# Local cross-encoder diagnostic

This standalone research tool never modifies production search or its Google
EmbeddingGemma 2 model. Do not confuse the cross-encoder's name with the historical
MiniLM embedding experiment. Publisher weights stay under ignored `.searchi` data.

Pinned publisher revision, asset bytes/SHA-256 and licence provenance are in
`model-manifest.json`; the pre-inference contract is in `compatibility-plan.json`.
The existing CPU Python environment can execute it without installing Torch or
changing production dependencies:

```powershell
python tools/rerank-eval/download.py
$env:SEARCHI_FORCE_CPU = '1'
& tools/semantic-proof/.venv/Scripts/python.exe tools/rerank-eval/compatibility.py --output .searchi/rerank-eval/compatibility-v1.json
```

The runner verifies all assets, rejects changed frozen inputs, enables only CPU
Execution Provider with one thread, bounds five pairs to 256 tokens including
special tokens and at most 64 query tokens, checks finite one-logit-per-pair
outputs, and records serial/batch parity and twenty warm timing samples. It creates
new reports atomically and refuses overwriting prior evidence. A failed parity
contract writes the observed report then exits nonzero. Never weaken its frozen
tolerance after observing output.

The first execution failed serial/batch logit parity; see
[recorded result](../../docs/LOCAL-RERANKING.md). It is not a product integration,
relevance calibration or full search performance benchmark. Scores are not
probabilities that the requested artifact exists. Any further experimental input
or scoring policy needs a separate version and output path.

The separate unquantized follow-up preserves the first failed evidence:

```powershell
python tools/rerank-eval/download.py --reference
$env:SEARCHI_FORCE_CPU = '1'
& tools/semantic-proof/.venv/Scripts/python.exe tools/rerank-eval/reference.py --output .searchi/rerank-eval/reference-v1.json
```

`reference-plan.json` retains five short pairs and adds a 256-token budget case.
F32 parity passed; quantized score differences persisted. Mixed-length padded
batching was slower. Neither result establishes general absence calibration or
sufficient default-search latency. Detailed costs and possible VM contention are
recorded in the linked report documentation.