SchemaForge V2

• A working hybrid, not just a measured residual. v1.5 measured the distilled model in isolation on the semantic residual and never built the merge. V2 builds it — static routing by declared field ownership, merge by precedence — and measures the combined system directly.
• The hybrid dominates both components on every metric, on every checkpoint tested. Field F1 0.6432–0.6827 against 0.2911 (rules alone) and 0.4461 (model alone); hallucination 0.0136 against 0.0852; schema validity 0.8333 against 0.7361.
• A severity-swept explanation for why rules cannot be trusted alone even where they are strong: precision decays 0.990→0.537 under corruption, while recall sits at a structural ~0.26 ceiling present even at zero corruption — the reason no rules/model crossover appears at any severity.
• The dominant failure mode named, measured, and left unsolved on purpose: silent omission at 55–62% of all classified failures, with four independent interventions reported as failing rather than quietly retried until one worked.
• A methodological audit that invalidated part of the project's own early evidence — non-deterministic teacher labelling, caught by comparing the full iteration history rather than adjacent runs, fixed, and its actual impact bounded by a controlled re-run (0.0003 F1) rather than assumed.

A tuned deterministic pre-pass resolves every field it owns by declared schema partition — dates, IDs, amounts, emails, phone numbers — asserted at import time never to fill a semantic field. Whatever remains goes to a ~1B distilled student (MiniCPM5-1B, teacher google/gemma-4-31B), trained by sequence-level distillation on teacher labels that must clear a four-check admission gate: JSON-parses, schema-validates, every semantic value is a literal substring of source text or a registered ontology derivation, nothing asserted beyond what the schema licenses.

Routing is static and precedence-based, not confidence-gated — the pre-pass's answer wins for fields it owns, the model fills the rest, and the merge is schema-validated. Confidence is measured for future use but is not part of this routing decision, because calibration (raw mean 0.978 against 47% correctness) is not yet trustworthy enough to route on.

Fifteen training iterations, greedy deterministic teacher sampling after iteration-history comparison found and fixed a non-deterministic labelling bug that had been silently affecting every earlier run.

On a held-out set spanning twelve schemas (three withheld from training entirely):

The hybrid claim held on every checkpoint measured across fifteen iterations, not only the best one.

The unsolved half: across every configuration, the dominant error category is not hallucination and not schema violation — it is silent omission, at 55–62% of all classified failures, more than every other category combined. Corpus scale (288→3,780 records), ontology-coverage fills, and a corpus concentrated on the two operators most linked to omission were all tried; none reduced the share below roughly half, and the targeted-corpus attempt made it worse (55.0%→60.1%). Working hypothesis: the admission gate is a precision filter by design, and a precision filter applied to training labels teaches the model to omit rather than guess.

Omission is unsolved, not merely unoptimized — four independent interventions, none of which worked, is the honest state of the art here.

Confidence calibration is measured but not usable for routing (ECE 0.16 at best after temperature scaling), so the confidence-routed escalation the project's charter specified is not built, and no latency, cost-per-document, or throughput figures exist for the system as configured.

A severity-sweep benchmark re-run did not reproduce the project's own earlier figures exactly, on an apparently identical recipe; the discrepancy is unexplained. The incorrect_normalization failure category reads exactly zero in every measurement taken, ontology fills or not — also unexplained.

The teacher-labelling non-determinism bug was caught and its impact bounded (0.0003 F1 at the corpus scale of later iterations), but earlier, smaller-corpus results in the project's history predate the fix and are not built upon, even though they are not retracted.

The hybrid-with-precedence-merge design sits closest to constrained-decoding and grounded-generation practice: every semantic value the model contributes must be traceable to source text or a declared ontology derivation, the same discipline v1.5 introduced for training labels, now also enforced at inference through the ownership partition.

The omission finding is the paper's most transferable result independent of this project's checkpoint: a system tuned toward precision on its training labels (via a grounding gate) reproducibly tends toward under-completion at inference, not over-generation. That is close to but distinct from the hallucination-avoidance literature's usual framing, which optimizes against false positives without separately measuring the false-negative cost of doing so.