# Calibrated 4B model

[model.pt](model.pt) is the completed OpenSysOne decision-scoring artifact. It
stores learned additive adapters, a scalar head, reconstruction metadata and a
global temperature. Pretrained backbone weights are required separately.

OpenSysOne's adapter/head additions and original reconstruction code and
documentation are licensed under [Apache License 2.0](../LICENSE); see
[NOTICE](../NOTICE). The required Qwen3-4B-Instruct-2507 base is separately
distributed under Apache 2.0. Upstream model and dataset terms remain applicable;
this release does not relicense third-party material.

| Property | Recorded value |
| --- | --- |
| Format | `opensysone-adapter-v1`, custom PyTorch checkpoint |
| Base | `Qwen/Qwen3-4B-Instruct-2507` |
| Base revision | `cdbee75f17c01a7cc42f958dc650907174af0554` |
| Precision | FP32 |
| Adapters | Rank 8, alpha 16; 16,517,633 trainable adapter/head parameters |
| Selected weights | Spark B refinement step 1,500, retained at expanded branch step 0 |
| Temperature | `1.7458220720291138`, fitted on 510 separate calibration decisions |
| Verified inference limit | 1,024 complete formatted candidate tokens; no silent truncation |

The calibrated artifact SHA-256 is:

```text
e270e3da905604d97bf5a8f380ea308133403d1c4790a5c012cb1c12e9b6f348
```

The original release remains at
[`final/20260916T194403396250Z-fleet/model.pt`](../final/20260916T194403396250Z-fleet/model.pt),
with [its immutable manifest](../final/20260916T194403396250Z-fleet/backup_manifest.json).
[FINAL_MODEL.json](../FINAL_MODEL.json) records the exact payload commit, model hash,
manifest hash and publication source. The `model/model.pt` front copy has identical
bytes; moving the presentation does not change the artifact.

## Source and lineage

The selected checkpoint records training source
`24b8ccf60d388f9cbb184e03a6ae260a1f5a8b86`; its warm-start parent's source was
`4a60423c39d70f8d50472ce4f4f7fa4a4bd9fce1`. Final evaluation used
`07f10e791061a679b829ed1dc5b33897e001d67d`.
The final [evaluation manifest](../final/20260916T194403396250Z-fleet/evaluation/manifest.json)
records source-file hashes, model pin, dataset signature, configuration and packages.

The selected branch step is zero because it retains the already-trained parent's
weights. CPU lineage checks confirmed all 506 trainable tensors equal the parent,
and that calibration leaves them unchanged. Later expanded-data checkpoint 159
was evaluated but not promoted. Its diagnostic results do not describe a different
deployed model.

## Reconstruction constraint

The existing loader in [experiment.py](../source/experiment.py) reads the base path
from checkpoint metadata. For this release that path is:

```text
/home/andy/ai/models/opensysone/Qwen3-4B-Instruct-2507-cdbee75f
```

That directory must contain the pinned local base, tokenizer and matching
`opensysone-provenance.json`. The loader uses local files only and verifies base,
prompt and adapter versions. The checkpoint itself may be downloaded elsewhere
and supplied through `--checkpoint`; relocating it does not relocate the saved
base path. The current CLI has no base-path override. Do not rewrite and re-save
the published checkpoint to disguise that constraint: doing so changes its hash.

Use the project's [reproduction guide](../docs/reproduce.md) and
[Jev-compatible harness](../source/docs/usage/jev-api.md). This is not a drop-in
Transformers or standard PEFT package, and its serialization is intended for
trusted, hash-verified project artifacts. The final artifact excludes optimizer
state; resumable checkpoints and sibling validation evidence are preserved in
the [historical bundles](../archive/README.md).

Probabilities are normalized over the supplied choices. Temperature calibration
does not change the chosen answer. It was fitted on known task families; Social
IQa holdout ECE remains 8.30%, so calibration on arbitrary tasks is unproven.
See the [full measured results](../results/report.md).