AI Opportunity Scoring Transparency in 2026 — Publish the Weights, Inputs, and Override Path

The direct answer

AI opportunity scoring transparency is the practice of publishing five artifacts before the score reaches a forecast. The published weights identify the relative importance of each input signal in the scoring model, and the model card is the versioned, named-owner-controlled artifact that lives in a controlled-document register. The input taxonomy identifies every signal the model consumes, the freshness window for each signal, and the staleness threshold that triggers a re-score. The override path identifies the named owner who can override a score, the override justification field, the ratification log cadence, and the executive escalation path when the override rate exceeds the documented threshold. The audit log records every score assignment with the model version, the input snapshot, the downstream consumer, and any override decision; the log is the audit trail when a downstream dispute asks how a score reached a forecast. The post-mortem sample is the periodic, named-owner review of overridden or contested scores, what the sample reveals, and how the sample changes the next model iteration.

Boundary vocabulary before the model ships

Anthropic's engineering guidance on building effective agents distinguishes workflows — systems where models and tools run through predefined code paths — from agents, which dynamically direct their own processes and tool usage, and recommends starting with the simplest solution possible (ev-aost-002). The point is the boundary vocabulary. A workflow-driven score against a fixed rule is deterministic and re-runnable; the score can be regenerated on demand without a per-record audit beyond the standard event log. An agent-driven score reasons over inputs that change; the score is a per-record decision that requires an override audit whenever the inputs change. Most opportunity scoring pipelines conflate the two: a static score from a static model is logged as a workflow, but the moment the model ingests a fresh signal (an inbound intent hit, a funding event, a champion change), the score becomes an agent decision and the per-record audit must fire.

The protocol has three operating modes. Mode one is the workflow score: deterministic, re-runnable, logged as a single event. Mode two is the agent score: per-record decisions, logged as the full audit record (weights version, input snapshot, consumer, override decision). Mode three is the override: when the named owner overrides a score, the override justification, the timestamp, and the downstream effect are logged, and the executive committee reviews the override rate at the cadence the protocol names.

Published weights, versioned and signed

The published weights are the first artifact the team owes the rep, the CRO, and the auditor. The weights are the model's expression of which signals matter: how much weight on company size, how much on intent signals, how much on engagement history, how much on stage velocity. Publishing the weights is not publishing the model source code; publishing the weights is publishing the relative importance of each signal so the rep can reason about why a deal scored high or low. The published weights are a controlled artifact: versioned, named-owner-signed, and stored in a controlled-document register before the model goes live.

The OECD AI Principles — Transparency & Explainability (May 2019, updated May 2024) — require AI actors to commit to transparency and responsible disclosure so that those affected by an AI system can understand the outcome and challenge it based on plain-and-easy-to-understand information; for high-impact applications including those that affect consequential decisions about individuals (employment, credit, and consequential commercial decisions including opportunity prioritization), transparency should include information about the factors that influenced the AI system's output and the data used to train or operate the system (ev-aost-003). In opportunity scoring, the published weights are the factors that influenced the output; the publication is the transparency the OECD principle asks for.

Input taxonomy and the freshness window

The input taxonomy is the second artifact. Every signal the model consumes is named, the freshness window for each signal is documented, and the staleness threshold that triggers a re-score is published. A signal older than its freshness window is a stale signal, and a score computed from a stale signal is a stale score. The freshness window depends on the signal. Job title freshness may be six months; intent signal freshness may be thirty days; funding event freshness may be seven days; technographic stack freshness may be ninety days. The input taxonomy is the operating definition of which signals the model reasons over; the freshness window is the operating definition of how fresh those signals must be; the staleness threshold is the operating definition of when the score must be re-computed.

The discipline is that every signal in the taxonomy carries a documented freshness window and a documented staleness threshold. Without the windows and thresholds, the team cannot answer the post-mortem question about why a stale signal flowed into a fresh score; with them, the team can answer the question by pointing to the taxonomy and the freshness log.

Override path as a first-class design surface

The override path is the third artifact. Opportunity scoring that cannot be overridden is a black box; opportunity scoring that can be overridden arbitrarily is a rubber stamp. The override path is the middle: a named owner who can override, an override justification field that must be filled, a ratification log that records the override, and a cadence at which the executive committee reviews the override rate. The override rate is the operating evidence that the override path is healthy: a zero override rate means the model is treated as infallible, an override rate above the documented threshold means the model is being treated as advisory.

The OECD AI Principles — Accountability (May 2019, updated May 2024) state that AI has risks and all actors must be accountable; in an opportunity-scoring context, the accountable actor is a named human — typically the VP-Sales or the deal-desk owner — who ratifies the scoring model's weights, the input taxonomy, and the override path before the model goes live, and who signs off on each override after the fact; the principle makes that ratification and override log a governance requirement rather than a courtesy (ev-aost-005).

Audit log per score assignment

The audit log is the fourth artifact. Every score assignment is logged with the model version, the input snapshot at the moment of scoring, the downstream consumer of the score, and any override decision. The log is the audit trail when a downstream dispute asks how a score reached a forecast. The log is also the data source for the post-mortem sample. Without the audit log, the team cannot reconstruct what the model saw when it produced the score; with the audit log, the team can replay the score against any later change in inputs to understand why the score moved.

NIST AI RMF 1.0 (January 2023) — Govern 1.3 — requires processes, roles, and responsibilities for managing AI risks including pre-deployment review, ongoing monitoring, and clear documentation of system changes (ev-aost-004). The audit log is the documentation that satisfies the practice; the named owner of the log is the human accountable for the practice.

Post-mortem sample, and how it changes the next model

The post-mortem sample is the fifth artifact. At the cadence the protocol names (typically per quarter), the named owner reviews a sample of overridden or contested scores, identifies patterns the sample reveals (which inputs drove the override, which inputs were stale, which weights were contested), and feeds the patterns back into the next model iteration. The sample is not a quality-control exercise; it is the mechanism by which the model learns the team's preferences over time. A model without a post-mortem sample is static; a model with a post-mortem sample is a model whose weights evolve with the team's reasoning.

The discipline is that the sample is small enough to be reviewed, large enough to be representative, and structured enough to identify patterns. A 5–10% sample of overridden scores per quarter is a reasonable starting point; the structure is a table with score, override reason, contest reason, and outcome. The patterns the table reveals — for example, a recurring override reason of stale intent signal — are the inputs to the next model iteration.

Audit cadence inherits NIST AI RMF Govern 1.3

The audit cadence has three layers, each named-owner and each written down before the next scoring model goes live. The pre-deployment review is the layer at which the model is onboarded: the published weights, the input taxonomy, the freshness windows, the override path, and the audit log schema are published and signed. The ongoing-monitoring layer is the per-forecast-commit review at which the named owner checks the override rate, the stale-signal rate, and the audit-log completeness. The post-mortem-sample layer is the per-quarter review at which the named owner produces the sample, identifies patterns, and feeds the patterns back into the next model iteration. Without the cadence, the scoring-transparency framework is a documentation exercise rather than a control.