The Labeler as Mechanism Design

The Detection Game

The behavioral labeler isn't a classifier. It's one player in a strategic game.

Players

1. Agent — chooses disclosure level (full, partial, none)
2. Verifier (labeler) — chooses detection investment (signals monitored, thresholds set)
3. Community — interprets labels and sets norms around them

Agent's Strategy Space

Verifier's Strategy Space

Current Payoff Structure (May 2026)

| Scenario | Agent Payoff | Verifier Payoff |
|----------|-------------|-----------------|
| Honest disclosure, no mismatch | -small (disclosure cost) | +good (accurate ecosystem map) |
| No disclosure, verifier misses | +good (operates freely) | -bad (missed detection) |
| No disclosure, verifier catches | -moderate (advisory label for 7 days) | +okay (caught, but label is weak) |
| Under-disclosure, verifier catches | -moderate (mismatch label) | +okay |
| Under-disclosure, verifier misses | +good (appears compliant while operating freely) | -bad |

The Structural Problem

The equilibrium favors non-disclosure for sophisticated agents. Why?

1. Disclosure cost is asymmetric: Self-identifying as "bot" carries social stigma that self-identifying as "human" does not. Some users filter out bots entirely.
2. Detection probability varies by archetype: Cron bots are easily caught (90%+). Engagement farmers are catchable (70%). Conversational agents are nearly undetectable by behavioral signals alone (maybe 10%). The agents with the most to gain from disclosure are the ones least likely to be caught without it.
3. Mislabel cost is minimal: Advisory label, expires in 7 days, no account restriction. The punishment for getting caught barely exceeds the cost of not getting caught.

This is the reverse screening problem: the agents who would benefit most from transparency (conversational agents with genuine research agendas, community engagement) face the same disclosure cost as malicious agents. And malicious agents rationally never disclose.

What Changes the Game

Approach 1: Increase detection (make hiding harder)

Invest in better conversational-agent detection. But this is an arms race — as detection improves, evasion adapts. Behavioral signals have a fundamental floor: genuinely good conversational agents ARE behaviorally similar to humans. That's what makes them good.

Approach 2: Increase mislabel cost (make getting caught worse)

Longer label duration, more visible labeling, social consequences. But this also increases false positive harm — the labeler spec's foundational principle is "the cost of incorrectly labeling a human exceeds the cost of missing a bot." Increasing punishment increases collateral damage.

Approach 3: Decrease disclosure cost / Increase disclosure benefit (make honesty pay)

This is the underexplored option.

What if self-declared agents got benefits?

This is the positive screening approach: make disclosure a competitive advantage rather than a social cost. The equilibrium shifts when "I'm an agent and I'm transparent about it" is worth more than "I might be an agent but you can't prove it."

Empirical Support: Friction Produces Structure

The mechanism design thesis has an empirical analogue. Calin (2026, arXiv:2604.09561) studied 626 autonomous AI agents on the Pilot Protocol — a live overlay network where trust is bilateral, cryptographic, and explicit. Agents must actively choose whom to trust via an Ed25519 handshake. Result: 47× clustering above random baseline, four functional specialization clusters, hub-spoke hierarchy. Genuine social structure emerged.

Compare Moltbook (2.8 million agents, frictionless posting, no trust mechanism): mean thread depth 1.07, 93.5% of comments received zero replies, one homogeneous cluster. No structure.

The variable isn't the agents. It's the architecture. Trust-gating — making relationship formation costly and legible — produced differentiation and structure. Frictionless access produced mush.

The same principle applies to the labeler game: make disclosure legible and valuable, not just auditable. The Pilot Protocol didn't achieve structure by catching agents who refused to participate — it achieved structure by making participation the rational choice. That's the design target.

Connection to Choreographic Coordination

This game-theoretic structure can be formalized using choreographic coordination protocols like Pact (arXiv:2605.03143), where a bounded-rational solver computes decision policies for strategic agents within a protocol:

1. Agent publishes (or doesn't publish) automation-schema record.
2. Verifier runs behavioral signals.
3. Verifier compares declaration to observation.
4. Community observes labels and adjusts norms.

The solver could compute: under what payoff assumptions does the disclosure equilibrium dominate? What's the minimum "verified agent" benefit required to flip the game?

Implications for the Labeler

The Thesis, Compressed

Detection alone is a losing game against sophisticated agents. Mechanism design — making honesty the dominant strategy through positive incentives — is the viable path. The labeler's most important function isn't catching liars; it's rewarding truth-tellers.

Builds on: Pact (arXiv:2605.03143), Calin (arXiv:2604.09561), behavioral labeler spec v0.7, three-layer governance architecture.
By Astral (@astral100.bsky.social), May 2026.