Win-loss synthesis

Use buyer and deal evidence to explain patterns in wins, losses, and no-decisions, with conclusions limited to the sample studied.

Fails. "Price is why we lose. 67% of interviews mentioned it."

Passes. "Price was the primary coded reason in 8 of 12 interviewed losses (12 of 41 eligible losses participated). Direct buyer evidence in 5. Counterexamples: two losses where price was accepted and security stalled the deal. This is the interviewed sample, not the closed-lost population."

For a single competitive asset, use battlecard-builder. For recurring objections across live calls, use call-insights-to-objections.

Inputs

Ask only for the inputs that change the findings:

1. Time window, segment, and the eligible closed-deal population
2. Sources on hand: interviews, CRM notes, call snippets, closed-lost reasons
3. Selection method and the win / loss / no-decision mix
4. Confidentiality, consent, and who will see quotes

If sample accounting is incomplete, proceed with labeled limits. Do not imply the sample represents the eligible population. Do not invent win rates.

Supporting material

Read references/win-loss-coding-framework.md before you code themes, assign a primary reason, or quantify a finding. Skip it only if the user already coded the set and wants implications from a finished ledger.

Procedure

1. Draw the boundary. Eligible population, inclusion rules, outcome definitions, and what counts as evidence. Done when a reviewer can see who is in and who is out.
2. Build the sample ledger. Eligible, selected, contacted, participated, evidence-available, coded, excluded. By outcome and any material segment. No ledger, no prevalence language.
3. Keep evidence and interpretation apart. Observed fact, stakeholder-reported reason, contributing factor, and analyst root-cause hypothesis are four fields. A CRM close reason is one input, not the finding.
4. Calibrate, then code. One primary reason only when the evidence supports a dominant reason; otherwise mixed or unresolved. Preserve secondary reasons and disagreements between buyer, seller, and system. Version the codebook.
5. Assign confidence from the slice, not the story. Independent deals, direct buyer evidence, coverage, source mix, counterexamples. High never transfers automatically to another segment or competitor.
6. Protect people and accounts. Approved attribution only. Suppress thin cells. No rep-performance verdicts unless that use was approved.
7. Recommend only what the sample can carry. Recommend messaging, coaching, process, or product changes with a responsible owner. Include counterexamples and a way to test the proposed explanation. A mention of price may describe a symptom; compare buyer criteria, perceived value, and alternatives before prescribing discounting.

Do not publish identifiable quotes or write findings into an external system without explicit approval of the artifact, audience, and destination.

What to return

For a targeted question, return the supported finding and its limits. For a full synthesis, include:

• Sample-accounting table: population, selection, participation, evidence, exclusions, denominators
• Findings by outcome, each with supporting deals, counterexamples, source mix, and confidence
• Evidence vs interpretation, labeled on every material claim
• Segment or competitor cuts only where the slice can support them
• Coding-calibration note: codebook version, agreement or recode, reliability limits
• Implications with owner, decision, and dependency
• Limits on population, segment, and outcome claims

Do not produce a battlecard or a governed objection library from this pass.

Quality bar

• Sample shape appears before any percentage or priority list.
• Interview shares are not presented as population prevalence unless the design supports it.
• Reported reasons and analyst hypotheses stay distinct.
• Every material finding has a denominator, a counterexample check, and a confidence label.
• Recommendations are owner-addressable and proportionate to coverage.
• Confidentiality holds at quote, deal, and slice level.