Plate-budget acquisition

[smcolby]

Replaces the flat k_per_iteration acquisition parameter with a physical plate-budget system and enhances the cost bar chart in the live dashboard to display per-category query counts alongside dollar costs on hover.

---

Motivation

k_per_iteration was a dimensionless count that didn't map to real experimental constraints. In practice, a wet-lab run fills a physical plate: each Primary Screen (PS) consumes 1 well, each Dose-Response Curve (DRC) consumes many more (e.g., 13 for a 13-point singlet DRC). Expressing the budget in wells rather than queries makes the config directly interpretable by experimentalists and naturally handles the asymmetric cost of DRC vs PS queries.

---

Changes

moal/config.py

ActiveLearningLoopConfig drops k_per_iteration and gains three new fields:

Field	Default	Description
plate_size	1536	Maximum wells per plate (one plate = one iteration)
wells_per_ps	1	Wells consumed by a single PS query
wells_per_drc	13	Wells consumed by a single DRC query

Backwards compatibility: plate_size=k, wells_per_ps=1, wells_per_drc=1
exactly replicates k_per_iteration=k behaviour.

moal/acquisition.py

CostAwareGreedyAcquisition.select() replaces the k: int parameter with plate_size, wells_per_ps, wells_per_drc. The selection loop now:

1. Walks the unified ranked candidate list (highest score first).
2. Tracks wells_used as each candidate is considered.
3. Hard-stops when the next candidate's well cost would push wells_used above plate_size — it does not skip that candidate and search for a cheaper one. Remaining candidates are deferred to the next iteration, where the model will rescore them on the updated labeled pool.

The undersupply warning message is updated to report plate_size, wells_per_ps, and wells_per_drc rather than counts.

moal/loop.py

• ActiveLearningLoop.run() signature updated: k_per_iteration → plate_size, wells_per_ps, wells_per_drc.
• Both acquisition.select() call sites updated (pre-loop seed + in-loop body).
• Campaign startup banner updated to display well budget instead of query count.
• Progress step label updated.
• Per-category query counts (iter_n_ps, iter_n_drc_new, iter_n_drc_upgrade) are computed from new_records after each oracle query and forwarded todashboard.update().

moal/dashboard.py

LiveDashboard.update() gains three optional parameters (defaulting to 0 for backwards compatibility with any existing serialized snapshot data):

• iter_n_drc_new — new first-pass DRC queries
• iter_n_upgrades — PS→DRC upgrade queries
• iter_n_ps — PS queries

These are stored per-iteration in the snapshot dict and surfaced as customdata on the three stacked bar traces in the per-iteration cost breakdown panel (panel 2). Hover tooltips now read, e.g.:

DRC
10 queries
$130

PS→DRC
3 upgrades
$39

PS
1200 queries
$1200

moal/cli.py

Extracts plate_size, wells_per_ps, wells_per_drc from config and passes them through to loop.run().

examples/default_config.yaml

k_per_iteration: 100 replaced with the three new fields and descriptive inline comments:

active_learning_loop:
  n_iterations: 20
  plate_size: 1536       # Maximum wells available per plate (per iteration)
  wells_per_ps: 1        # Wells consumed by a single PS query
  wells_per_drc: 13      # Wells consumed by a single DRC query

---

Tests

tests/test_acquisition.py

All existing select() call sites updated to the new signature. Two new tests
added:

• test_drc_cost_stops_at_plate_boundary — verifies that a DRC candidate that would overflow the plate causes the loop to stop, even though there are remaining candidates with lower scores.
• test_wells_used_never_exceeds_plate_size — property-style test across a range of plate sizes confirming wells_used never exceeds plate_size.

tests/test_loop.py

K = 5 constant replaced with PLATE_SIZE = 5, WELLS_PS = 1, WELLS_DRC = 1 (equivalent behaviour). All loop.run() calls updated.