Phase 8 — Evidence-First Drawing Intelligence: From One-Shot Guesses to Traceable Quantities

A quantity is not trustworthy because a model can name it. It is trustworthy when the system can show where it was seen, how it was counted, how repeated views were reconciled, and when a human should doubt it.

We designate this work Phase 8: Evidence-First Drawing Intelligence. Phase 6 completed the live Peak tile accuracy sprint (pooled detection 78.2%, QA@20 38.0% — see /blog/peak-phase6-accuracy-sprint). Phase 7 independently repaired BOQ and ground-truth integrity. Phase 8 turns those lessons into an opt-in production evidence path that must pass its own frozen ablation before any additional headline uplift is claimed.

The bottleneck is evidence resolution, not one more score

Phase 5 made the denominator honest: Peak pooled QA@20 was 33.8% and detection 55.1% across 234 drawable ground-truth rows. P&D was the limiting sheet: 33.8% detection and 14.1% QA@20. Phase 6 live tiles moved P&D detection to 71.8% and pooled detection to 78.2%, while QA@20 only reached 38.0% pooled / 21.1% P&D — proving that finding symbols is not the same as usable quantities. The recurring missing mass for small fittings remains partly physical: A0 sheets at ~7021×4967 px still lose detail under aggressive cost ceilings. No ranking rule can recover an object that was removed by the input transform.

These are deliberately heterogeneous studies. Peak tests a dense multi-discipline tender, N23 tests Chinese P&D and cross-page aggregation, and CPS tests an unseen PDF bill. They do not average into a product claim. Together they prevent a Peak-only optimisation from masquerading as general drawing understanding.

A platform architecture for traceable quantity

The implementation is opt-in behind EVIDENCE_TILED_EXTRACT=1 and EVIDENCE_TARGETED_RECOUNT=1. It preserves TeraQuant’s rotation-aware display coordinate system, uses the tested tile planner and local-to-global transform, rejects tile-only unlocated claims, and records evidence-pass cost and latency beside the existing model run. That is a controlled experiment surface, not a silent production switch.

Annotation integrity is part of accuracy

We also corrected the visual evidence standard. Earlier Phase 3 imagery was withdrawn because an offline demo fallback could draw plausible but fabricated positions, and it had no real measurement geometry. The renderer now has no demo path; it refuses empty source data, reads actual length polylines, computes metres in the rotation-aware display space, suppresses count/location-inconsistent pins, and keeps project-level SOR comparison separate from each page-local position.

Evaluation protocol and the next claim we are allowed to make

• Freeze the existing extraction and score it beside the Phase 8 evidence path on the same drawable rows. Report detection, QA@20, QA@0, selective coverage, latency, image count and estimated cost together.
• Run the primary ablation on Peak P&D, where fittings and valves dominate the remaining miss mass; keep FS and MVAC as regression sheets rather than tuning targets.
• Repeat only the frozen protocol on N23 and CPS. Label N23 as offline re-aggregation until a fresh live extract exists, and keep CPS INFO until its PDF-BQ parser is independently audited.
• Promote the flags only after a pre-registered comparison improves measured quantity quality without concealing a cost, coverage or cross-project regression.

Phase 6 published measured live-tile gains on Peak P&D; Phase 8 still reports no separate product-default uplift until its opt-in path is ablated on frozen rows. The present achievement is a deployed, reversible evidence path; three-project baselines; an annotation standard that cannot fabricate geometry; and a protocol under which the next improvement can be believed or rejected. The goal remains significant per-line quantity accuracy, but the platform will earn that claim one visible item at a time.