Voc Synthesize

/voc-synthesize — Cross-Source Pattern Detection

One mention is an anecdote. Three mentions across three different channels is a pattern. That distinction is the entire product, and this skill is built so it cannot be quietly relaxed.

Clustering is a judgment task, so subagents do it. Counting is not, so code does it. The confirmation rule lives in scripts/synthesize.js reading config/thresholds.json, and nowhere else. If you find yourself reasoning about whether something "feels confirmed," you have left the rule behind.

Requires

Phase 1 — Gather

Collect everything in the window (default from windows.default_days):

Do not read the quarantine folder. Tier-U signals are excluded by design.

Report which sources were empty. A source that returned nothing because its connector is broken looks identical to a source where nothing happened, and the difference matters — say which you believe it was and why.

Phase 2 — Cluster in parallel

Spawn one subagent per dimension, all in a single message:

pain · request · praise · objection · competitive

Each subagent receives the full corpus and returns clusters for its dimension only:

Group signals in the corpus below into themes for the "<dimension>" dimension.

Cluster on underlying meaning, not wording. "Export is slow", "downloads time out"
and "can't get data out for big events" are ONE theme. Err toward merging: two
themes that should have been one is a worse error than one theme that should have
been two, because splitting hides frequency.

Return ONLY this JSON:
{"clusters":[{
  "theme": "short specific name, written as the customer would recognize it",
  "dimension": "<dimension>",
  "signals": [{"channel":"...","weight_tier":"...","company":"...",
               "date":"YYYY-MM-DD","verbatim":"exact quote","source_file":"..."}]
}]}

Rules:
- Copy verbatim quotes EXACTLY. Do not clean up grammar or trim for length.
- Every signal keeps its original channel and weight_tier. Do not reassign them.
- Do not invent signals to strengthen a cluster.
- A signal may appear in only one cluster within your dimension.
- If a quote is ambiguous between two themes, put it in the more specific one.
CORPUS:
[...]

Merge all five results into one clusters.json.

Phase 3 — Apply the rule in code

node scripts/synthesize.js outputs/voc/synthesis/clusters.json

This writes <date>-confirmed.json, <date>-watchlist.json, <date>-dropped.json and prints the tally. It handles four things you should not re-derive by hand:

• Independent source means distinct channel. Two calls are one source.
• Tier U never counts toward confirmation, but is reported per theme so you can see how much unattributed volume sits behind it.
• One tier-A signal alone never confirms. Seniority is not corroboration.
• A theme where every mention comes from one account gets downgraded, because one loud customer is not a market pattern. This is the check most VoC tooling omits, and it is the one that most often prevents a bad roadmap decision.

Take the script's output as fact. Do not re-rank it because a dropped theme seemed interesting.

Phase 4 — Write the report

outputs/voc/synthesis/<date>-synthesis.md:

# VoC Synthesis — week of <range>

## Coverage
| Channel | Signals | Status |
|---|---|---|
| Calls | N | ok |
| Reviews | N | ok |
| Chat | 0 | connector not configured |
...
**Quarantined (unattributed):** N signals excluded from all counts.

## Confirmed themes
### <Theme name>
**High confidence** · N mentions · N channels (call, review, chat) · N accounts · tiers A:1 B:4
**First seen** YYYY-MM-DD · **Most recent** YYYY-MM-DD

> "verbatim quote"
> — Company, YYYY-MM-DD (call, tier A)

<3-5 quotes>

**What this means:** <one sentence, concrete, naming who should do what>

## Medium confidence
<same shape>

## Watchlist
<single-channel themes, one line each, with the channel named>

## What did not confirm
<count, plus two or three notable near-misses and what they were short of>

Lead with the confirmed themes. Put coverage above them so nobody reads a thin week as a complete picture.

What this gets wrong

• The rule is a heuristic, not a truth test. Three channels can all reflect one loud market segment. Check the account spread the script reports before acting.
• Merging hides nuance. Aggressive clustering is right for finding frequency and wrong for finding the specific edge case that is actually blocking deals. Read the raw quotes.
• Absence proves nothing. A theme that did not confirm may simply be one nobody asked about. Your discovery script shapes this corpus more than you think.
• Recency bias is built in. A weekly window favours whatever happened this week. Run a quarterly window occasionally to see what has been steadily true.
• The rule can be gamed by adding connectors. Wiring up more channels makes confirmation easier, not truer. If you add sources, consider raising min_independent_sources.