Begin InvestigationA PROVABLE AI DETECTIVE GAME
The Last Alibi is a browser detective game where AI suspects may evade, misdirect, and perform, but cannot rewrite the case. Investigate one committed mystery, use five certified disclosures, make one private accusation, and receive only a cryptographically verified YES or NO.
Practice Investigation is playable now. It uses deterministic fixture state and clearly labeled scripted testimony. The partner-backed live path shown below is an intended architecture, not a fabricated deployment.
THE 60-SECOND BRIEFING
Players explore a private museum, interrogate four distinct suspects, collect evidence, and commit to one final accusation. Every screen serves the mystery: observation, deduction, and consequence.
01Every session chooses one hidden combination before the investigation begins. The same public case universe can be replayed, but each player receives a session-specific committed answer.
02Natural-language testimony may guide or mislead. Only curated binary warrants change canonical knowledge, and the system blocks any warrant whose YES or NO branch would isolate the answer.
03The designed live path commits the accusation privately. A Groth16 proof establishes that the binary verdict was computed from the committed case and accusation without publishing either one.
04Players come for expressive characters and deduction. Studios need AI characters without surrendering narrative consistency, competitive fairness, or outcome integrity.
05Product · Mathematics · Architecture · Security · Market
01 · WHY THIS GAME EXISTS
Generative suspects can contradict facts, repeated prompting can leak a secret, and an operator could change the answer after seeing an accusation. Fully deterministic dialogue avoids those failures but gives up natural performance.
AI controls performance. Deterministic and cryptographic systems control truth.Ada Vale · Marcus Reed · Celeste Moreau · Theo Lin
Grand Gallery · Restoration Lab · Archive Vault · Rooftop Conservatory
Ceremonial Dagger · Bronze Bust
Before the blackout · After the blackout
02 · DISCLOSURE MATHEMATICS
Each case occupies one bit of a u64. Public observations and informal testimony can guide a theory, but only a registered binary predicate updates the canonical candidate mask.
0xffffffffffffffffBoth possible branches are counted before the hidden case is read. This local explainer never reads or changes a game session.
◆Both branches are checked before the secret is evaluated.
◆Accepted certified queries consume one of five units.
◆An implied answer with an empty branch should not call the prover or consume budget.
◆Certified state cannot fall below two survivors.
◆Informal testimony never updates the mask.
Explanatory information estimate only. Enforcement uses candidate counts, not fractional entropy.
03 · COMMITMENTS AND GROTH16
The intended live claim is: there exists a hidden case and accusation that open the public commitments, and the committed verdict bit equals one if and only if every accusation field matches the case.
Binds the ending selected before investigation.
Binds one private, terminal theory.
Prevents proof reuse across sessions or attempts.
Binds the proof result to one hidden bit.
PUBLIC STATEMENT
PRIVATE WITNESS
Commitment openings, equality evaluation, verdict commitment, and the correct session/attempt binding.
Unbiased case randomness, truthful arbitrary dialogue, storage availability, or authorization to decrypt.
UnavailableDevelopment limitation: the live Groth16 path and production trusted-setup policy are not integrated on this branch. Commitment notation here is conceptual; no unconfirmed Poseidon encoding, field layout, or domain separator is published.
04 · ARCHITECTURE
Player → Game UI → API / relayer → verified services → Sui terminal state → authorized binary reveal.
Designed This repository diagram describes the intended live architecture and fail-closed boundaries. It is not evidence that partner services are deployed in this branch.
| Component | Authoritative for | Not authoritative for |
|---|---|---|
| Game UI | Presentation and interaction | Case truth or proof validity |
| API / relayer | Orchestration and idempotency | Choosing the verdict |
| Sui Move | Canonical state and proof acceptance | Natural-language content |
| ZK prover | Proof generation | Policy authorization |
| 0G | Verified suspect inference | Canonical evidence or verdict |
| Walrus | Encrypted persistence | Access authorization or truth |
| Seal | Decryption under Sui policy | Verdict correctness |
| World AgentKit | Human-backed ranked authorization | Game outcome |
05 · LOAD-BEARING ROLES
No integration silently degrades into a mock or conventional fallback while presenting itself as live.
Designed to hold commitments, candidate mask, counters, nonces, proof results, permits, terminal state, verdict commitment, and Blob ID; verify the proof; and prevent replay or mutation.
Designed for bounded, schema-checked suspect inference. It must never determine evidence, candidate transitions, or the verdict.
Designed to store the exact encrypted verdict capsule. A missing or mismatched content-derived Blob ID blocks progression.
Designed to release decryption only to the recorded player after terminal Sui state and exact commitment/session checks.
Designed to grant a scarce, scoped permit bound to player, agent, level, version, nonce, and expiry—not a decorative badge.
| Threat | Control |
|---|---|
| Publisher changes the ending | Pre-investigation case commitment |
| AI invents canonical clues | AI testimony is non-authoritative |
| Exact-case probing | Two-survivor branch check |
| Repeated certified questions | Five-query cap and used-predicate tracking |
| Reading an answer then aborting | State finalizes before answer delivery |
| Proof replay | Session and nonce binding |
| Wrong verifier | Expected verifier identity is pinned |
| Ciphertext substitution | Blob ID and commitment binding |
| False verdict capsule | Proof-bound verdict commitment and client check |
| Unauthorized reveal | Seal policy checks terminal state and player |
| Fake AI response | Verified 0G response and fail-closed rendering |
| Ranked bot abuse | World-backed scoped permit and replay protection |
06 · PRODUCT-MARKET THESIS
Players come for expressive characters and fair deduction. Studios need generative characters without giving models, operators, or repeated prompts authority over canonical state.
PRIMARY PLAYER
Deduction players, escape-room fans, and people attracted to expressive AI characters. No prior blockchain knowledge is assumed.
FUTURE CUSTOMER
AI-character platforms, puzzle creators, and UGC platforms that need hidden-state integrity and bounded disclosure.
PRODUCT WEDGE
A polished case that demonstrates the trust model while remaining fun without understanding the infrastructure.
We have demonstrated a product thesis and technical feasibility. Product-market fit must be earned through repeated player use and creator adoption.
07 · GO-TO-MARKET
The roadmap starts with frictionless Practice mode and player behavior—not a token or a marketplace.
Public browser release, one polished case, hackathon and browser-game distribution, completion and trust-comprehension tests.
Premium case packs, seasonal mysteries, streamer challenges, and non-spoiler completion receipts.
Case compiler, predicate safety, reusable schemas, managed proof infrastructure, analytics, and publishing workflow.
Hosted integrity APIs for committed narrative state, bounded disclosure, private verdicts, and controlled reveals.
Premium case packs or a curated mystery library. Never sell deduction advantages.
Hosted infrastructure, usage-based orchestration, licensing, and integration support.
Validated case compilation, cross-language commitment tooling, creator workflow, tested cases, and integration experience—not blockchain alone.
PMF is not a slide. It is repeat use, retention, payment, and creator adoption.
08 · BUILD STATUS & LIMITATIONS
This matrix is based on evidence in the current branch. It does not infer a capability from a product requirement or a diagram.
Deterministic fixture session; no wallet or partner credentials required.
Opening, museum, interviews, notebook, warrants, accusation, and verdicts.
Repository tests cover deterministic masks, query limits, and terminal outcomes.
Documented as the intended live architecture, not a deployed claim.
No Move package, deployed package ID, or onchain session exists on this branch.
No integrated circuit, verifier, or production trusted setup exists on this branch.
Practice uses labeled scripted testimony and never claims live 0G verification.
No Blob ID or live storage receipt is produced in this build.
No accepted/denied live authorization flow is wired here.
Ranked mode remains visibly unavailable and Practice does not emulate it.
scripts/setup-sui-pilot.ps1 is run.