The Composite Portfolio Score: a methodology paper

Why a composite, not a single framework

RICE and WSJF are both useful and both incomplete. RICE rewards reach and confidence but undervalues time-sensitivity. WSJF surfaces time-sensitivity through cost of delay but is notoriously sensitive to gameable inputs. Running either in isolation produces predictable distortions in the resulting backlog.

A composite that combines both, then tempers the result with a bounded AI judgment input trained on prior portfolio decisions, produces a score that survives the most common gaming patterns and remains interpretable to a human reviewer.

Inputs the model uses

Every scored initiative provides the four RICE inputs (reach, impact, confidence, effort) and the three WSJF inputs (business value, time criticality, risk reduction or opportunity enablement, plus job size). The platform sources these from the intake submission, the finance review, and the strategic objective link.

A bounded AI judgment input is computed last. It is constrained to a ±15 percent adjustment on the composite, intentionally limited so the model can apply contextual reasoning without dominating the human-supplied inputs.

01 · Math· original to Vantage

The composite formula

02 · Constraints· original to Vantage

The bounded AI judgment layer

The judgment layer reads the initiative's narrative fields, the connected strategic objective, the prior six months of similar initiatives, and the current portfolio mix. It returns a single signed adjustment in the interval from negative fifteen to positive fifteen percent, with a one-sentence rationale that is logged alongside the score.

The ±15 percent bound is deliberate. It is large enough to break ties and correct for obvious framework distortions, and small enough that the score remains attributable to the human-supplied inputs. The rationale string is part of the audit trail and is included in every score export.

03 · Robustness

Failure modes the system tolerates

A scoring system is only as useful as its behavior on bad inputs. The composite is built to survive five common failure modes.

1. 01

   Input gaming on a single framework.

   A submitter who inflates RICE reach to push a ticket up the queue moves the composite by half as much as they would move a pure-RICE score, and the bounded judgment layer flags the result as an outlier when the inflated input is inconsistent with the narrative.

2. 02

   Missing inputs.

   When fields are blank, the platform substitutes organization medians and lowers the confidence multiplier on RICE. The score still computes, with the confidence reduction visible to reviewers.

3. 03

   Stale strategic weights.

   When a strategic objective is updated, every score that references it is recomputed in the background and the activity log records the change. The portfolio view shows a small indicator next to recomputed scores until they are reviewed.

4. 04

   Model drift.

   The AI judgment layer is calibrated against prior leadership decisions. When the calibration window exceeds the configured drift threshold, the layer's contribution is automatically reduced to zero until the calibration is refreshed.

5. 05

   Disagreement between human reviewer and model.

   Reviewers can override the score. Overrides are tracked and feed back into the calibration loop. Persistent disagreement on a class of initiatives is surfaced as a calibration alert to admins.

Calibration and audit

Calibration is continuous. Every reviewer override, every promoted-to-shipped initiative, and every retrospective rating feeds the calibration store. The judgment layer's contribution is re-weighted against this store on a rolling basis.

Every score is auditable. The composite, the per-framework subscores, the judgment adjustment, and the rationale string are all retained for the life of the initiative and exported together when leadership requests a portfolio audit.