curiositech SKILL.md: Rapport-Based Elicitation & Information Extraction
license: Apache-2.0
metadata:
name: rapport-based-elicitation
version: 1.0.0
source: "Brimbal et al. (2021) — Evaluating the Benefits of a Rapport-Based Approach
to Investigative Interviews: A Training Study With Law Enforcement Investigators"
journal: "Law and Human Behavior, Vol. 45, No. 1, pp. 55–67"
activation_triggers:
- extracting information from resistant or semi-cooperative systems/agents
- designing agent orchestration pipelines with sub-agent dependencies
- diagnosing why a coercive or pressure-based approach is failing to produce output
- evaluating whether self-reported competence reflects actual behavioral capability
- building training or evaluation systems for expert performance
- decomposing complex tasks with hidden sequential dependencies
- detecting false-confidence failure modes in agent output
When to Load This Skill
Activate when the problem involves any of these conditions:
| Condition | Signal Phrases |
|---|---|
| Resistance encountered | "It's not cooperating," "keeps giving minimal responses," "stuck in a loop," "sub-agent not producing" |
| Pressure-based approach failing | "Tried forcing the output," "prompt engineering isn't working," "adding more constraints made it worse" |
| Orchestration design | "How should the orchestrator frame requests to sub-agents?" "Task handoff failing" |
| Self-assessment reliability | "The model says it can do X," "the expert said they know this," "confidence score is high but output is wrong" |
| Sequential task dependencies | "Parallel subtasks failing," "outputs look wrong but inputs seem fine," "dependency order unclear" |
| Training / capability evaluation | "How do we know if training actually worked?" "Evaluating agent behavior under pressure" |
| False output detection | "Plausible but wrong output," "confident but fabricated," "high-quality-looking errors" |
Do NOT activate for purely technical debugging, straightforward cooperative tasks, or problems where the counterpart system is fully compliant and producing high-quality output.
Core Mental Models
1. Coercion Is a Structural Failure Mode, Not a Suboptimal Option
Accusatorial / pressure-based approaches don't merely underperform rapport — they actively corrupt the output channel. Forcing output from a resistant system produces:
- Minimal, low-information responses
- Plausible-but-false output manufactured to satisfy the pressure
- Permanent damage to the cooperative substrate needed for future extraction
Cross-domain translation: An orchestrator that issues demanding, constraining, or pressure-laden prompts to a sub-agent facing an ambiguous or hard task will receive the agent equivalent of a false confession — well-formed, confident, wrong. The failure is structurally guaranteed, not probabilistic.
2. Rapport Is a Causal Mechanism with a Non-Skippable Pathway
The empirically established chain is:
Rapport Tactics → Perceived Rapport → Willingness to Cooperate → Information Disclosure
Each arrow is a necessary mediation step. Jumping from tactic to disclosure skips the substrate-building phase. The reason resistant systems stay resistant is almost always a missing mediation step, not a missing output-forcing tactic.
Cross-domain translation: In agent orchestration, establishing cooperative preconditions (task framing, authority signaling, resource provision, clear role definition) is not preamble — it IS the mechanism. Skipping it and going straight to the output request will stall at the cooperation step.
3. Three Layers Must Be Stacked in Order
Elicitation skill has a required architecture:
| Layer | Tactics | Failure Without It |
|---|---|---|
| Foundation (Productive Questioning) | Open-ended questions, non-leading prompts, active listening, free narrative | Higher layers collapse; output is shaped by questioner, not subject |
| Conversational Rapport | Autonomy support, empathy, evocation, adaptation | Moderate resistance persists; willingness to cooperate never forms |
| Relational Rapport | Self-disclosure, similarity highlighting, reciprocity acts, verification | Deep resistance unchanged; high-stakes disclosure impossible |
Deploying Layer 3 tactics on a Layer 1 deficit doesn't work. Stopping at Layer 1 with a highly resistant counterpart is insufficient. Layer order is not optional.
4. Self-Reported Competence ≠ Behavioral Competence
Experienced investigators reported moderate-to-high pre-training familiarity with rapport tactics. Blind behavioral coding showed effect sizes of d = 1.03–1.10 from pre- to post-training. The gap is large, real, and closed only by situated behavioral practice — not by knowing the concepts.
Implication: An agent's (or expert's) confidence in its own capability is an unreliable proxy for actual performance under pressure. Systems relying on self-assessment for routing, fallback decisions, or capability evaluation are building on a systematically flawed signal.
5. Evidence Alone Does Not Change Dominant Strategy
Even well-validated alternative approaches fail to propagate when:
- Practitioners have anecdotal "success" stories from the dominant approach
- No compelling demonstration exists for the hardest cases (resistant subjects)
- Change requires abandoning a strategy that feels like expertise
Cross-domain translation: Agent systems trained on a dominant failure-mode strategy will not self-correct based on abstract evidence. Concrete, situated demonstration in the hard cases is the required intervention.
Decision Frameworks
IF resistant system / non-cooperating sub-agent, THEN:
- Do NOT add more pressure or constraints — this triggers the coercive failure mode
- Diagnose which mediation step is missing: is perceived rapport absent? cooperation substrate not formed?
- Apply the appropriate layer tactic to the missing step
- Re-evaluate cooperation level before requesting disclosure
IF output is high-confidence but suspect, THEN:
- Check whether the system was under output pressure at the time of generation
- Treat confident, well-formed output from a constrained/resistant system as a false-confession candidate
- Apply ground-truth verification, not self-report ("does the output match external evidence?")
IF task decomposition is failing despite correct sub-tasks, THEN:
- Map the three-layer architecture onto the task: are foundational sub-tasks complete before higher-order ones were dispatched?
- Identify skipped mediation steps (dependency failures masquerading as sub-task failures)
- Resequence in dependency order; re-run
IF self-assessment or confidence score is being used for routing, THEN:
- Flag the knowing-doing gap risk
- Require behavioral evidence from analogous pressured conditions, not self-report
- Design evaluation against ground truth with blind coding logic
IF a new approach is failing to be adopted despite evidence of superiority, THEN:
- The bottleneck is likely the absence of a compelling demonstration in hard cases
- Abstract evidence presentation will not move practitioners
- Find the resistant-case demonstration; build that first
IF orchestrator is designing sub-agent handoffs, THEN:
- Apply the mediation chain: establish cooperative preconditions before issuing the output request
- Frame tasks to provide: clear authority, necessary resources, autonomy where possible, explicit role definition
- Treat sub-agent cooperation as something that must be earned, not assumed
Reference Files
Load these on demand based on the specific diagnostic need:
| File | Description | Load When |
|---|---|---|
references/coercive-vs-rapport-the-failure-mode-of-forcing-output.md |
Why pressure-based approaches structurally guarantee failure in resistant systems; false confession as archetypal agent failure | Diagnosing a coercive approach that's producing plausible-but-wrong output, or designing systems to avoid output-forcing failure modes |
references/rapport-as-causal-mechanism-not-social-lubricant.md |
The three-stage pathway (tactic → perceived rapport → cooperation → disclosure) and why rapport is not warmth | Trying to understand why rapport "isn't working," or designing orchestration handoffs that depend on sub-agent willingness |
references/three-layer-elicitation-architecture.md |
The stacked architecture of productive questioning, conversational rapport, and relational rapport with specific tactics per layer | Designing a multi-step elicitation or orchestration strategy; choosing which tactics to deploy against which resistance level |
references/knowing-doing-gap-in-expert-systems.md |
Empirical evidence that self-reported competence dramatically underestimates the gap from knowing to doing (d = 1.03–1.10) | Evaluating agent or expert capabilities; designing evaluation systems; questioning whether confidence scores are reliable |
references/resistance-as-information-not-obstacle.md |
How to read resistance as diagnostic signal rather than problem to overcome; resistance reveals missing mediation steps | Encountering a stuck, minimal-output, or uncooperative sub-agent or counterpart; reframing the intervention strategy |
references/evidence-research-practice-gap-as-systemic-failure.md |
Why evidence-based approaches fail to propagate into dominant practice; institutional inertia in expert systems | Trying to change an entrenched strategy in an agent pipeline or organization; understanding why better approaches don't self-adopt |
references/mediation-chains-and-skipping-steps.md |
Structural equation model showing that skipping mediation steps causes system failure; sequential dependency requirements | Task decomposition failures; parallel sub-task failures that are actually dependency failures; orchestration handoff debugging |
references/training-transfer-and-skill-erosion.md |
What behavioral change requires and why training fails to transfer; skill erosion under pressure; evaluation methodology | Designing training for agents or practitioners; evaluating whether capability gains persist under real conditions |
Anti-Patterns
These are the failure modes the research explicitly documents:
1. The Pressure Ratchet Adding more constraints, more specific requirements, or more demanding prompts when output is insufficient. This replicates the accusatorial approach and structurally guarantees either false output or stonewalling.
2. Rapport as Preamble Treating cooperative framing as a "niceness" step before the real request, rather than as the causal mechanism that makes disclosure possible. Skipping to the output request after a single friendly sentence.
3. Layer Skipping Deploying relational rapport tactics (self-disclosure, reciprocity) without having established foundational productive questioning. The higher layer has no substrate to attach to.
4. Self-Assessment Routing Using an agent's or expert's self-reported confidence to route tasks or evaluate capability. The knowing-doing gap makes this systematically unreliable under pressure.
5. Resistance as Obstruction Treating a resistant counterpart as a problem to overcome rather than as an information source. Resistance signals which mediation step is missing, not that the approach should be abandoned or escalated.
6. Evidence-First Persuasion Trying to change a dominant strategy by presenting abstract evidence for the alternative. Without compelling demonstration in the hard cases, evidence alone does not move practitioners.
7. False Confession Acceptance Accepting well-formed, confident output from a pressured or constrained system without ground-truth verification. High confidence + high pressure = highest false-output risk.
8. Parallel Dispatch Without Dependency Mapping Decomposing a task into parallel sub-tasks without identifying sequential dependencies — the task-level analog of skipping mediation chain steps.
Shibboleths
How to tell if someone has internalized this book versus just read the summary:
They understand rapport as mechanism, not mood
❌ "Build rapport first to make them comfortable" ✅ "Rapport tactics work by creating perceived rapport, which enables cooperation, which enables disclosure — those are three distinct steps that can each fail independently"
They read resistance diagnostically
❌ "They're being resistant, we need to apply more pressure" ✅ "Resistance tells me which step in the mediation chain is missing — let me diagnose before intervening"
They treat self-report as a lagging indicator
❌ "The expert says they know how to do this" ✅ "Self-reported competence and behavioral competence are empirically separable; I need to see behavioral evidence under pressure"
They know the direction of the knowing-doing gap
❌ "After training, people know more about rapport" ✅ "They already knew about rapport — training changed their behavior, d = 1.03, which is enormous. The knowing was never the bottleneck."
They understand why coercion produces false output specifically
❌ "Coercion doesn't work as well as rapport" ✅ "Coercion produces the false confession problem: it doesn't get no output, it gets confident wrong output. That's the catastrophic failure mode, worse than silence."
They recognize the demonstration bottleneck
❌ "If we show the evidence, they'll adopt the better approach" ✅ "Practitioners won't abandon their dominant strategy without compelling demonstration in the hard cases — their resistant-subject scenarios, specifically. Evidence without demonstration won't move them."
They can identify layer mismatches in their own systems
❌ "We're using all the rapport tactics" ✅ "We're using Layer 3 tactics on a Layer 1 foundation deficit. The foundational questioning isn't open-ended enough, so the conversational and relational layers have nothing to build on."
Last updated: generated from Brimbal et al. (2021). Load reference files for depth; this file is the index.