How the pWin capture assessment works

pWin is FedGrade's capture-decision engine. For a single (opportunity, vendor) pair it computes a win-probability score out of 100, a bid / no-bid recommendation, and — the part that matters — the named, individually-graded factors that produced the score. It is a Team+ capability.

What you get back

An assessment carries:

• A composite score (0–100), computed as baseline 50 + the sum of each factor's contribution, clamped to the range.
• A set of factors, each with a label, a direction (helps / hurts), a contribution to the score, an evidence grade, and a one-line rationale.
• A rolled-up evidence grade for the whole assessment.
• A recommendation: bid, lean-bid, no-decision, lean-no-bid, or no-bid.
• A deterministic summary naming the top two drivers.
• A model version and an inputs hash so every score is reproducible and traceable to the evidence that made it.

The deterministic core

The score is not a black box and not an LLM guess. It is a transparent weighted sum: baseline plus each factor's contribution. The same inputs always produce the same score. You can read every factor's contribution and see exactly why the number landed where it did.

An optional LLM narration can write a 2–4 sentence capture-analyst rationale on top of the already-computed factors. The narration is commentary only — it can never change the score, the recommendation, or any factor grade. If the language model is not configured, the assessment still returns in full, just without the narrative paragraph.

How the grade rolls up

This is the anti-blackbox guarantee. The assessment's grade is the weakest grade among the factors that actually carried weight:

• If every factor is insufficient-evidence, the whole assessment is insufficient-evidence.
• If any weighted factor is only modeled (from priors or registry rather than real award data), the assessment is at best modeled.
• Only when all weighted factors are verified (award-anchored) is the assessment verified.
• If fewer than half the factors carried weight, the grade is capped at modeled no matter how strong the few were.

A high score sitting on thin evidence is honestly downgraded — never surfaced as a confident bid.

How the recommendation is gated

The grade gates the strength of the call, not just the score:

• On insufficient-evidence, the recommendation is no-decision regardless of the score — underwriting a bid on data you don't have is the exact failure this engine exists to prevent.
• On modeled, the call is softened by a notch (lean-bid / lean-no-bid / no-decision).
• On verified, the full-strength call is available (bid / no-bid down to lean calls near the middle).

Calibration

When pWin assessments resolve to real outcomes (won / lost), FedGrade scores how well-calibrated each model version has been for your workspace. This is tenant-private — your calibration is never pooled with another workspace's and never published. Pre-data, the ledger ships correct-but-empty (no fabricated track record). See The pWin calibration ledger.

Where to use it

• From a Case at /cases, run an assessment against a pursuit and pin it to the pursuit's evidence.
• It feeds the Opportunity ranking — the rank engine consumes pWin read-only.
• It rides inside the capture package export.
• Over the API at POST /api/v1/pwin (scope pwin:assess) and as the assess_capture MCP tool.

Honest limits

pWin factors deepen as award data ingests. Early in a research library, more factors land at modeled or insufficient-evidence, and you will see more no-decision recommendations — that is the engine being honest, not broken. As vendor and agency award history accumulates, more factors become award-anchored and the calls sharpen.