# Suite24 frozen design

## Why this experiment exists

Suite23 did not justify a larger global graph. Its strongest positive evidence came from (1) correct student-history identity and (2) direct structural/prerequisite course relations. Its major failures came from the full C design and its coarse faculty-level fusion. Suite24 is therefore a **target-conditioned local structural-path model** with a **Committee-anchored bounded residual** rather than a generic graph classifier plus post-hoc blend.

## Variants

### A — `suite24_history_residual`
- target-course-conditioned attention over up to 16 historical courses;
- exact historical grade/pass/credit/recency attributes;
- relation-to-target code retained;
- bounded history correction to Committee;
- no structural correction.

### B — `suite24_structural_residual`
- A plus up to 10 direct prerequisite/structural neighbors of target course;
- bounded history + structural corrections;
- fixed evidence availability gates.

### C — `suite24_evidence_residual` — PRIMARY
- B plus monotone row-wise history and structure authority gates;
- gates depend only on evidence availability/relevance summaries, not faculty identity;
- each branch is bounded to ±0.75 logit;
- primary final logit is `zC + aHΔH + aSΔS`.

## Non-primary outputs

- `suite24_graph_standalone`: auxiliary relational classifier from the same representation but without Committee.
- `suite24_safe_scale`: global chronology-selected shrinkage of C's learned total correction. Diagnostic/deployment fallback only.
- `suite24_posthoc_blend`: old-style Committee + standalone blend. Kept only to test whether the new integrated residual actually improves on simple fusion.

## Falsifiable expectations

1. B should improve A if direct structural relations are truly useful.
2. C should improve B if row-wise evidence authority solves Suite23's coarse-fusion problem.
3. Shuffling student history should materially hurt C.
4. Shuffling direct structural neighbors should hurt when structural support is present.
5. Removing a correction branch should hurt if the branch learned generalizable information.
6. C should outperform Committee broadly and should at minimum improve on Suite23.
7. The standalone head should tell us whether future work can remove Committee entirely.
