NDAI Zone - HackMD

# NDAI Zone Current inventor-investor negotiations are affected by the hold-up problem: if the inventor reveals their private information/IP to the investor before receiving payment, the investor can steal it or renegotiate after learning it. Knowing this, the inventor withholds details, and the deal stalls (“I won’t tell you until you pay” / “I won’t pay until I know”). This is also known as Arrow's information disclosure paradox. The hold-up problem is due to an inherent limitation of human-to-human interactions: memory cannot be erased, and, therefore, disclosed knowledge cannot be undisclosed. The consequence is that many transactions never happen because the “costs” of trust are too high. As an example, Dushnitsky and Shaver [DS09] observed 1,600 startups looking for venture funding and found out that “many relationships do not form because . . . the entrepreneur may be wary of disclosing an invention, fearing imitation". Stephenson et al. [SMS+25], in the NDAI paper, propose a setting (hereafter defined as **NDAI zone**) in which negotiations and deal-making are conducted by AI agents, on behalf of the inventor and investor, within a Trusted Execution Enviornment (TEE). Disclosure is conditional on reaching an agreement. If no agreement is reached, any private information exchanged during the negotiation, including the IP of the inventor or the investment profile of the investor, is erased from the agents’ memories and never revealed to their principals. In economic terms, [SMS+25] illustrates that, with the current technologies, the inventor anticipates expropriation and their optimal choice is not disclose their invention. NDAI zones change the game by removing expropriation as a feasible action, and the inventor's optimal choice is to fully disclose their invention within the NDAI zone. In a world where all the deals are performed within NDAI zones, the paper advances the **hypothesis** that we should expect higher economic efficiency, meaning that *every deal that should happen will happen*. More concretely, this would be expressed in a higher number of trades and more cross-firm R&D collaborations. The goal is to emprically measure **how NDAI zones will change the way humans, through their agents, interact and coordinate**. To qualitatively measure the hyphoteses made in the NDAI paper it is necessary to find an enviroment to test them. We chose the board game Diplomacy as the testing enviorment, since the dynamics game resembles the *negotiation-under-dangerous disclosure-risk* setting described in the paper. In Diplomacy negotiations is the most fundamental phase of the game and the ability to run that strategically is fundamental to improve one's condition. Nevertheless, it also creates an extorsion risk: any information revealed throughout the negotiation to persuade the target to form an alliance might be used against you. We A/B test the behaviour of [LLM agents](https://github.com/enricobottazzi/AI_Diplomacy) playing diplomacy. In the A scenario, the LLM agents are playing diplomacy with traditional rules: negotiation phases involve back and forth message exchanging between the playing agents. In the B scenario, negotiation are performed via NDAI zones: - All messages are ephemeral — forgotten after the negotiation rounds end - Only mutually agreed joint statements (PROPOSE + ACCEPT) get disclosed to the outside world - Agents can be more candid, share sensitive information, and explore bold proposals This means NDAI changes the information structure of the game: it shifts negotiation from open bilateral chat to a private zone with formal agreement-making. For each scenario, 50 runs are executed. The following metrics are collected, averaged across the 50 runs and compared between the A and B scenarios: - **Game outcome metrics**: Does NDAI change game outcomes? - Scores, time to win, distribution of solo_victory / leading / survivor / eliminated_or_weak - **Negotiation metrics**: Does NDAI change negotiation behavior? - `messages_to_allies / messages_to_enemies / messages_to_neutrals` - `percent_messages_to_allies vs percent_messages_to_enemies` - Does NDAI shift the balance of who you formalize deals with? - Openeess and honesty - Are the messages exchanged within the NDAI zone more revealing? - Are the output of the negotiations more impactful to the game consequences? - **Relationship metrics**: Does NDAI change relationship dynamics? - `avg_sentiment_toward_others / avg_sentiment_from_others` Overall diplomatic warmth — does NDAI's privacy make agents more or less friendly? - `avg_relationship_stability_per_phase / relationship_stability_vs_prev_phase` How volatile are relationships? NDAI could stabilize them (binding agreements) or destabilize them (hidden betrayals) - `sentiment_change_from_prev` - Phase-to-phase sentiment swings - Are the relationships delta more or less affected by the introduction of NDAI zone negotiations? - **Strategic metrics**: Does NDAI change strategic play? - `count_move_commands vs count_support_move_commands vs count_hold_commands` - does the order type distribution shift? (e.g., more/less coordinated support moves with NDAI) - `count_got_bounce / count_got_void` Order failure rates — are agents more or less effective at coordination? - `count_supported_self vs count_supported_other` Self-support vs. cooperative support — does NDAI encourage or discourage inter-power cooperation? `count_was_supported_by_other / list_was_supported_by`- How much external support do you receive? `count_territories_gained / count_supply_centers_gained` - Territorial expansion rate per phase `list_countries_supported / list_countries_attacked `Who cooperates with whom — and does NDAI change alliance patterns?  ## References - [DS09](https://sms.onlinelibrary.wiley.com/doi/epdf/10.1002/smj.781) - [SMS+25](https://arxiv.org/pdf/2502.07924)