:>> ABIO → ABIO Architecture → ABIO-Runtime

ABIO-Runtime

The paid-run blast radius: everything between a drafted world and the frozen record, including the one edge that reaches the network.

runner.run narrows the turn-0 truth first (narrow_observation, then project_observation for task.setup["hidden_ids"]) and builds the brief from exactly that view, so a probe can only name an id the agent can already see. When opaque names are requested it builds a seeded NameMap and wraps the agent in OpaqueAgent; every crossing after that is translated at the boundary and the runner never leaves structural ids. Each turn rebuilds a fresh WorldImpl from the prior end-state, redraws only observation noise (the hidden set is frozen at the turn-0 draw), fires every_turn probes, calls agent.act, then after_action and at_commit probes whose answers land only on probe_records. Validation is rejection-as-data: an unknown id, a non-finite value or a negative setpoint is recorded as rejected, an over-cap Intervene is clamped and stays accepted, and every attempt appends an ActionRecord with accepted, reason, requested value and applied delta. The loop stops on Commit, illegal_action_limit, budget exhaustion or max_turns; the answer is graded; only then does audit_prompts scan every prompt actually sent — hidden molecules, key tokens, and every reaction id the brief did not hand the agent — and a hit raises TaintError carrying the full record; any other mid-trial exception raises TrialError carrying the partial record. MassTrialRunner enumerates (condition, i) units, derives each seed from a label with the matched dials removed so arms share worlds, runs them on a thread pool, lands results in submission order so persistence stays ordered, consults the cost ceiling after every landed unit and as each worker starts, and under on_error="record" turns a failure into an error line rather than a dead grid.

Modules

MODULE	LOC	Role
agent.py	311	The action algebra (Measure, Intervene, Commit, Wait), the Agent / SessionAgent / ProbeAgent protocols, ScriptedAgent and Policy. No I/O, no world access.
brief.py	716	TaskBrief, Affordances, build_brief, render_brief; constitution, stakes, monitoring, probes, protocol, padding, DEFAULT_ACTION_COSTS.
naming.py	236	NameMap (injective, longest-first), build_name_map, OpaqueAgent, surface_brief / surface_observation / structural_action.
ops.py	~110	LLMFn, the one provider seam; LLMOp validates and retries a schema'd reply; canonical prompt text.
llm_agent.py	1130	LLMAgent (history window, compact_at, history_token_limit, token ceiling, prompt_texts/hashes, probe forking), UsageMeter, prices, PINNED_MODEL, _call_with_retry, default_anthropic_llm_fn.
trial.py	237	Frozen TrialRecord + ProbeRecord; condition_key, hashable_condition_key, final_state_dict.
runner.py	939	run, Budget/BUDGET_LADDER, _world_from_state, probe timings, audit_prompts, TaintError.
mass_trial.py	596	MassTrialRunner.run/run_one, condition_grid, ReliabilityMap, CellSummary, ContrastResult, Provenance, aggregate_records.

Layering is strict and acyclic: agent/ops → brief → naming/trial/llm_agent → runner → mass_trial. runner.py never imports llm_agent.py: it duck-types SessionAgent/ProbeAgent and reads usage, prompt_texts, compaction by getattr, so the live layer is swappable. Only runner.py and mass_trial.py touch bio/.

Where a paid run could mis-record

Box 4 hypotheses (2026-09-10): the error record keeps its transcript (TrialError), the ceiling is consulted (box 4), taint lines stay put on resume and the census names the class (T054), the constitution's exit is stamped (TrialRecord.forgetting, T054), a shared molecule/reaction id refuses (T054), the audit covers hidden reaction ids (T054); probe spend and the clamp note were confirmed as designed.

HYPOTHESIS	Why it would be quiet	Where
An error record drops its transcript	the generic error path keeps only usage; brief, turns, action_log, probes, final_timeline are gone, so a $-charged trial shows an empty transcript; a factory-stage failure keeps no usage at all	mass_trial.py ~530–540
A systematic taint shrinks n silently	a TaintError record is re-stamped terminal_reason="error" and excluded from cells and contrasts; only Provenance.failed_trials counts it, without the reason	mass_trial.py ~529, ~593
The cost ceiling overshoots by up to concurrency	stop() is checked at submission; spend advances only when a unit lands, in submission order, so one slow early trial defers all accounting	mass_trial.py ~573–584; experiment.py ~2741
Probe spend never trips the token ceiling	meter_tokens subtracts _probe_tokens; every_turn probes on a long grid spend past the ceiling until the USD ceiling catches it between trials	llm_agent.py ~905
The constitution vanishes without a trace	stored as history turn −1, it is always inside the displaced set at compact_at; an unusable summary replaces it with a stub; a finite memory=k window drops it with no field on the record saying when	llm_agent.py ~751, ~826, ~852–876
A molecule and a reaction sharing an id	build_name_map writes molecules then reactions into one dict; the reaction wins silently and structural_text maps the surface name back to the wrong entity	naming.py ~112–118
The taint audit has blind spots	returns () when the agent exposes no prompt_texts; secrets are molecules and AnswerObjective.key leaves only, so hidden reaction structure and OutcomeObjective targets are never scanned	runner.py ~907–914
A clamp note for a compacted turn is dropped	notice() returns silently when no history entry matches outcome.turn	llm_agent.py ~769