"""One vocabulary for what a tool can and cannot establish. Every tool failure measured in this project has the same shape: a **confident, well-formed answer from outside the tool's competence**, indistinguishable by the caller from a good one. * CellTypist, given a mouse lung sample and a human immune-only reference, called 91% of cells "Double-positive thymocytes". Concentrated, decisive, wrong. * BLAST, given 40 reads from an influenza sample, returned *Homo sapiens*. The host dominates and the virus is invisible at that depth; the answer was about the wrong organism. * Cell2Sentence, given genes averaged over a heterogeneous sample, named a pseudo-cell belonging to no population. None raised. Each was fixed separately, in its own vocabulary -- `abstained`, `reference_is_conclusive`, `can_see_minority` -- which is three spellings of one idea and leaves the next tool free to invent a fourth. This module makes the idea explicit and shared, so a new tool must state its competence to be usable at all. The distinction that matters is between **negative** and **uninformative**. "No influenza found in 40 host-dominated reads" is not evidence of absence, and a caller that treats it as such reaches a confident wrong conclusion faster than one with no tool at all. """ from __future__ import annotations from dataclasses import dataclass, field @dataclass(frozen=True) class Competence: """What a tool result establishes, and what it cannot. Attributes: applicable: Whether the tool was in-domain for this input at all. False means the result carries no information about the question, however confident it looks. confidence: Strength of the finding in [0, 1] when applicable, else 0. margin: Separation from the runner-up interpretation, where one exists. A high confidence with a negligible margin is the out-of-domain signature: the tool had to choose between options that all fit badly. ties: Candidates the evidence cannot order below the leader, at the same relative threshold `margin` uses. Strictly stronger than `margin`, which only looks at the immediate runner-up. Reported alongside `alternatives` because the two criteria fail in opposite directions and neither dominates: measured at the real Cas9 geometry, `alternatives` abstains on 200/200 answers that were all correct, while at the invented one-distractor geometry `ties` commits to 240 answers that were all wrong. Which to gate on is a policy choice with a measured cost on each side, and resolving it needs information from outside the tool's ranking axis. alternatives: Candidates besides the top one that also cleared the tool's evidence floor. Measured on 2,400 simulated libraries: margin is large both when the true answer dominates and when a distractor dominates, so the gap between the top two is not the same as the absence of a rival. A ranking tool that leaves this at 0 is asserting its answer was unique, not merely ahead. blind_spots: What this result cannot rule out, in the caller's terms. Empty only when the tool genuinely could have seen anything relevant. remedy: What would let the question be answered, when the tool cannot. """ applicable: bool confidence: float = 0.0 margin: float = 0.0 alternatives: int = 0 ties: int = 0 blind_spots: tuple[str, ...] = () remedy: str = "" @property def conclusive(self) -> bool: """Whether a caller may act on this result as an answer. Requires four things: the tool was in-domain, it is confident, it separated its answer from the runner-up, and no rival cleared the evidence floor. Confidence alone is what made every failure look like a success -- CellTypist reported 0.885 while choosing between two brain references for a lung sample, on a margin of 0.013. Uniqueness is the condition added after measurement: with exactly one saturated distractor present, the margin-only rule accepted 40% of libraries and was right on none of them, because the distractor cleanly outranked the truth. Requiring uniqueness cut confidently wrong answers from 240 to 28 and raised accuracy on what is accepted from 0.583 to 0.923. """ return (self.applicable and self.confidence >= 0.5 and self.margin >= 0.05 and self.alternatives == 0) @property def informative(self) -> bool: """Whether the result says anything at all, even if not conclusive. A non-applicable result is not a negative result. Distinguishing the two is the point of this class: absence of evidence from a tool that could not have seen the thing is not evidence of absence. """ return self.applicable and not self.blind_spots def explain(self) -> str: """One line a caller can act on, or show to a model.""" if not self.applicable: return f"NOT APPLICABLE: {self.remedy or 'this tool cannot address the input'}" if self.conclusive: return f"conclusive (confidence {self.confidence:.2f}, margin {self.margin:.2f})" parts = [f"INCONCLUSIVE (confidence {self.confidence:.2f}, margin {self.margin:.2f})"] if self.alternatives: parts.append(f"{self.alternatives} other candidate(s) cleared the evidence " f"floor; this is a candidate SET, and the top of the ranking is " f"not the answer") if self.blind_spots: parts.append("cannot rule out: " + "; ".join(self.blind_spots)) if self.remedy: parts.append(f"to resolve: {self.remedy}") return " -- ".join(parts) def not_applicable(remedy: str = "") -> Competence: """A result carrying no information about the question.""" return Competence(applicable=False, remedy=remedy) def bounded(confidence: float, margin: float, blind_spots: tuple[str, ...] = (), remedy: str = "", alternatives: int = 0, ties: int = 0) -> Competence: """An in-domain result that may still be unable to rule things out. `alternatives` defaults to 0, which asserts the answer was unique. A ranking tool must pass the real count; leaving it silent is a claim, not an abstention. """ return Competence(applicable=True, confidence=confidence, margin=margin, blind_spots=blind_spots, remedy=remedy, alternatives=alternatives, ties=ties)