The loop was missing its transition, and two demos were flattering themselves

[di-omics]

The observation

This repo already had every stage of a perception-action loop lifted into a named, swappable module except one.

labcv.synth.tip_view and column_view are observation models: they take a scalar physical state and render the frame a camera would see. verify reads a frame back. decide chooses an action. What was missing was the function that says what the action did, and it lived as two magic constants at demos/pipette_cam/run.py:47 applied inline nine lines later:

rewash_leave: float = 0.12   # a re-aspirate leaves ~12% of the residual
drydown: float = 0.35        # an extend-dry evaporates down to ~35%

That function is a world model. It now has an interface.

What is added

• labcv/bridge.py -- promotes the action and verdict types out of the demo, because a library must not import from a demo. Load-bearing: dynamics.py cannot be written without it.
• labcv/dynamics.py -- Transition protocol, AnalyticTransition (the constants, finally named), PlantedTransition, InjectedTransition.
• labcv/episodes.py -- action-conditioned rollouts. This is the new capability: the repo produced passive observations, and these are sequences where the scene changed because of a commanded action.
• eval/rollout_metrics.py -- terminal error, in-tolerance, trajectory RMSE, first divergence step.

Unknowability is constructed, not asserted, and the build refuses rather than warns: an occluder must cover 100% of the target ROI in every frame, checked pixel-wise against the boxes synth.microplate returns and again by byte-identity of that region against a render at the other level.

A scoring pack withholds coefficients and future frames but still carries knowability and the hidden latent, because the consumer's separation gate needs that latent to check the labels at all. That leaves a real gaming channel, so the manifest now names it in disclosed_to_harness: a submitter reading the file directly can abstain on exactly the right episodes, so a scoring split is an adversarial holdout only where the harness mediates access. Better stated in the file than in a README beside the data.

Two demos were asserting things they had not measured

run_qc.py set the corrected volume to the target exactly, which is a transition model claiming every correction lands with zero error. It now commands against the camera reading. Every published number is unchanged (the residual stays inside tolerance) and the demo now says corrected wells land 0.24 uL from target rather than on it.

vocab_vlm computed an escalation mask and then sent every box to the VLM anyway. The published "52% of VLM calls saved" described a branch nothing took, and no test could have caught it because the accuracy was identical either way.

The gate now gates. At the published config the numbers hold and are finally true: 13 calls made instead of 27, accuracy still 1.000, 52% saved.

At other configs, enforcing it exposed what the un-enforced version was hiding:

config	accuracy before	after
--seed 11	1.000 PASS	0.889 FAIL
--distractors 8	1.000 PASS	0.750 FAIL

Three specks score above the threshold, take the cheap path, and that path has no word for "bubble". Raising the threshold would fix the number by fitting it to the data it judges, so the demo prints the cost of the saved calls and fails honestly instead.

CI

This repo had house-rules CI before it had test CI, and a red suite sat in it unnoticed, because the unittest runner these modules use aborts without printing a total and tailing its output looks exactly like a pass. Adds test and lint jobs, including a step that counts scored verdicts so a demo cannot stop scoring silently.

Verification

83 tests pass, make all exits 0 with all 8 demos passing, ruff clean under the same rule set the sibling repo enforces, house-rules clean. The pipette_cam block of make all is byte-identical to before this change.