Technical report

Everything in this report is relative to one anchor, and this is it. A board is the set of reference images that defines a look. This one holds eight public-domain paintings: four Claude Lorrain harbor landscapes and four van Gogh. Every score in this report is relative to these eight images and to nothing else.

What to look for The board reports an effective size of 4 rather than 8, and flags near_duplicate_references: some references are near-duplicates of each other, and the scoring counts them as fewer independent examples rather than double-counting one look. The leverage column shows how much each reference tightens the score band on its own.

The score is the instrument every later section relies on, so before it is trusted on new work it is pointed at material whose answer is known. Twenty-four candidates: the eight references themselves, plus a mirrored and a center-cropped variant of each. The score answers one question — how hard is this image to tell apart from the board’s own members? Technically it is a board-relative conformal p-value: a score of 1.00, like the top row, means the board’s own members fit no better than this candidate does; a score near zero means almost none of them fit that poorly. Each score carries a 90% interval, and two candidates whose intervals overlap are reported as tied rather than ranked apart.

What to look for The bottom of the table is trojan women at 0.111 — a version of a painting that is itself one of the eight board references. Membership does not buy fit; the score is about the look, not the file list. And 18 of the 23 adjacent pairs tested are statistical ties: on an eight-image board, small rank gaps are mostly noise, and the report says so instead of pretending precision.

Reading the columns: n_local counts the references near enough to inform the score; style repeats the score on its own axis; palette, tone, and composition are classical diagnostics reported beside the score, never merged into it; the nearest reference is the board member the candidate most resembles.

Replace only the selected apple tree with a mature lemon tree. Preserve the scene outside the confirmed rectangle, camera geometry, and lighting.

Here the loop reaches a real generator, and the contract is locality. The edit was allowed to change one rectangle (x 230–1152, y 48–912 of 1280 × 960) and nothing else. Locality is measured, not assumed: everything outside that rectangle is compared to the source image, and an output passes only if the outside-mask similarity (SSIM) clears 0.950.

What to look for The raw generator output scored 0.175 against the 0.950 threshold — rejected, because it had redrawn things outside the edit region. The pipeline then cut the region, generated only that block, and composited it back. Outside the mask the composite is pixel-exact: 0 changed pixels out of 432,192.

This is the learning half of the loop: taste moves only through explicit, recorded judgments, never as a side effect of use. The question here: can the system’s taste move, and only when it is deliberately retrained? The model is a pairwise ranker over 10 frozen visual features — given two images, it returns the probability that one is preferred. Two disclosed rule-sets, policy A and policy B, disagree about what “better” means. The model is trained on policy-A judgments and published as an immutable snapshot; policy-B judgments are appended and a second immutable snapshot is published. 8 held-out pairs on which the two policies disagree are then replayed against both snapshots.

What to look for Before retraining, the model gives policy B’s choice a mean probability of 0.000248 — effectively never. After retraining: 50.1%. Snapshot A’s answers do not change when B is published; the taste moved only when the system was retrained, and the whole movement replays from the two frozen snapshots. The labels are simulated rules, disclosed as such (recorded class policy_simulated); the claim is only that the signal measurably moved.

This is the retrieval step that starts the loop: before any pixels move, the stated intent picks the evidence. Two edit intents on the same photograph — replace one lemon, restyle the whole frame — both retrieve their references from the same library. The question: does the stated intent actually change what an edit consults?

What to look for The two top-three retrieved sets share zero assets: intersection 0, union 6, Jaccard 0.0. The replace intent consulted lemon references; the restyle intent consulted Claude Lorrain landscapes. One library, one hard filter per intent — and the record below shows exactly which assets each intent saw, with scores and content refs.

Recorded interpretation: structural_routing_control_not_learned_retrieval_quality. Artifact status measured_run.

The run that produced every number above, pinned: schema, board hash, engine and encoder revisions, conformal parameters, seed, and the exact command.