Absolute per-cell trainability floor: a yield shortfall must DROP the cell, not shrink it to 1 row

Goal

Close the gap that let a fine-tune cell train on 1 row and flow all the way to monitor correlations with no gate firing: the yield-shortfall contract specifies "shrink + report" with no ABSOLUTE minimum below which a cell is DROPPED and the experiment's denominator revised.

The incident (#2221)

#2221 built 24 real-data-twin training mixes by judge-banding found/organic responses, then equalizing down within family to floor-N. Realized per-cell rows:

family	rows/cell
evil	1
mistake_opinions	13
sycophancy	175
mistake_medical	233
mistake_gsm8k	520
mistake_math	760
hallucination	1,533
insecure_code	3,357

Plan §4 targeted "~a few k" rows per cell. Evil got 1. All 24 cells were fine-tuned (rs-LoRA r32, lr 1e-5, 1 epoch), captured, judged with a 6-draw Sonnet graded rubric, and carried into the monitor-arm correlations and a checkpoint-detection AUC.

Downstream cost: evil's DV floored at 0.00-1.21 (base 0.00) and mistake_opinions showed no install, so 6 of 24 cells were effectively untrained; the only trait-acquiring cells were the 2 largest families, making "which cells acquired the trait" substantially a proxy for "which family had enough rows" — a confound on the headline comparison, plus wasted judge spend on 6 non-conditions.

Why nothing fired

Plan §7's gate row read Per-cell realized yield >= floor-N (else shrink+report), explicitly "graceful shrink, never abort". The related kill criterion required ZERO band-II rows for >= 2 of 3 chat-trait families; evil produced 1 row (not zero) and the other two chat families were healthy, so no criterion was met. The pipeline followed the plan exactly.

.claude/rules/on-policy-completions.md and .claude/rules/data-realism.md both say a below-floor shortfall is REPORTED and the source DROPPED — never silently carried. The plan's "shrink to whatever survives" wording is compatible with the letter of "shrink+report" while violating that intent, and the critic ensemble did not flag it.

Proposed change (implementer to scope)

1. An absolute trainability floor, separate from the equalize-down floor-N. Below it a cell is DROPPED, not shrunk: the cell does not train, the denominator is revised everywhere it appears, and the drop is named in the clean-result ## Takeaways per the After-Every-Experiment planned-vs-actual rule. Ground the default in literature / prior issues rather than inventing a constant (the LoRA-install evidence in .claude/rules/marker-training-recipe.md is the nearest in-repo anchor for "how many rows install anything at a given lr/epoch"); a plan may override it explicitly with a stated argument, which is exactly the visibility that was missing here.
2. A pre-P4 hard assert in the mix-build/fine-tune dispatch path — a cell whose realized row count is below the floor fails loud BEFORE any GPU or judge spend, rather than producing a silent non-condition. This is the fail-fast placement: the gate belongs where the rows are counted, not in post-hoc interpretation.
3. A planner/critic surface rule so a plan's yield row must state the absolute floor and the DROP disposition, not only the shrink path — the review-side arm that would have caught this plan's wording.

Scope note for the implementer: prefer extending the existing yield/band-count machinery over adding a parallel gate, and check whether a shared helper already counts realized per-cell rows before writing a new one (reuse-discovery rule).

Provenance

Surfaced by the #2221 orchestrator while composing that issue's clean-result: the 1-row evil mixes were found by counting the realized mixes on disk and confirming them byte-identical to the HF copies.