field-feedback-synthesizer

Field feedback synthesizer

Turn a defined set of field feedback into traceable findings and actions. Weigh independent sources and contradictory evidence before recommending a response.

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.

Inputs

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 or question is unresolved, return an analysis plan. If only the denominator is missing, qualitative findings may still be useful; label coverage limits and avoid prevalence claims. Do not invent usage, win rates, or quotes.

Supporting material

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.

Procedure

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.