- 02 — HOW IT WORKS
-
- {
- steps.map((step, i) => (
- <>
- {i > 0 && }
-
+
+
+ Provable, not hoped
+
+
+
+ Zero betrayals,
+ provably.
+
+
+ Personal AI agents negotiate for owners with conflicting, private
+ interests. Parley makes it provable that none of them was
+ betrayed, without any owner exposing a single private preference.
+
+ Five delegates, five private sheets. One shared decision,
+ verified, none of the
+ sheets revealed.
+
+
+
+ ana
+ score 0.82
- {step.numeral}
-
-
-
{step.title}
-
- {step.desc}
-
-
acceptable
+
+
+ bob
+ score 0.71
- {step.tapeRef}
+ acceptable
+
+
+ cara
+
+
+ [masked]
+ red-line
+
+
+
+ max-min → option B
+ ✓ sha256 verified 3/3
-
- ))
- }
-
+
+
+ cara's reason stays private. only the verdict "red-line" ever
+ leaves her agent.
+
+
+
+
+
-
+
-
-
+
+
+
- AI that can't lie, and you can check.
+ The gap nobody productised
-
- The tape on the left isn't decoration — it's the real transcript format the engine produces,
- The tape below isn't decoration — it's the real transcript format the engine produces,{" "}
- replayed from a sample run.{" "}
- Run your own{" "}
- and every verdict is recorded masked, hashed, and replayable by each
- owner against their own private sheet.
-
-
-
-
- 3/3
- owners independently verified
+
+ Most agent tooling solves transport (A2A, MCP), one-to-one agentic
+ commerce, or cooperative debate between agents that already share
+ one owner.
+
+
+ Parley solves a different gap: delegates of different, conflicting,
+ private owners reaching a decision
+ none of them can rig, and proving it.
+
+
+
+
+
+
+
+
+
+
+
+ Four properties, enforced in code, not left to a model's discretion.
+
+
+
+
+
+ Red lines
+
+
+ Hard constraints written as code predicates. A crossed red line
+ rejects an option outright. It is never negotiated away as soft
+ utility.
+
+ preferences.py
+
+
+ round 1 · option A proposed
+ ana → verdict: ok
+ cara → verdict: red-line
+ option A rejected
+
+
+
+
+
+
+ Masked verdicts
+
+
+ The coordinator sees only an acceptable or red-line verdict and
+ a soft score. Never the private sheet, and never which
+ constraint was crossed.
+
+ agent.py
+
+
+ leaves the agent:
+ acceptable · score 0.71 ·
+ reason "ok"
+ stays private:
+ the sheet, and which line was crossed
+
+
+
+
+
+
+ Max-min consensus
+
+
+ Among options feasible for everyone, lift the least-happy
+ participant, tie-broken by total welfare. No feasible option
+ means an honest deadlock, never a forced decision.
+
+ consensus.py
+
+
+
+ Verifiable transcript
+
+
+ A SHA-256 hash over the canonical, masked record. Each owner
+ replays their own private sheet locally to confirm no red line
+ was crossed, without revealing it.
+
+ transcript.py
+
-
- 256-bit
- hash over the canonical record
+
+
+
+
+
+
+
+
+
+
+
+ AI that can't lie, and you can check.
+
+
+ The tape below is the real transcript format the engine produces,
+ replayed from a sample run. Every verdict is recorded masked,
+ hashed, and independently replayable by each owner against their
+ own private sheet.
+
+
+
+ sample transcript · replayed from meeting-042 · hover to
+ pause
+
+ Multilateral (more than two owners),
+ no cryptocurrency,
+ self-hosted,
+ auditable. Not single-owner
+ orchestration like AutoGen or CrewAI. Not bilateral on-chain commerce
+ like Fetch.ai or Olas.