# EXP3 — Sense calibration, run 1: the instrument catches its own ground truth

Channels vs the deterministic SVG primitive graph (7 nodes / 10 edges):

| channel | WL vs truth | spectral dist | graph size |
|---|---|---|---|
| vision-on-raster | 0.094 | 10.09 | 4n/6e |
| language-named | 0.021 | 7.10 | 6n/7e |
| language-nameless | 0.025 | 9.22 | 4n/3e |

## Frozen prediction scorecard

- **P1 FAILED**: vision WL 0.094, far below the predicted 0.3–0.7 band.
- **P2 inconclusive at floor**: nameless 0.025 vs named 0.021 — the
  predicted direction, but both at noise level; no call.
- **P3 FAILED**: named language is spectrally CLOSER to the deterministic
  graph than vision (7.10 vs 10.09) — opposite of prediction.

## The finding that explains the failures

**"Ground truth" is basis-relative.** The deterministic channel delivers
truth about the artifact's ENCODING — SVG document topology (root
contains group contains primitives, sibling order) — while every model
channel described the DEPICTION — scene topology (panels meet at a cap,
parts inside an outline). Comparing them treated primgraph_v0 and
relgraph_v0 as the same basis, exactly what ADR-0002 said not to do
("the absence of a comparator between two schemes is a recorded fact").
The instrument's numbers went to the floor because the experimenter
compared across bases without a transformation — §23 (canonicalization
is a transformation whose losses must be observable), demonstrated on
the experimenter.

A beach ball's document structure and its scene structure are two
different shapes. Neither is "the" truth of the artifact.

## Consequence (next run)

svg sense v1: derive the SCENE graph deterministically as well — spatial
containment from geometry (area inclusion), adjacency from boundary
sharing, meeting-points from coordinate coincidence. Then sense
calibration compares scene-parse to scene-truth and document-parse to
document-truth, with the document↔scene transformation stored explicitly,
carrying its own loss vector. Both ground truths coexist as competing
representations — which is what the kernel was built for.

---

# Run 2 — rescored within the scene basis (svg sense v1, no new model calls)

Scene truth (deterministic geometry): 5 nodes / 7 edges — outline
contains three panels and a cap; panels pairwise meet.

| channel | WL | spectral | preserved | destroyed | invented-by-truth |
|---|---|---|---|---|---|
| vision-on-raster | 0.031 | 5.83 | 6 | 4 | 6 |
| language-named | 0.119 | 6.36 | 7 | **6** | 5 |
| language-nameless | 0.031 | 5.43 | 7 | **0** | 5 |

## The result that matters: the vector and the scalar disagree — the vector is right

The nameless parse is a clean SUBGRAPH of the scene truth: everything it
asserted is true (0 destroyed); it simply saw less. The named parse is
bigger and wronger: the word "beach ball" imported sand, a beach, and
resting-relations — six structures not in the scene. **Scalar similarity
ranks the confabulator higher (0.119 > 0.031). The loss vector exposes
it.** §11/§27's "one destroyed constraint can outweigh many preserved
trivia" is no longer a design principle; it is a measured event, three
days into the project's life.

Sense ordering by honesty (destroyed-count): nameless-language (0) >
vision (4) > named-language (6). Names are the least honest sense.

## Prediction rescoring (fair basis)

- P1 still FAILED on WL (0.031, band was 0.3–0.7) — though the edit view
  shows genuine partial recovery (6 preserved); WL-Jaccard is harsh on
  small graphs of unequal size. Metric choice is itself under test.
- P2 SUPPORTED decisively on the vector (0 vs 6 destroyed), REVERSED on
  the scalar — the prediction's truth value depends on the metric, which
  is a finding about metrics.
- P3 mixed: vision spectrally closer than named language, farther than
  nameless. No call.