feat: negotiation discourse analysis — affect-text discordance, linguistic evasion, and manipulation pattern detection [INTERNAL] #21

New issue

Open

opened 2026-04-12 10:19:26 -07:00 by pyr0ball · 0 comments

pyr0ball commented

2026-04-12 10:19:26 -07:00

Owner

Summary

Personal-use feature for high-stakes negotiations with corporate partners. Not lie detection (acoustic stress is not deception-specific). Instead: real-time discourse analysis that detects evasion patterns, manipulation signatures, and affect-text discordance in a counterparty's speech.

Scope: BSL 1.1. Local inference only. No server-side logging of analysis output. Consent-first design.

Signal Types

Affect-text discordance

Compare tonal valence (from existing tone classifier) against sentiment of transcribed words
High discordance: saying positive words in a stressed/negative vocal affect → flag

Linguistic hedge density

Rolling window over Whisper transcript
Score density of hedges: "kind of", "sort of", "I think", "probably", "it depends", "that's not really my area", "we would need to check"
Spike in hedge density while topic is under negotiation = evasion signal

Question deflection rate

Small LLM pass (local) on rolling 60s transcript window
Detect: direct question posed by user → counterparty response does not answer, pivots topic, or asks a counter-question
Accumulate deflection count per session

Baseline drift

Establish baseline vocal affect + hedge density in first 2 minutes of session
Flag when later segments deviate significantly from baseline (topic-correlated drift is more meaningful than absolute levels)

Named pattern signatures

Flattery pivot: positive affect spike immediately before or after a concession request
False urgency: temporal pressure language ("by end of week", "we need to decide now", "the window is closing") without substantive justification
Minimization + reframe: "it's really not a big deal" framing applied to items the user raised as concerns
DARVO adjacency: Deny, Attack, Reverse Victim and Offender — detectable via sentiment reversal after user raises concern

Session Mode

Add mode: "negotiation" parameter to session start. Negotiation mode:

Enables discourse analysis pipeline alongside standard tone annotation
Does NOT change standard Elcor annotation for the transcript
Adds NegotiationSignal event type to SSE stream

NegotiationSignal {
  pattern: string        // e.g. "hedge_spike", "flattery_pivot", "deflection"
  confidence: float      // 0.0–1.0
  explanation: string    // human-readable, e.g. "High hedge density while discussing timeline"
  segment_start_ms: int
  segment_end_ms: int
}

UI

Subtle concern indicator strip alongside NowPanel (not alarming, not accusatory)
Framed as: "discourse pattern detected" not "lying detected"
Session summary: pattern frequency table, timeline of signal spikes
All UI copy must be carefully neutral — this is analytical, not judgmental

Negotiation mode requires explicit one-time opt-in with plain-language explanation
Jurisdiction note: one-party recording consent laws vary; show a reminder before enabling
All analysis runs locally; no analysis output is stored on CF servers
Raw signals (tone vectors, transcript) follow standard Linnet retention policy
Analysis results (pattern detections) are session-scoped and discarded on session end

Implementation Notes

Reuse existing _classify_real_async pipeline — NegotiationSignal is a post-processing pass on the existing outputs
Hedge density and question deflection require a lightweight LLM call; use a local model (llama.cpp or ollama) — never cloud for this feature
Baseline establishment: accumulate first N segments, compute baseline stats, then enable drift detection
Pattern signatures: rule-based first (faster, more explainable), ML upgrade later

Out of Scope (this issue)

No export of raw analysis to third parties
No training data collection from user sessions
No cloud inference for negotiation analysis
No public-facing version of this feature without separate design review

Dependencies

#5 YAMNet acoustic events (baseline affect signals)
#8 Speaker diarization (needed to attribute signals to counterparty vs. self)
Existing tone classifier (already working)
Existing Whisper STT (already working)

## Summary Personal-use feature for high-stakes negotiations with corporate partners. **Not lie detection** (acoustic stress is not deception-specific). Instead: real-time discourse analysis that detects evasion patterns, manipulation signatures, and affect-text discordance in a counterparty's speech. Scope: BSL 1.1. Local inference only. No server-side logging of analysis output. Consent-first design. ## Signal Types ### Affect-text discordance - Compare tonal valence (from existing tone classifier) against sentiment of transcribed words - High discordance: saying positive words in a stressed/negative vocal affect → flag ### Linguistic hedge density - Rolling window over Whisper transcript - Score density of hedges: "kind of", "sort of", "I think", "probably", "it depends", "that's not really my area", "we would need to check" - Spike in hedge density while topic is under negotiation = evasion signal ### Question deflection rate - Small LLM pass (local) on rolling 60s transcript window - Detect: direct question posed by user → counterparty response does not answer, pivots topic, or asks a counter-question - Accumulate deflection count per session ### Baseline drift - Establish baseline vocal affect + hedge density in first 2 minutes of session - Flag when later segments deviate significantly from baseline (topic-correlated drift is more meaningful than absolute levels) ### Named pattern signatures - **Flattery pivot**: positive affect spike immediately before or after a concession request - **False urgency**: temporal pressure language ("by end of week", "we need to decide now", "the window is closing") without substantive justification - **Minimization + reframe**: "it's really not a big deal" framing applied to items the user raised as concerns - **DARVO adjacency**: Deny, Attack, Reverse Victim and Offender — detectable via sentiment reversal after user raises concern ## Session Mode Add `mode: "negotiation"` parameter to session start. Negotiation mode: - Enables discourse analysis pipeline alongside standard tone annotation - Does NOT change standard Elcor annotation for the transcript - Adds `NegotiationSignal` event type to SSE stream ``` NegotiationSignal { pattern: string // e.g. "hedge_spike", "flattery_pivot", "deflection" confidence: float // 0.0–1.0 explanation: string // human-readable, e.g. "High hedge density while discussing timeline" segment_start_ms: int segment_end_ms: int } ``` ## UI - Subtle concern indicator strip alongside NowPanel (not alarming, not accusatory) - Framed as: "discourse pattern detected" not "lying detected" - Session summary: pattern frequency table, timeline of signal spikes - All UI copy must be carefully neutral — this is analytical, not judgmental ## Privacy and Consent Design - Negotiation mode requires explicit one-time opt-in with plain-language explanation - Jurisdiction note: one-party recording consent laws vary; show a reminder before enabling - All analysis runs locally; no analysis output is stored on CF servers - Raw signals (tone vectors, transcript) follow standard Linnet retention policy - Analysis results (pattern detections) are session-scoped and discarded on session end ## Implementation Notes - Reuse existing `_classify_real_async` pipeline — NegotiationSignal is a post-processing pass on the existing outputs - Hedge density and question deflection require a lightweight LLM call; use a local model (llama.cpp or ollama) — never cloud for this feature - Baseline establishment: accumulate first N segments, compute baseline stats, then enable drift detection - Pattern signatures: rule-based first (faster, more explainable), ML upgrade later ## Out of Scope (this issue) - No export of raw analysis to third parties - No training data collection from user sessions - No cloud inference for negotiation analysis - No public-facing version of this feature without separate design review ## Dependencies - #5 YAMNet acoustic events (baseline affect signals) - #8 Speaker diarization (needed to attribute signals to counterparty vs. self) - Existing tone classifier (already working) - Existing Whisper STT (already working)