SchemaForge V3 (draft, iterations 1–4)

• The recipe grid, fully bracketed. Epochs 1/2/3 and LR 1e-5/2e-5 swept at a fixed corpus; 2 epochs at LR 2e-5 confirmed as the optimum V2/V3 had already been using — the sweep found no improvement in either direction, which is itself a result: the default recipe was not the bottleneck.
• The lowest omission share measured in the project, 52.3%, the first result below V2's stable 55–62% band across fifteen iterations — reported alongside the fact that it is still not below half of all failures and still 0.0082 F1 behind the V2 release candidate.
• A deadlock recovered by exploiting the pipeline's own determinism, not by re-running the expensive step: greedy, seeded teacher generation means saved outputs from a prior run are byte-identical to a fresh run's, so a gate-only re-run on saved rejections recovered 37% of them at CPU cost instead of a full GPU re-generation.
• Every result reported against the V2 release candidate, not a moving baseline — no iteration here has been used for a Hugging Face upload, and that is stated per-iteration in the run manifests, not just in prose.

Same pipeline as V2 — greedy, seeded teacher labelling (the determinism V2's own audit established as load-bearing), a four-check admission gate, sequence-level distillation. Corpus held at n=100 documents (up from V2-FINAL's n=75) across the sweep, with the four iterations varying one training-recipe axis at a time: corpus scale in isolation, then epochs 3→2→1, then learning rate 2e-5→1e-5 at the 2-epoch point.

Every result is cross-checked on both a 72-record and a 288-record held-out eval, and every run is provenance-recorded with a manifest that states explicitly whether the checkpoint was used for a Hugging Face upload — none, so far, have been.

Corpus scale alone regresses the result (iteration 1, -0.0246 F1) — the fourth scale-up attempt in this project's history to fail against omission, and the one that isolated total training steps as the more likely mechanism: iteration 1's 3-epoch run over the larger corpus produced more steps than V2-FINAL, into an over-converged regime.

2 epochs is a clean interior optimum across the sweep (1ep 0.6597 / 2ep 0.6745 / 3ep 0.6581), and LR 2e-5 beats 1e-5 at that point (0.6745 vs 0.6671, with the lower rate costing 0.0833 in schema validity — the largest single regression measured).

52.3% missing-field share is a new low and the first result under V2's 55-62% band — three to nine points better than any V2 configuration. It is also still above half of all failures, and iteration 2's field F1 remains 0.0082 below the V2 release candidate on the directly comparable eval.

No V3 checkpoint beats V2-FINAL on the comparable 72-record eval. This paper reports a recipe-grid sweep, not a solved omission problem.

The experiment that most directly tests V2's open hypothesis has not run: whether stacking delabel/implicit into the full ten-operator corruption mix, rather than isolating them (which made things worse in V2), reduces omission. Everything reported here is upstream of that question.

The gate-loosening question is untouched. Whether the admission gate's source-support check can be safely loosened to admit legitimate denoising corrections without also admitting wrong labels — flagged as a real false-rejection source in two V2 iterations — has not been attempted in V3 either.

The 0.0082 F1 gap to V2-FINAL is small relative to a 72-record eval's resolving power and should not be read with more confidence than the eval set supports in either direction.

incorrect_normalization reading exactly zero remains unexplained, carried over unresolved from V2.

This paper extends schemaforge-v2 directly rather than standing alone — same architecture, same evaluation protocol, same held-out schemas. It is reported as a separate draft rather than folded into a V2 revision because V2's own status (methodologically closed, release candidate prepared) should not be reopened by an in-flight sweep that has not yet beaten it.

The omission-as-gate-artifact hypothesis this paper inherits from V2 sits closest to the selective-prediction and abstention literature, where a model trained to say nothing under uncertainty can systematically under-report exactly where labels are hardest to ground — the mechanism V2 proposed and V3's next experiment (stacking rather than isolating the omission-linked operators) is designed to test.