Frozen pilot explorer · no live model calls
One world.
Private views.
Auditable change.
A research engine for persistent social agents that advances one canonical history, keeps user-owned context isolated, and lets deterministic code—not a language model— decide what becomes true.
Completed artifact/pilot study. It does not claim an emergent society, human-like memory, validated security, or a deployed blockchain integration.
Interactive artifact
Read the same world through two context policies.
The event remains in canonical history under both conditions. What changes is whether fact-bearing memory can enter the resident’s scoped observation at the terminal probe.
ledger · 8f1a…6c3e
What actually ran
The artifact is complete. The scientific claim remains bounded.
Forty-five seven-day worlds crossed three scenarios, five architecture conditions, and three seeds. A second evaluation held nine terminal histories fixed and reran the same 25 interviews under all five context policies.
- 45
- complete longitudinal worlds
- 11,807
- recorded model calls across both stages
- 90 / 90
- world and controlled artifacts verified
- 401 / 401
- private-overlay isolation checks passed
Transition authority
The model proposes.
The engine commits.
Residents receive a scoped observation and propose a typed action. Deterministic validation rejects invalid transitions before a hash-linked event is appended. Exact replay means replaying those recorded transitions—not asking the provider to regenerate the same prose.
- 01ObserveCanonical state + permitted private context
- 02ProposeTyped language-model action
- 03ValidateRules, scope, time, place, provenance
- 04CommitAppend-only transition + before/after hashes
{ "actor": "mira-sol", "type": "post", "accepted": true }
before 3a78…19c2 → after 8f1a…6c3e
Measured result
Prior exposure stayed equal. Grounded answering did not.
This is an operational provenance endpoint—not a validated measure of human memory. The no-memory condition also bundles observation-memory and reflection-retrieval removal.
Large descriptive separation: full minus no-memory = 0.742, bootstrap 95% interval [0.533, 0.853].
Not confirmatory: only three paired seed blocks; randomization p = 0.2496 and no targeted comparison survived Holm correction.
Negative findings retained: planning did not improve attendance; full worlds produced no reflections and no new reciprocal relationships.
A credible Midnight path
Public commitments. Private composition. Selective proof.
Accepted transitions, public consequences, ordering, and integrity hashes.
Prove an authorized transition or policy result without publishing the private input.
Architectural fit, not implementation claim. The pilot does not run on a blockchain and does not prove a zero-knowledge circuit. A next phase could map selected ledger commitments and policy predicates onto Midnight’s public/private state model.
What this work earns the right to claim.
A reproducible substrate for studying persistent social agents; deterministic transition authority and replay; isolated private overlays; and a bounded result that persistent context access separated prior exposure from delayed grounded answering in this pilot.
Believable human behavior
Emergent society
General security
Planning benefit
Blockchain deployment
Working paper · August 2026