# EXP9 — grading the child's taste, and the rulers bent

## Raw table (solar graph vs nursery ground truths)

| target | volunteered | GED cost | preserved | WL |
|---|---|---|---|---|
| feedback3 | YES | **1.0** | **0.833** | 0.050 |
| cycle3 | YES | 5.0 | 0.375 | 0.050 |
| tree | YES | 8.0 | 0.273 | 0.036 |
| inversion | no | 5.0 | 0.375 | 0.050 |
| diamond | no | 7.0 | 0.300 | 0.042 |
| chain | no | 6.0 | 0.333 | **0.287** |

## Verdict: INCONCLUSIVE — the measurement failed before the child could

1. **GED unreliability exposed in its own table**: cycle3 and feedback3
   are ISOMORPHIC (the designed near-miss family) yet scored 5.0 and 1.0
   against the same probe. optimize_edit_paths' first-yield
   approximation is not stable enough to grade taste. A comparator
   scoring isomorphic targets differently is disqualified from verdicts.
2. **The metrics rank oppositely**: by GED, volunteered beats control
   (4.67 vs 6.0 avg cost); by WL, control beats volunteered (0.126 vs
   0.045) — driven by chain, the path-shaped twin of the solar graph,
   which the child did NOT volunteer. Third consecutive experiment where
   metric choice flips the verdict (EXP3-run2 P2, EXP4, now EXP9).

## What survives the wreckage

- Tantalizing, unconfirmed: the child's tank-gauge-valve overlay
  (feedback3) is the CHEAPEST mapping on the board under GED — if that
  holds under a repaired comparator, the child's first volunteered
  correspondence was also its best. Held, not believed.
- **Metric honesty is promoted from queue to BLOCKING.** Three
  experiments have now had their verdicts hostage to comparator choice.
  No further taste/correspondence claims until the comparators are
  calibrated on known-isomorphic and known-distant pairs with stability
  tests. The instrument must distinguish its own lenses before it may
  grade a child's eyesight — §116's founding demand, now overdue.