,
- H100 / Llama-3.1-8B / vLLM,
- profile in `swebench-multiturn-short` or `terminalbench-multiturn-short`,
- request success rate at least 0.75,
- stage in `prefix_aware`, `apc_off_v2`, `morphology`, `source_locked_baseline`, `paper_terminalbench`, or `paper_swe_noise`.

Excluded from the main table:

- smoke tests,
- older broad validation sweeps before the final generator framing,
- A100 real-only validation runs,
- the superseded non-v2 APC-off run,
- Python caches and temporary source-session files.