Signal Lab

The Signal Lab — the minigames that asked you to compare papers, the model fitted from your answers, and the ranking heads it fed — is not part of this line of ALMa. ALMa's core is Feed, Discovery and Library (product decision D24); the Lab was extracted under D25 so the core carries no game surface, no fitted-model view and no additive taste heads on the score.

Nothing was lost:

• The complete implementation is preserved on the git branch feature/signal-lab, at its best measured state.
• Your answers are preserved. The signal_lab_rounds table and its migrations are left in place; nothing reads them here, and nothing deletes them.

Why it left

It was measured before it was removed, and the measurements are the reason:

• Nobody played it. Every organic round was answered in a two-day window in late July 2026. None in the eight weeks that followed.
• The fit was starved by design. One three-way comparison carries roughly one bit of information. A few dozen rounds were being asked to fit several hundred free parameters — one offset per region, one per venue, plus a 768-dimensional direction. An honest estimator reports silence on that ratio, and it did: the region head was silent by its own shrinkage, the author head had no reach, the utility head measured null-to-negative.
• The instrument could not settle it either way. ALMa's outcome evaluation resolves differences of about ±0.011 AUC, bounded by how many papers you have rejected — not by how much you played. Every head's effect was smaller than that. More rounds would not have changed the answer.

What stayed

The shared machinery the Lab used but did not own is still here: the beta-binomial shrinkage in core/scoring_math.py, the per-install scoring calibration, and the outcome evaluation that fits family weights from what you actually keep and reject.

The semantic partition and its regions left with it — see Maps for that history — because with the Lab gone they had no consumer, and a clustering job that feeds nothing is not worth running.

If it comes back

It would not come back in this shape. The design that would earn its place is written down: three parameters instead of three hundred, matched pairs that isolate one axis, questions chosen for how much they would actually teach, and output scoped to a single Discovery branch rather than applied globally. That plan lives with the task notes, not in the shipped docs.

What is already decided is where its output would go: into branch-local ranking weights (D26), never a global taste term.