head-exit-gold — examflow early-exit policy head (Laya fine-tune)

Fine-tuned convaiinnovations/laya (Apache-2.0) for one job: after Tier-1 (text layer + fast CPU), decide among exit-accept, run-one-more, full-fanout — in a single encoder pass (~ms on GPU). The cost saver: exits when cheap agreement suffices even at moderate conf, refuses to exit when the text layer is blind (handwriting on forms).

Training (all $0: Kaggle T4 x2)

• 2,000 gold construction-truth cases (incl. blind-layer traps), 0 eval overlap
• Full fine-tune, 3 epochs, lr 2e-5, batch 8, bf16; option order shuffled per sample + 3 instruction variants (anti-prior-collapse)
• Held-out synthetic eval (n=150): 1.0000 vs heuristic 0.740 (+26pp)

Scope & limits (read before use)

• SYNTHETIC distribution: proves the loop, not real-world accuracy.
• Confidence is temp-uncalibrated until per-head refit (base checkpoint ships invalid temperatures — refit before trusting it).
• Never final-judge duty: dispatcher/signal layer only, abstain below tau.
• Safe format: model.safetensors (no pickle, no code execution on load).

Load

from laya import Agent
agent = Agent(model_id_or_path="ngdghfdc/head-exit-gold")
out = agent.predict(state, {"exit": {"type": "choice",
    "instructions": "Pick the best early-exit action.",
    "criteria": {o: o for o in ["exit-accept", "run-one-more",
                                "full-fanout"]}}})