Restore full model card (results, training, usage, license frontmatter)

README.md

@@ -1,98 +1,94 @@
 ---
-base_model: Qwen3.5-2B
+base_model: Qwen/Qwen3.5-2B
 library_name: peft
+license: apache-2.0
 pipeline_tag: text-generation
 tags:
-- base_model:adapter:Qwen3.5-2B
 - lora
-- verdict-slot
-- opencodifier
+- transformers
 - decision-model
-inference: false
+- opencodifier
 ---
 
-# OpenCodifier E1-E 2B Decision Model
-
-A verdict-slot LoRA adapter on Qwen3.5-2B for typed decision-making (choice, boolean, score).
-Part of the OpenCodifier decision runtime: opencodifier.pages.dev.
-
-## Model Details
+# OpenCodifier E1-E — Qwen3.5-2B verdict-slot decision LoRA
 
-- **Model type:** Decision classifier — verdict-slot LoRA on Qwen3.5-2B
-- **Base model:** Qwen3.5-2B (Apache-2.0)
-- **Adapter architecture:** LoRA r=16, alpha=32, dropout=0.05
-- **License:** Apache-2.0
-- **Developed by:** TaskWizerAI
-- **Released by:** TaskWizerAI
+A LoRA adapter that turns [Qwen/Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B)
+into a Jev-style **decision model**: probabilities are read from the
+log-probs of verdict-slot label tokens at the answer position — nothing is
+generated, nothing is parsed. It is the decision rung of the
+[OpenCodifier](https://github.com/aliasfoxkde/OpenCodifier) ladder, and the
+first own-trained arm to pass both release-gate legs.
 
-## Gates Passed
+## Results
 
-| Benchmark | Result | Gate | Status |
-|---|---|---|---|
-| Suite holdout (231 items) | 0.8083 accuracy | ≥ 0.75 | ✅ PASS |
-| JevBench public 231 | 0.6840 accuracy | ≥ 0.6494 | ✅ PASS |
+All numbers from the OpenCodifier evaluation records
+([BENCHMARKS.md](https://github.com/aliasfoxkde/OpenCodifier/blob/main/docs/BENCHMARKS.md),
+TRAINING.md §9.9.1). Public 231-item JevBench split, deterministic
+readout, fp16, single V100.
 
-Full evaluation evidence: `benchmarks/decision-model/results/CALIBRATION.md` and
-`benchmarks/decision-model/results/e1e-slot-231/results.jsonl`.
+| metric | value |
+|---|---|
+| JevBench-231 accuracy | **0.6840** (publish bar 0.6494 — clear) |
+| Internal holdout suite | **0.8083** (gate ≥ 0.75 — GATE_PASS; seed-replicated 0.8000) |
+| Raw board ECE | 0.207 |
+| Holdout-fitted temperature (T=3.362, analysis-only on board) | ECE 0.207 → 0.127 |
+| Auto-decidable @5% risk | 0.100 (23/231) |
+| Brier | 0.504 |
+| p50 latency | 0.58 s |
+
+Per-family shape: routing / tool-selection / ordinal families at 1.00;
+relational/compositional families are the weakness (multi_hop 0.389,
+judge_hard 0.412, long_policy 0.368). The board is lost on compositional
+families, not on routing — the candidate lever is relational-heavy
+soft-label data, not parameters.
 
 ## Training
 
-- **Corpus:** 187,822 decision items (e1-softlabel-v2), teacher distributions from MiMo
-- **Teacher-gold agreement:** 96.1–97.7% argmax match
-- **Hyperparameters:** lr=1e-4, batch=2, accum=8, max_len=2048, fp16, sdpa, seed=42, 1 epoch
-- **Platform:** T5500 (2× V100, 120W GPU cap)
-
-## Evaluation Results
-
-### JevBench Public 231 (231 items, authors' harness)
+- **Base**: Qwen3.5-2B, LoRA verdict-slot readout
+- **Corpus**: merged-v3 (173,452 records; composition and join rates in
+  TRAINING.md). The corpus itself is **not** released with this adapter.
+- **Schedule**: 11,009 steps, single V100 (T5500), fp16
+- **Calibration**: temperature T=3.362 fitted on a labelled holdout
+  (holdout ECE 0.179 → 0.117); artifact ships in the OpenCodifier repo
+  at `calibration/e1e-qwen3.5-2b-holdout.json`
 
-| Metric | Value |
-|---|---|
-| Accuracy | 0.6840 |
-| Macro | 0.6840 |
-| ECE | 0.127 (calibrated T=3.362) |
-| Brier | 0.332 |
-| p50 latency | ~1.7 s (CPU, single decision) |
-| Replay determinism | 231/231 |
-
-### Suite Holdout (231 items, internal)
+## Intended use
 
-| Metric | Value |
-|---|---|
-| Accuracy | 0.8083 |
+Typed multiple-choice decisions (routing, tool selection, ordinal
+grading) where calibrated token-level probabilities — not free text —
+are the product. Pairs with the OpenCodifier engine's confidence gate:
+low-confidence reads abstain or escalate to a stronger rung.
 
-## Uses
+## Out of scope
 
-### Direct Use
+Free-form generation; the base model's general chat ability is not
+preserved or claimed. Not evaluated for safety of generated content —
+the adapter is a readout head, not a chat model.
 
-Use via the OpenCodifier binary:
+## How to use
 
-```bash
-# Engine-only (no model needed)
-opencodifier decide --input request.json
+```python
+from peft import PeftModel
+from transformers import AutoModelForCausalLM, AutoTokenizer
 
-# With E1-E model (requires llama-server running the decision fork)
-opencodifier decide --input request.json --ladder-url http://127.0.0.1:8080
+base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-2B", torch_dtype="float16")
+model = PeftModel.from_pretrained(base, "TaskWizerAI/opencodifier-e1e-2b")
 ```
 
-### Out-of-Scope Use
-
-- Not a general-purpose language model
-- Not trained for prose generation
-- Not calibrated for non-decision classification
+Scores come from label-token log-probs at the answer position; see the
+OpenCodifier repo for the readout harness and the Jev adapter wire shape.
 
-## Calibration
+## License
 
-Temperature fitted on the suite holdout: T = 3.362.
-Apply the calibration artifact from `results/calibration/qwen3.5-2b-q4_k_m/fit.json`
-before evaluating on any surface other than the holdout.
+Apache-2.0, matching the OpenCodifier project.
 
-## How to Cite
+## Citation
 
-OpenCodifier decision runtime: https://github.com/TaskWizerAI/OpenCodifier
-
-## Remarks
-
-- Corpus withheld (per OpenCodifier model release policy)
-- Calibration artifacts published separately
-- Explicit opt-in pull only — never auto-downloaded at build time
+```bibtex
+@software{opencodifier_e1e_2026,
+  title  = {OpenCodifier E1-E: a gate-passing 2B decision-model LoRA},
+  author = {Kinney, Michael},
+  year   = {2026},
+  url    = {https://github.com/aliasfoxkde/OpenCodifier}
+}
+```