Coverage

dial	what it varies
chain_length	Number of steps in the causal event chain (object moves/transfers) the model must track.
scramble	0.0-1.0: how much the narration order is shuffled relative to true chronology (every sentence still carries its real timestamp).
coref_pressure	Number of extra coreference-stressing mentions (pronouns, repeated near-identical entities) injected into the story.
distractor_events	Number of plausible-looking but query-irrelevant events involving other entities/objects mixed into the narration.
filler_events	Number of narratively neutral padding events that lengthen the story without touching the query's answer.
witness_cutoff	0.0-1.0: how early an entity's last witnessed exposure to the object is cut off relative to the object's later moves -- stresses whether a model over-generalizes 'last witnessed' into 'still watching'.

trap	mechanism
final_true_state	The object's actual final location -- the most damaging trap for belief queries: what a model says if it collapses belief into reality.
last_mentioned_location	The last location named in narration order, regardless of true chronology or who witnessed it.
last_witnessed_by_anyone	The destination of the last move ANYONE witnessed, not necessarily the query agent.
query_agent_own_last_move	If the query agent themselves moved the object, their own last such move's destination (only applies when the agent moved it).
most_frequent_location	The most-mentioned destination across the whole story, trusting mention frequency over recency or truth.

The plausible-but-wrong answers a shallow reader gives on a belief query; adversarial seed search picks the worlds that pit a trap most sharply against the truth. Every other province has its own dials and traps, listed per campaign in docs/campaign-catalog.md.

tool	what it does
escalation	Walks a single ray through the full joint difficulty space (all dims at once, base to generator ceiling) to find where a model actually breaks -- see cartographer/schedule/escalate.py.
resample-gate	The scheduler's explore/refine waves re-probe weak/ambiguous signals with additional resamples before committing to a boundary, instead of trusting a single trial.
ablation	Deterministic, answer-key-preserving prompt transforms over the same latent world, used to causally isolate which surface factor a failure actually depends on -- see cartographer/ablate/ablations.py.
fresh-seed confirmation	Re-tests a decompose hypothesis on seeds that were never used during the search that produced it (seed_base >= 10,000 by convention), reporting an effect size + p-value rather than a bare pass/fail -- see cartographer/confirm/confirmation.py.