JOB136 INDEPENDENT AUDIT PLAN — WRITTEN BEFORE NEW IMPLEMENTATION / RESULTS

The user's desired typed-link F1 > 0.91 is a useful threshold, not a result
to be manufactured. Job135 COPY scored 0.905596 on its now-public 600 scenes.
The new system must be evaluated against COPY on the SAME fresh fields. A
new easier exercise mixture can raise COPY above 0.91 without any learned gain.
The historical 0.55 and 0.91 are context, never interchangeable denominators.

1. Report graph preservation and required change separately:
   - typed-edge micro precision, recall, F1; exact graph accuracy;
   - persistent-edge recall (source intersection target);
   - added-edge F1 and removed-edge F1;
   - missed necessary edits and unnecessary edits;
   - per-relation results, preserving direction for types 2–5.
   COPY is a strong preservation baseline and has zero required-edit recall.

2. Freeze architecture, data policy, checkpoint selection and any confidence
   thresholds before final outcomes. Development-only selection is allowed.
   Preserve unsuccessful trials. Freeze is author-maintained chronology, not
   externally enforced or cryptographically authenticated preregistration.
   Fresh scenes test generalization within the authored teacher; they do not
   make that teacher independent or make its rules semantic ground truth.

3. All methods see the same available source facts. Target count, target node
   validity, target edge mask, or target edit identity may not enter inference.
   Source-observed counts and deterministic legal output bounds are allowed,
   but distinguish those engineered priors from learned counting/validity.
   Inference must not call the teacher or receive target arrays.

4. If the task changes to fixed-cell graph editing, explicit edit requests,
   candidate-limited additions, or teacher-designed geometric features, label
   the changed scope. Better scores on that reduced or informed problem do
   not establish a better autonomous complete-field generator than Job135.
   State any candidate-scaffold coverage ceiling and compare identical inputs.

5. Report raw learned edit proposals and any protected-mode adapter separately.
   A lock copied by construction is an engineering guarantee, not something
   the model learned. A guarantee to preserve unedited fields is nevertheless
   useful when its scope, conflicts, and intervention authority are explicit.
   Do not hide unsafe proposals behind successful postprocessing.

6. Run deterministic regression fixtures for:
   symmetric orientation versus directed reversal; duplicate edges; no-op;
   deliberate deletion/retyping; protected edge survival; capacity boundaries;
   input purity; parent/slot identity; malformed requests; conflicting locks.
   Validate that fake COPY and empty-graph predictors cannot earn change credit.

7. Repeated use is a separate condition. Follow the model's own previous output
   for multiple steps and evaluate at each step. Compare the teacher response
   to that current input (local one-step error) AND an intended teacher-only
   trajectory (accumulated divergence) where feasible. Do not average away a
   failure late in an interaction. No-op chains must preserve protected state.

8. Report seed variation and paired scene-level differences when available.
   An aggregate F1 > 0.91 alone says neither all scenes are correct nor that a
   new artistic composition will receive a faithful interpretation.

The audit does not certify subjective colour experience, philosophical
understanding, artistic merit, semantic entailment, or general intelligence.
It can establish a bounded graph-editing capability and explicit guarantees.
