# Study status after Study 04

Study 04 has completed the proposed independent implementation check, locally frozen fresh-input evaluation, and comparison under matched solve ceilings. The [earlier review](STUDY_STATUS_PRE_STUDY04.md) is retained as a historical decision record. The [frozen protocol](../validation/PROTOCOL.md) and [report](../validation/output/REPORT.md) distinguish planned procedures from the post-run timing-scope clarification.

## What the new evidence establishes

- 120 new input files: 80 witnessed-feasible requests, 20 unknown-feasibility requests, 10 proven-impossible and 10 unsupported controls.
- The same initial candidate and at most two follow-up solves per policy. Restart certifies 21/80 feasible requests, nested sampling 32/80, and witness refinement 50/80.
- Paired comparison: refinement succeeds alone on 18 cases versus nested sampling; nested sampling succeeds alone on zero. Both succeed on 32; neither on 30.
- Exact quadratic-radical extrema and Bernstein strip moments provide a separate implementation of gap and moment checks. No contradictory decisive results in development or evaluation.
- 2,022 development polynomial decisions, 42 earlier export entries, 764 newly replayed paths (120 inputs + 80 witnesses + 564 candidates), and 41 passing behavioral tests across all four studies.

These counts refer to different units. The evaluation has 360 workflow rows, including 60 preflight stops; 300 eligible workflows reuse 100 initial candidates. Its 764 history entries include repeated references to shared initials and must not be confused with its independently replayed 764 input/witness/candidate paths. Identical totals here are coincidental.

## What remains unresolved

Thirty requests proven feasible by saved witness shapes still fail the minimum gap after two refinement rounds. All other preservation checks pass. This is a limitation of the controller/budget, not an impossibility proof. The generator retains an affine common centerline, even while varying asymmetric gaps, translation, widths and landmarks. General curves, real-world SVG coverage, perception and attribution remain separate questions.

The exact-checker comparison is independent implementation work by the same author, sharing Python and XML dependencies and the declared shape model. It is not formal verification or external review. Single-host subsecond timings are descriptive, and equal solve ceilings do not mean identical CPU work.

## Decision record added by this round

| Evidence | Decision |
|---|---|
| No contradictory decisive checker result | Keep both implementations as an audit cross-check; retain the acceptance gate. |
| 18 paired refinement wins and zero paired losses versus denser sampling | Further develop witness feedback within this supported model. |
| 30 witnessed-feasible gap failures remain | Diagnose and address this limitation before broadening the geometry family. |
| Fresh fixtures still use an affine centerline | Do not claim arbitrary asymmetric-boundary coverage. |

## Recommended next step — not yet performed

Use the 30 feasible failures as explicitly labeled development cases. Trace gap-minimum and witness locations across rounds, then evaluate whether more rounds, a stronger numerical clearance buffer, or direct stationary-point constraints address the failures efficiently. Freeze a new protocol and a separate untouched input set for any comparative claim. Do not silently tune or replace Study 04.

Broader shape support should follow a credible coverage improvement and a concrete need that two graphs cannot express. The current work supports geometrically verified editing and traceable local operation records; it supplies no evidence of human authorship, ownership or training-data influence.