feat: negotiation discourse analysis — affect-text discordance, hedge density, pattern signatures [INTERNAL] #4

New issue

Open

opened 2026-04-12 10:56:21 -07:00 by pyr0ball · 0 comments

pyr0ball commented

2026-04-12 10:56:21 -07:00

Owner

Context

Linnet#21 (INTERNAL). cf-voice side of the negotiation discourse analysis feature.

cf-voice responsibilities

Baseline establishment: accumulate first N segments (configurable), compute baseline vocal affect + hedge density stats
Baseline drift detection: flag windows where affect or hedge density deviates significantly from session baseline
Expose NegotiationSignal output from classify_chunk_async() when mode="negotiation" is set
Rule-based pattern signatures (no cloud LLM — local inference only):
- Flattery pivot: positive affect spike immediately before/after a concession request in transcript
- False urgency: temporal pressure language scored by hedge density delta
- Minimization+reframe: sentiment reversal when user-concern keywords appear in rolling window

Linnet side

Question deflection rate (requires conversational context) is handled on the Linnet side via a small local LLM pass on the rolling transcript — not cf-voice.

Privacy constraints

Local inference only
No export of raw analysis to any server
No training data collection from sessions
Consent-first: mode="negotiation" must be explicitly set per session

Tracking

Linnet#21

## Context Linnet#21 (INTERNAL). cf-voice side of the negotiation discourse analysis feature. ## cf-voice responsibilities - Baseline establishment: accumulate first N segments (configurable), compute baseline vocal affect + hedge density stats - Baseline drift detection: flag windows where affect or hedge density deviates significantly from session baseline - Expose `NegotiationSignal` output from `classify_chunk_async()` when `mode="negotiation"` is set - Rule-based pattern signatures (no cloud LLM — local inference only): - Flattery pivot: positive affect spike immediately before/after a concession request in transcript - False urgency: temporal pressure language scored by hedge density delta - Minimization+reframe: sentiment reversal when user-concern keywords appear in rolling window ## Linnet side Question deflection rate (requires conversational context) is handled on the Linnet side via a small local LLM pass on the rolling transcript — not cf-voice. ## Privacy constraints - Local inference only - No export of raw analysis to any server - No training data collection from sessions - Consent-first: `mode="negotiation"` must be explicitly set per session ## Tracking Linnet#21