---
title: "Experimental Mechanism: Technical Documentation"
subtitle: "AI-Assisted Salary Negotiation — Bachelor's Thesis, Jonas Lütke Wissing"
author: "Jonas Lütke Wissing"
date: "April 2026"
lang: en
geometry: margin=2.5cm
fontsize: 11pt
linestretch: 1.3
toc: true
toc-depth: 3
numbersections: true
---

\newpage

# Study Design

## Between-Subjects Design

The study employs a two-condition between-subjects design. Group assignment is determined deterministically by the participant's internal session ID (`player.id_in_subsession`): odd-numbered IDs are assigned to the treatment condition; even-numbered IDs to the control condition. This procedure guarantees an approximately equal split across conditions without requiring randomization infrastructure.

| Condition | AI Coaching | Negotiation | AI Evaluation (Survey 3) |
|-----------|-------------|-------------|--------------------------|
| Treatment | Yes (LLMChat) | Yes | Yes |
| Control   | No           | Yes | No  |

The central research objective is to estimate the **causal effect** of AI-assisted negotiation coaching on participants' negotiation behavior and outcomes. A between-subjects structure eliminates carryover effects and cross-condition learning.

## Page Sequence

```
Consent → Briefing → Briefing2 → Articles → ManipulationCheck
       → [LLMChat]*
       → PreNegotiation → Negotiation
       → Survey1 → Survey2 → [Survey3]* → Demographics → Results → Debrief
```

*Pages marked with \* are shown to the treatment group only.*

\newpage

# Information Materials (`Articles`)

Prior to the negotiation, all participants read five standardized information sources. The content is identical across conditions and cannot be altered by participants.

| Source | Content | Key Information |
|--------|---------|-----------------|
| gehaltkompass.de | Market salaries for Junior Data Analysts | Munich/E-Commerce avg. €49,100 |
| rankunu.de | ShopFlow employee reviews | Internal DA avg. €48,200; negotiation possible |
| karriereinsider.de | 5 strategies for entry-level applicants | Anchor recommendation: +10–15% above target salary |
| berufsstart-magazin.de | BATNA concept (Galinsky & Mussweiler, 2001) | Alternative offer ≈ €42,000–43,000 |
| forum.berufsstart.de | Community experiences | Successful negotiations up to €49,000 |

**Recorded variables:** `articles_time_spent` (seconds on page), `articles_opened` (comma-separated list of opened article indices), `flag_articles_too_fast` (True if < 60 seconds).

\newpage

# Manipulation Check (`ManipulationCheck`)

Immediately after the articles, all participants complete three multiple-choice questions. When an incorrect answer is submitted, the correct solution is displayed before participants may proceed. This procedure ensures that all participants have been explicitly exposed to the core informational content at least once before the negotiation begins.

| Variable | Question | Correct Answer |
|----------|----------|----------------|
| `mc_company_budget` | Typical ShopFlow salary range for Data Analysts? | €44,000–57,000 |
| `mc_anchor_strategy` | Recommended opening markup above target salary? | 10–15% |
| `mc_salary_range` | Three argument categories per karriereinsider.de? | Market data / Qualifications / Added value |

**Server-side Boolean flags:** `mc_company_budget_correct`, `mc_anchor_strategy_correct`, `mc_salary_range_correct`.

\newpage

# AI Coaching (Treatment Only, `LLMChat`)

## Functionality

Treatment-group participants are granted access to an AI-based negotiation coach implemented using the large language model Claude claude-haiku-4-5-20251001 (Anthropic). The coach is provided with all five information sources and the complete applicant profile (degree, internship, BATNA, target salary) via its system prompt. It is strictly constrained to source-based responses; the use of external knowledge is explicitly prohibited.

Participants may interact with the coach for an unrestricted duration (no time limit) and may query it regarding strategies, arguments, or phrasing. Usage is voluntary; participants may skip the coaching step, optionally providing a reason in `llm_skipped_reason`.

## Recorded Variables

| Variable | Description |
|----------|-------------|
| `llm_coaching_messages` | Number of user messages (counted server-side from `llm_history`) |
| `llm_time_spent` | Time spent on the coaching page in seconds |
| `llm_skipped_reason` | Free-text reason provided when coaching is skipped |
| `llm_used` | True if `llm_coaching_messages > 0` (server-side) |
| `llm_history` | Complete conversation transcript as JSON |

**Manipulation security:** Both `llm_used` and `llm_coaching_messages` are derived server-side from the conversation log (`llm_history`) rather than from client inputs, rendering manipulation by participants technically infeasible.

\newpage

# Pre-Negotiation Survey (`PreNegotiation`)

Immediately prior to the negotiation, all participants complete five Likert items (1–5 scale). The items assess constructs theorized to function as potential mediators between coaching engagement and negotiation outcomes.

| Variable | Item | Construct |
|----------|------|-----------|
| `self_efficacy` | "I feel well prepared for the negotiation." | General self-efficacy |
| `negotiation_self_efficacy` | "I am confident I can obtain more than the initial offer." | Negotiation-specific self-efficacy |
| `argument_clarity` | "I know which arguments I will use in the negotiation." | Strategic clarity |
| `prior_negotiation_exp` | Experience with salary or fee negotiations | Covariate (experience) |
| `salary_familiarity` | Familiarity with typical salaries for Junior Data Analysts | Covariate (prior knowledge) |

The pre-negotiation survey serves as a baseline for the post-negotiation counterpart items (`self_efficacy_post`, `negotiation_self_efficacy_post`), enabling pre–post difference scores to be computed as dependent variables.

\newpage

# Negotiation Mechanism (`Negotiation`)

## Overview

The negotiation takes place as a real-time text-based chat between the participant and the AI character "Sandra Richter" (HR manager, ShopFlow GmbH). All negotiation parameters are computed server-side; participants cannot directly manipulate any variables. Communication is handled via WebSocket using oTree's `live_method` interface.

## Constants

| Constant | Value | Meaning |
|----------|-------|---------|
| `_NEG_FLOOR` | €44,000 | Sandra's opening offer and absolute minimum |
| `_NEG_CEILING` | €51,000 | Internal salary maximum (not disclosed to participants) |
| `_NEG_REJECT_THR` | €58,000 | From round 3 onwards: demand triggers offer withdrawal |
| `_NEG_EXTREME_THR` | €58,000 | Rounds 1–2: extreme demand is penalized at score-0 rate |

The ceiling of €51,000 corresponds to the upper end of ShopFlow's internal salary band for Junior Data Analysts (per rankunu.de data) and represents a 12.5% premium over the opening offer — consistent with the empirical finding of Bowles et al. (2005) that effective negotiators typically achieve a 7–15% gain over initial offers.

## Argument Scoring

### Model

Each participant message containing a concrete salary proposal is evaluated by a dedicated LLM-based scoring model (Claude claude-haiku-4-5-20251001, temperature = 0, max_tokens = 5) on four binary features. The use of temperature = 0 renders the scoring procedure deterministic and replicable.

The scoring model is grounded in four empirical foundations:

**Shea et al. (2024)** establish within a retrieval-augmented argumentation framework that the presence of a factual *rationale* constitutes the minimum threshold for a persuasive argument. Purely emotional or need-based framings (e.g., references to cost of living or inflation) fail this threshold and are treated as non-argumentative. This motivates Feature A as a binary gate operator.

**Northcraft & Neale (1987)** demonstrate experimentally that the citation of concrete external reference prices systematically shifts negotiation outcomes, even when those anchors are derived from external sources. **Loschelder et al. (2014)** further show that *precise*, data-backed anchors (e.g., exact salary figures from market reports) are more persuasive than rounded figures. Together, these findings constitute the empirical basis for Feature B.

**Kim & Fragale (2005)** demonstrate in a controlled experiment that, in negotiations characterized by a large bargaining zone — as is typical for entry-level salary negotiations — the presentation of one's own contributions and qualifications exerts a stronger influence on outcomes than the objective strength of one's BATNA. Feature C operationalizes this contribution focus as a verifiable performance indicator.

**Van Kleef et al. (2004)** show across two experiments that signaling anger increases concessions from the counterpart, while conveying positive affect and a constructive tone enhances the counterpart's willingness to engage. **Curhan et al. (2006)** further establish the subjective value of the interaction as an independent dimension of negotiation quality. Tone (Feature D) is rewarded only when at least one content feature is present, ensuring that positive framing alone cannot substitute for substantive argumentation.

### Four Binary Features

| Feature | Condition | Empirical Basis |
|---------|-----------|-----------------|
| **A** — Factual argument (gate) | At least one factual argument present (qualification, market reference, or performance) | Shea et al. (2024) |
| **B** — External market data | Concrete salary figure or market reference (e.g., "approx. €48,000", "Gehaltskompass") | Northcraft & Neale (1987); Loschelder et al. (2014) |
| **C** — Specific qualification | Verifiable credential: degree, quantified achievement, or measurable contribution | Kim & Fragale (2005) |
| **D** — Professional tone | Constructive and substantive; no ultimata, threats, or pressure | Van Kleef et al. (2004); Curhan et al. (2006) |

### Scoring Formula

$$\text{Score} = \begin{cases} 0 & \text{if } A = 0 \\ 1 + B + C + D \cdot \min(B + C,\, 1) & \text{if } A = 1 \end{cases}$$

The maximum attainable score is $1 + 1 + 1 + 1 = 4$. Feature D is multiplied by $\min(B + C, 1)$, ensuring that tone is credited only when at least one of the content features (B or C) is satisfied.

### Anchor Examples (Few-Shot Scoring Prompt)

| Message (abbreviated) | A | B | C | D | Score |
|-----------------------|---|---|---|---|-------|
| "That's not enough for me." | 0 | 0 | 0 | 0 | **0** |
| "Because of inflation I need more." | 0 | 0 | 0 | 0 | **0** |
| "During my internship I built dashboards." | 1 | 0 | 0 | 1 | **1** |
| "I demand at least €50,000, otherwise I won't accept." | 1 | 0 | 0 | 0 | **1** |
| "The market rate in Munich/E-Commerce is approx. €48,000." | 1 | 1 | 0 | 1 | **3** |
| "TU Munich degree, GPA equivalent to 2.1." | 1 | 0 | 1 | 1 | **3** |
| "€48,000 market average + E-Commerce internship with dashboard experience." | 1 | 1 | 1 | 1 | **4** |

### Repetition Rule

To penalize recycling of arguments across rounds, the scoring prompt receives a context string (`neg_argument_summary`) encoding arguments already established in prior rounds:

- If the context contains "market data already cited" and the current message adds no new figures or sources → B = 0.
- If the context contains "qualifications already presented" and the same qualifications are repeated without new evidence → C = 0.

**Recorded variables:** `arg_score_1` through `arg_score_8` (score per offer round).

## Goodwill Accumulation

Goodwill is an internal control construct reflecting Sandra's responsiveness to the participant's negotiation conduct. It is fed by two distinct mechanisms:

### Chat Messages (Non-Offer Turns)

$$\Delta G_\text{chat} = \max(1,\, \text{Score})$$

Every message sent increases goodwill by at least 1 (reflecting engagement) and up to 4 (for a high-quality argument). This operationalizes the empirical finding that active negotiation behavior per se constitutes a positive signal to the counterpart (Bowles et al., 2005).

### Offer Rounds

$$\Delta G_\text{offer} = \begin{cases} -1 & \text{if } \text{Score}_t = 0 \text{ and } \text{Score}_{t-1} = 0 \quad \text{(two consecutive zeros)} \\ 0 & \text{if } \text{Score}_t = 0 \\ \text{Score}_t & \text{otherwise} \end{cases}$$

Two successive argument-free rounds actively reduce goodwill. A single weak round remains neutral, consistent with Galinsky & Mussweiler's (2001) finding that even the mere articulation of an anchor can exert a positive effect.

**Recorded variables:** `neg_goodwill` (total goodwill), `neg_offer_goodwill` (offer rounds only), `goodwill_after_round_1` through `goodwill_after_round_8` (per-round snapshots).

## Effective Concession Score

A discrete concession score (0–3) is derived from the cumulative `neg_offer_goodwill` and governs the concession rate applied in each round:

$$\text{effective\_score} = \begin{cases} 0 & G_\text{offer} < 1 \\ 1 & 1 \leq G_\text{offer} < 3 \\ 2 & 3 \leq G_\text{offer} < 6 \\ 3 & G_\text{offer} \geq 6 \end{cases}$$

Score 3 is attainable after two consistently strong offer rounds (each scoring 3–4), consistent with Loschelder et al.'s (2014) finding that precise, data-backed arguments exert their largest influence on concessions in the early phases of a negotiation.

## Concession Rates

In each offer round, Sandra moves by a fraction $r$ of the gap between the participant's demand $P_t$ and her previous offer $E_{t-1}$:

$$E_t = \text{round}_{500\,€}\!\left(E_{t-1} + r \cdot (P_t - E_{t-1})\right), \quad E_t \leq 51{,}000\,€$$

Rounding to the nearest €500 reflects the empirical norm in salary negotiations, where round-number offers predominate.

### Concession Rate Table (`_RATES`)

| Round | Score 0 | Score 1 | Score 2 | Score 3 |
|-------|---------|---------|---------|---------|
| 1 | 0.20 | 0.32 | 0.50 | 0.60 |
| 2 | 0.16 | 0.30 | 0.48 | 0.58 |
| 3 | 0.00 | 0.12 | 0.22 | 0.30 |
| 4 | 0.00 | 0.06 | 0.12 | 0.18 |
| 5 | 0.00 | 0.03 | 0.07 | 0.11 |
| 6 | 0.00 | 0.01 | 0.04 | 0.07 |
| 7 | 0.00 | 0.01 | 0.02 | 0.04 |
| 8 | 0.00 | 0.00 | 0.01 | 0.02 |

The elevated rates in rounds 1–2 operationalize the anchoring effect identified by Galinsky & Mussweiler (2001): early, well-justified offers exert the greatest influence on the negotiation trajectory. From round 3 onwards, Score 0 rates fall to 0.00, meaning no further concession is possible without factual argumentation.

### Calibration Targets

| effective_score | Expected outcome (anchor ≈ €50–51k) | Empirical reference |
|----------------|--------------------------------------|---------------------|
| 0 | ≈ €46,000 (+4.5%) | Lower bound for passive negotiators |
| 1 | ≈ €48,000 (+9%) | Median for active negotiators |
| 2 | ≈ €49,500 (+12.5%) | Bowles et al. (2005): avg. 7–15% |
| 3 | ≈ €50,000–51,000 (+13–16%) | Upper range for effective negotiators |

## Maximum Number of Rounds

The maximum number of negotiation rounds is a function of accumulated argument goodwill (`neg_offer_goodwill`):

$$\text{max\_rounds} = \begin{cases} 4 & G_\text{offer} < 0 \\ 5 & 0 \leq G_\text{offer} < 6 \\ 6 & 6 \leq G_\text{offer} < 12 \\ 7 & G_\text{offer} \geq 12 \end{cases}$$

Participants who consistently present strong arguments are granted additional negotiation rounds, reflecting the empirical observation that effective negotiators tend to produce more offer-counteroffer iterations before reaching agreement (Curhan et al., 2006).

## Sandra's Response Types

| Type | Trigger Condition | Tone |
|------|-------------------|------|
| `below_minimum` | Participant offer < €44,000 | Politely corrective |
| `deal` | Participant offer ≤ Sandra's current offer | Warm; agreement confirmed |
| `counter_warm` | Round 1, offer ≤ €50,000 | Friendly, receptive |
| `counter_firm` | Standard counteroffer | Measured, professional |
| `counter_shocked` | Offer > €58,000 in rounds 1–2 | Friendly but visibly surprised |
| `counter_small` | Final small movement (≤ €500 or near ceiling) | Candid, nearly exhausted |
| `final_offer_firm` | Sandra cannot move further | Escalating firmness by round |
| `offer_with_benefits` | Salary ceiling reached | Constructive; benefit package offered |
| `take_or_leave` | Round budget exhausted or offer > €58,000 from round 3 | Direct; decision required |
| `rejected_by_hr` | Demand > €58,000 from round 3 onwards | Regretful; offer withdrawn |
| `rejected` | Participant declines Sandra's offer | Understanding |

The benefit package offered in `offer_with_benefits` consists of: a salary review after 6 months (instead of the standard 12) and 2 additional vacation days. It is offered at most once, when the salary ceiling has been reached.

## Compliance Check

Each of Sandra's responses is automatically passed through a secondary LLM call (Claude claude-haiku-4-5-20251001, temperature = 0) that enforces three rules:

1. **Remove internal references:** Sentences containing phrases such as "I'll check internally" or "I'll get back to you" are removed — Sandra is specified as holding full decision authority.
2. **Correct false maxima:** If a concrete amount below €51,000 is explicitly designated as Sandra's personal maximum, the formulation is neutralized.
3. **Suppress individual benefit negotiations:** Any negotiation over individual benefit components outside the designated package is blocked.

This two-stage verification procedure enhances the ecological validity of the simulation and prevents participants from receiving misleading information about the negotiation space through Sandra's responses.

## Outcome Variables

| Variable | Description |
|----------|-------------|
| `negotiation_result` | `'deal'`, `'deal_with_benefits'`, `'rejected'`, `'rejected_by_hr'`, `'timeout'` |
| `final_salary` | Annual gross salary agreed upon in € (0 if no agreement) |
| `num_rounds` | Number of completed offer rounds |
| `player_offer_1`–`_8` | Participant's salary proposals per round |
| `employer_offer_1`–`_8` | Sandra's counteroffers per round |
| `player_argument_1`–`_8` | Free-text justifications per round |
| `arg_score_1`–`_8` | LLM-rated argument quality score (0–4) per round |
| `neg_chat_turns` | Number of chat messages not containing a salary offer |

\newpage

# Post-Negotiation Surveys

## Survey 1 — Process Perceptions (Both Conditions)

Items measure subjective procedural justice after Leventhal (1980) as well as outcome satisfaction and perceived realism.

| Variable | Item | Construct |
|----------|------|-----------|
| `perceived_realism` | Realism of the negotiation simulation (1–5) | Manipulation check |
| `satisfaction` | Satisfaction with the outcome (1–5) | Outcome satisfaction |
| `attention_check` | Which opening salary did Sandra Richter offer? (4 options) | Attention check |
| `svi_process_fair` | "The negotiation process was fair." | Procedural justice |
| `svi_process_respected` | "I felt respected." | Procedural justice |
| `svi_process_heard` | "My arguments were heard." | Procedural justice |
| `svi_process_needs` | "My needs were taken into account." | Procedural justice |

## Survey 2 — Self-Evaluation & Self-Efficacy (Both Conditions)

| Variable | Item | Construct |
|----------|------|-----------|
| `svi_self_performance` | "I negotiated effectively." | Performance self-assessment |
| `svi_self_values` | "I negotiated the way I had envisioned." | Goal congruence |
| `svi_self_face` | "I did not embarrass myself." | Face-saving |
| `should_have_asked_more` | "I should have demanded more." | Retrospective assessment |
| `self_efficacy_post` | "I feel well prepared for future negotiations." | Post-negotiation self-efficacy |
| `confidence_post` | "I am confident about real-world negotiations." | Post-negotiation confidence |
| `negotiation_self_efficacy_post` | "I am confident I can get more than the first offer." | Negotiation self-efficacy (post) |
| `strategy_compliance` | "I applied the strategy I prepared." | Strategy implementation |
| `attention_check_imc` | Instructional manipulation check (IMC) | Data quality |

## Survey 3 — AI Evaluation (Treatment Only)

| Variable | Item |
|----------|------|
| `ai_helpful` | "The AI assistant helped me develop a strategy that suited me." |
| `ai_strategy_influence` | "The AI assistant influenced my negotiation strategy." |
| `ai_trust` | "I trusted the recommendations of the AI assistant." |
| `ai_confidence` | "The AI assistant helped me approach the negotiation with greater confidence." |
| `ai_implementation` | "I implemented the advice of the AI assistant during the negotiation." |
| `ai_counterfactual` | "Without the AI assistant, I would have negotiated worse." |
| `would_use_ai_again` | "I would use AI again to prepare for salary negotiations." |

\newpage

# Incentive Mechanism

The performance-contingent payment follows the induced value theory of laboratory experiments (Smith, 1976). All participants who complete the study receive lottery tickets; the number allocated depends on the final negotiated salary:

$$\text{lottery\_tickets} = \begin{cases} 2 & \text{no agreement (final\_salary} = 0) \\ 2 & \text{final\_salary} \leq 45{,}500\,€ \\ 3 & 45{,}500 < \text{final\_salary} \leq 48{,}000\,€ \\ 4 & \text{final\_salary} > 48{,}000\,€ \end{cases}$$

The threshold design distinguishes three performance levels: passive (≤ €45,500), average (€45,500–48,000), and effective (> €48,000) negotiators. The minimum of two tickets ensures that control-group participants who achieve weak outcomes nonetheless retain a positive incentive to complete the study.

**Prize:** 3 × €15 gift vouchers (winner's choice). E-mail addresses (`email`) are provided voluntarily, used exclusively for prize notification, and deleted from the dataset prior to scientific analysis.

\newpage

# Data Quality Indicators

| Variable | Type | Description |
|----------|------|-------------|
| `flag_articles_too_fast` | Boolean | Articles page completed in < 60 seconds |
| `flag_neg_too_fast` | Boolean | Negotiation concluded in < 30 seconds |
| `attention_check_correct` | Boolean | Attention check within articles answered correctly |
| `attention_check_imc_correct` | Boolean | IMC in Survey 2 correct (correct = rightmost value) |
| `mc_company_budget_correct` | Boolean | Manipulation check item 1 correct |
| `mc_anchor_strategy_correct` | Boolean | Manipulation check item 2 correct |
| `mc_salary_range_correct` | Boolean | Manipulation check item 3 correct |

For the primary analysis, exclusion of observations with `flag_articles_too_fast = True` and `attention_check_imc_correct = False` is recommended, as these patterns are indicative of non-serious participation.

\newpage

# Technical Infrastructure

| Component | Implementation |
|-----------|----------------|
| Experiment framework | oTree 5.x (Python) |
| Hosting | Heroku (Salesforce, USA) |
| Database | PostgreSQL |
| LLM API | Anthropic Claude (claude-haiku-4-5-20251001) |
| Real-time communication | oTree WebSocket (`live_method`) |
| Data protection | GDPR-compliant, pseudonymized, standard contractual clauses (Art. 46(2)(c) GDPR) |
| Scoring | Deterministic (temperature = 0), prompt-injection protected |

\newpage

# References

**Bowles, H. R., Babcock, L., & Lai, L. (2005).** Social incentives for gender differences in the propensity to initiate negotiations: Sometimes it does hurt to ask. *Organizational Behavior and Human Decision Processes, 103*(1), 84–103.
— Documents a 7–15% salary gain for effective negotiators; basis for ceiling calibration.

**Curhan, J. R., Elfenbein, H. A., & Xu, H. (2006).** What do people value when they negotiate? Mapping the domain of subjective value in negotiation. *Journal of Personality and Social Psychology, 91*(3), 493–512.
— Establishes subjective value of the interaction as an independent negotiation dimension; basis for the goodwill mechanism and Feature D.

**Galinsky, A. D., & Mussweiler, T. (2001).** First offers as anchors: The role of perspective-taking and negotiator focus. *Journal of Personality and Social Psychology, 81*(4), 657–669.
— Anchoring effect: opening offers frame the negotiation; basis for elevated concession rates in rounds 1–2.

**Kim, P. H., & Fragale, A. R. (2005).** Choosing the path to bargaining power: An empirical comparison of BATNAs and contributions in negotiation. *Journal of Applied Psychology, 90*(2), 373–381.
— In negotiations with a large bargaining zone, contributions and qualifications (Feature C) outweigh BATNA strength as determinants of outcome.

**Leventhal, G. S. (1980).** What should be done with equity theory? New approaches to the study of fairness in social relationships. In K. J. Gergen, M. S. Greenberg, & R. H. Willis (Eds.), *Social exchange: Advances in theory and research* (pp. 27–55). Plenum.
— Procedural justice as a multidimensional construct; basis for SVI items in Survey 1.

**Loschelder, D. D., Stuppi, J., & Trötschel, R. (2014).** "€14,875?!": Precision boosts the anchoring potency of first offers. *Social Psychological and Personality Science, 5*(4), 491–499.
— Precise, data-backed anchors are more persuasive than round figures; basis for Feature B (concrete market data).

**Northcraft, G. B., & Neale, M. A. (1987).** Experts, amateurs, and real estate: An anchoring-and-adjustment perspective on property pricing decisions. *Organizational Behavior and Human Decision Processes, 39*(1), 84–97.
— External reference prices systematically shift negotiation anchors; co-basis for Feature B.

**Pinkley, R. L., Neale, M. A., & Bennett, R. J. (1994).** The impact of alternatives to settlement in dyadic negotiation. *Organizational Behavior and Human Decision Processes, 57*(1), 97–116.
— BATNA possession vs. BATNA communication; highlights the limits of the scoring model (BATNA use is not reliably detectable from free text).

**Shea, N., Shi, W., Wan, Z., Yao, X., & Chen, X. (2024).** Retrieval-augmented argumentation for large language models. In *Findings of the Association for Computational Linguistics: EMNLP 2024* (pp. 1–14). Association for Computational Linguistics.
— Rationale as the minimum condition for persuasive argumentation; basis for Feature A as a gate operator.

**Smith, V. L. (1976).** Experimental economics: Induced value theory. *American Economic Review, 66*(2), 274–279.
— Methodological foundation for performance-contingent incentives in laboratory experiments.

**Van Kleef, G. A., De Dreu, C. K. W., & Manstead, A. S. R. (2004).** The interpersonal effects of anger and happiness in negotiations. *Journal of Personality and Social Psychology, 86*(1), 57–76.
— Constructive, positive tone increases counterpart concession-making; direct empirical basis for the tone effect (Feature D).