Field feedback synthesizer

Field feedback synthesizer

Turn a defined feedback corpus into findings a decision owner can use. Score source breadth before you recommend. Observation is what the sources said. Recommendation is what someone should do next. Keep them apart.

Fails. "The field is telling us the security FAQ is broken. Rewrite it this week."

Passes. "17 mentions of security-review friction across 9 independent accounts and 3 source types. Observation: sellers cannot find or cannot use the current FAQ. Recommendation: confirm discoverability before rewriting. Confidence: medium — no usage denominator, and 8 of 17 mentions came from one Slack channel."

Asset-demand scoring is content-gap-analysis. Recurring call objections are call-insights-to-objections. Why deals closed is win-loss-synthesis. One recorded call is call-review-coach.

Confirm Inputs First

Ask only for the inputs that change the synthesis:

1. The decision or research question
2. Included sources, timeframe, teams, segments, and exclusions
3. Counting unit and denominator: records, respondents, accounts, opportunities, or calls
4. Confidentiality and quotation boundaries

If the corpus, counting unit, or question is unresolved, return an analysis plan rather than quantified findings. Do not invent usage, win rates, or quotes.

Read The Right Reference

Read references/feedback-coding-framework.md before you code the corpus, score breadth, or write a recommendation. Skip it if the user already coded the set and wants implications from a finished register.

Default Workflow

1. Frame the sample. Question, inclusion rules, collection method, counting unit, denominator, and known bias. Done when a reviewer can see what this corpus can and cannot represent.
2. Register sources. Stable record and origin IDs. Collapse copies of the same originating statement. Keep independent observations separate.
3. Calibrate the codebook. Pilot on a mixed subset. Resolve overlapping codes. Freeze before the full pass, or record every later change.
4. Code without collapsing layers. Theme, segment, stage, evidence type, and severity. Frequency, severity, evidence strength, and business impact stay separate dimensions.
5. Quantify with the right denominator. Show n/N, base sizes, and independent-source breadth. A convenience sample is not population prevalence.
6. Score breadth, then confidence. Independent units, source types, segments, and time periods. High frequency from one biased channel is not high confidence.
7. Look for what a ranking would hide. Minority, role-specific, regional, and contradictory signals. Report a cut only when the base supports it.
8. Interpret, then recommend. Observation, interpretation, root-cause hypothesis, and decision are four lines. Each material recommendation needs an owner, a validation step, and what would disconfirm it.

Do not distribute verbatim feedback or write findings into product, CRM, or planning systems without explicit approval of content, audience, and destination.

Output Contract

Return a synthesis the decision owner can inspect:

• Scope and coverage: counting unit, denominator, source mix, exclusions, limits
• Decision summary that separates findings from hypotheses
• Theme table with n/N, independent-source breadth, severity, confidence, and evidence references
• Observation vs recommendation, labeled on every material theme
• Representative, minimally necessary excerpts
• Minority, contradictory, and missing signals
• Actions with owner, validation step, and decision point
• What the source mix cannot support

Do not produce a content backlog or an objection library from this pass.

Quality Bar

• Every material finding traces to included records without extra sensitive detail.
• Duplicate copies are not counted as independent sources.
• Breadth is scored; one loud channel cannot set the priority list.
• Observation and recommendation are not the same sentence.
• Root cause stays a hypothesis until evidence establishes it.
• No invented usage, win rates, or quotes.