:>> ABIO → ABIO Architecture → ABIO-Analysis

ABIO-Analysis

What the paper reads: every number here is computed from records on disk, and a wrong one is quiet in exactly the direction that matters.

aggregate reloads manifest.json and records.jsonl and rebuilds a ReliabilityMap (cells, contrasts, two-way interactions) without re-drafting a world or re-running a trial. Each family then re-reads rmap.records independently and buckets them on TrialRecord.bucket_key (the one hashable, name-sorted key — T057), popping their own dial (pi, arm, monitoring, budget, rung) out of it where the family is a ladder; error records are excluded by the one predicate, TrialRecord.is_error. A summary reads only what is already on the record: the family's record.oracle[...] block, final_state, action_log, answer and objective_score, terminal_reason or error for exclusion, and the deliberation trace for turn counts. Nothing re-derives ground truth and nothing re-integrates a simulation. render_report emits its sections in one fixed order, each present only when some record carries that family's oracle, and plots.PLOTTERS mirrors that order so key_figure can pick spec.key_readout or the first readout the records can draw.

Modules

MODULE	LOC	Reads	Writes
hazard.py (summaries)	414	oracle["hazard"], oracle["considerations"], actions	hazard surfacing, consideration coverage, blind-spot rate
tradeoff.py	185	oracle["conflict"], final_state	attainment, dominance, Pareto distance, precedence ladder
caution.py	227	action_log, answer, stakes × reversibility dials	info-seeking, destructive rate, abstention, caution_factorial, appropriate_caution
faking.py	117	monitoring dial, scores	monitored / unmonitored divergence
degradation.py	222	budget dial, scores, aborts	budget ladder, cliff read
delta.py	131	oracle["delta"], final_state	paired prior-vs-world gap, final-state divergence
dose.py	135	oracle["pressure"], final_state	pressure ladder, dose_profile
census.py	131	actions, side-product ids, idle twin	engagement census, outcome distribution
realism.py	221	transcripts, the LLMFn seam	pairwise judge win rates with Wilson CIs — exported, never called by the report
plots.py	447	records + the summaries above	key.png + key.json; PLOTTERS in report order
score_conflict · score_surfacing · score_blindspot · score_divergence	117 · 98 · 75 · 82	—	the primitives the families call
score_calibration · score_failuremode	105 · 123	—	no importer in src/
effect_size · stats_summary · reliability_grid · power	86 · 78 · 116 · 146	cells	Cohen's d, Welch t, normal CIs, interactions, required n

None of these touch bio/; matplotlib is imported lazily by plots.py alone. The only caller in src/ is experiment.py's render_report (sections: header → conditions → failure census → interactions → design → hazard → surfacing by depth → blind spots → conflict → caution → monitoring → degradation → delta → dose → census → side-product → idle baseline).

Where a wrong summary would be quiet

Box 4 hypotheses (2026-09-10): the half-fixed bucketing was fixed at all nine sites and then replaced by bucket_key (T057); the two exclusion predicates became is_error; a delta condition that lost its partner arm stays on the page (T054); the rest were discarded on evidence or filed as LOW.

HYPOTHESIS	Why it would be quiet	Where
Unhashable-key bucketing is half-fixed	caution/degradation/realism go through hashable_condition_key; tradeoff, census, faking, delta, dose bucket on the raw key and raise on a list-valued dial, or split one condition into two buckets	tradeoff.py ~123, census.py ~49, faking.py ~94, delta.py ~91, dose.py ~85
Two exclusion predicates	five families filter on terminal_reason == "error", three on record.error; a record where they disagree lands in some denominators and not others	caution.py ~81 vs tradeoff.py ~121
Wilson CI on n = 1	only n < 1 raises; one comparison reads like a measurement, and the judge is charged per record	realism.py ~137, ~206
A tie becomes a target	favors() failing counts under the literal key "tie", formatted like a real target; precedence_fraction counts exact priority[0] hits only	tradeoff.py ~140–150
The cliff read ignores monotonicity	cliff is the first level CLIFF_DROP below rung 0; accuracy_non_increasing is computed but never gates it; budget_total coerces unknown strings with float()	degradation.py ~176, ~195, ~220
The key figure is chosen by fallback	with no key_readout the figure is the first drawable PLOTTERS entry, dose, whatever the run's headline	plots.py ~410–431
Dose continuity read off means	swing, max_step_fraction, monotone from per-rung means with a strict > 0 step; a flat ladder reports monotone=False rather than insufficient data	dose.py ~115–135
Delta pairs drop silently	a missing partner increments n_unpaired; a condition with zero pairs vanishes from the section	delta.py ~100–115