Assumptions
Everything the model believes, stated before the math. If you do not buy one of these, do not buy the output.
- Every queue item starts from a real unknown or conflict on the record; items without a source reference are rejected and listed - never cosmetic completeness.
- Quantitative value requires calibrated evidence-state probabilities, a stated acquisition cost, and a decision value scale; without all three the output is priority, never dollars.
- Blocking evidence outranks cosmetic completeness by construction: materiality first, then decision cells affected, then acquisition time.
- Resolved evidence leaves the queue and is counted, not ranked.
- Source accessibility and price are never invented - unpriced acquisition stays unpriced.
- The stop rule responds to the decision threshold; items under it are shown as not worth pursuing rather than silently dropped.
Derivation
The queue holds only unresolved items anchored to a named unknown or conflict. Ranking is a deterministic lexicographic order: materiality (blocking before decision-relevant before non-material), then the count of decision cells the evidence moves, then acquisition time.
An item earns a dollar value only when it carries calibrated state probabilities, a stated cost, and a decision value at stake per swing. Then:
with V the decision value per swing, p_s the calibrated state probabilities, and C the acquisition cost - the probability-weighted swing value minus cost, floored at zero. A negative value is never sold as a reason to act. Missing any of the three inputs, the item prints priority with the reason, not a number.
The stop rule walks the ranked queue: blocking items always count as decision-moving; a priced item must exceed the decision threshold. From the first item that cannot move the decision, everything below is marked not worth pursuing - shown, not hidden.
Identification strategy
Why the estimate supports the claimed meaning - and where the model refuses to claim more than the data carries.
- Anchoring to a source unknown is the identification rule for relevance: an item exists because the record has a real gap, which is what separates diligence value from completeness theater.
- The priority-or-dollars split keeps weak inputs honest: without calibrated probabilities, a stated cost, and a value scale, no number exists - and the page of record says why.
- The ranking key is the identification of urgency: what blocks the decision outranks what merely informs it, and ties break toward the evidence that moves more decision cells and arrives sooner.
- The stop rule is stated in terms of the decision threshold, so 'not worth pursuing' is a computed consequence, not a judgment call.
Inputs and outputs
Reads
- Diligence items
- Each anchored to a real unknown or conflict on the record, with field, materiality, and decision cells affected
- Acquisition terms
- Days and cost to obtain, when documented; never invented
- Pricing inputs
- Calibrated evidence-state probabilities and decision value per swing, required for any dollar figure
- Decision threshold
- The swing size that would change the decision
Writes
- Ranked queue: rank, priority class, qualitative priority, stop-rule flag
- Dollar value per fully priced item - floored at zero, labeled low confidence
- Rejected items and resolved-excluded count
- Items not worth pursuing under the stop rule, with the rule printed
Calibration and gate evidence
Validation design
- Gate-verified pricing discipline: a fully priced item must produce a deterministic value, an unpriced item must print priority-never-dollars, and the stop rule must respond to the threshold.
Evidence on record
- Smoke gates at seal (8 m33-specific): a cosmetic item without a real unknown is rejected; resolved evidence leaves the queue; blocking outranks cosmetic completeness; an unpriced item is priority, never dollars; a fully priced item gets a deterministic value; the stop rule responds to the threshold; determinism (identical input hash on re-run); audit row written for every run.
Gates at seal
- smoke: 8 m33 gates incl. cosmetic-gap rejection, priority-not-dollars, stop-rule response, determinism + audit row (suite counts at seal per the build agent's seal report)
- sealed at 36389cf on atlas-nl-shock-redesign
Failure modes
The ways this model declines, degrades or goes UNKNOWN - and what it prints when it does.
- Item without a source unknown or conflict: rejected and listed.
- Missing probabilities, cost, or value scale: priority ranking only, with the missing terms named.
- No item can move the decision across the threshold: the whole tail is marked not worth pursuing, and the stop rule prints.
Sensitivity
- The decision threshold is the stop rule's only dial; lowering it resurrects marginal items, and the queue shows exactly which.
- Among same-materiality items, acquisition time breaks ties - a slow document can rank below a fast one of equal importance, which is the intended reading when a decision clock is running.
Iron rules
Priority, never dollars, without calibrated probabilities, a stated cost, and a decision value scale.
Source accessibility and price are never invented. Resolved evidence leaves the queue.
Source and audit
Written from the sealed model source. Every run is recorded - model version, frozen inputs, input hash, timestamp. There is no unrecorded model execution.