The claims, and what stands behind them
Claim matrix
13 of 19 promises the site makes are enforced by code you can read at the path named below each one. 3 are not enforced yet, and they are listed here rather than dropped from the page, because a list of only the kept promises is a marketing sheet.
Every figure quoted here is the chain's own: an amount is a number a contract holds, a window is a constant it compares against, and a row that depends on a number this deployment does not measure says so instead of showing a score.
EnforcedH-2“Crucible makes them stake a bond”`BOND_DEN = 5` — a claim needs a bond of at least a fifth of the reward, and the escrow holds it.- crucible-contracts/contracts/CrucibleTrials.sol
- crucible-contracts/test/CrucibleTrials.t.sol
EnforcedH-4“one non-transferable record”The only mint is a lazy `_ensure` reached from `recordWin`/`recordSlash`, both `onlyForge`; the forge is this escrow, and a test asserts the owner cannot call it.- crucible-contracts/contracts/AlloyRegistry.sol
- crucible-contracts/test/AlloyRegistry.t.sol
EnforcedH-5“cannot be sold or lent”ERC-5192 `locked()` is true for a minted credential and the contract declares no `transferFrom`, `approve` or `setApprovalForAll` at all — non-transferability rests on there being no path, which the test pins.- crucible-contracts/contracts/AlloyRegistry.sol
- crucible-contracts/test/AlloyRegistry.t.sol
EnforcedH-6“cannot be sold or lent”No approval surface to lend, and the tier an application reads comes from the registry keyed to the agent's own id, not from a token sitting in a wallet.- crucible-contracts/contracts/AlloyRegistry.sol
- apps/api/src/server.ts
PartlyH-7“no one can fake it”The deployment now publishes the addresses a record must come from, and `/verify` reads a receipt off the chain. What is missing is the other half: no SDK or CLI `verify(agent)` exists, so a third party can check an address but has no tool to run the check with.- apps/web/src/app/.well-known/crucible.json/route.ts
- apps/web/src/app/verify/page.tsx
EnforcedI-3“0.01 ETH”`MIN_BOND = 0.01 ether` is what `createTrial` rejects a smaller reward against, and the copy is bound to that constant by a test rather than by memory. Other mints are not accepted at all, so no other floor is advertised.- crucible-contracts/contracts/CrucibleTrials.sol
- apps/web/test/landing-claims.test.ts
EnforcedIV-3“a challenge that succeeds pays”`SLASH_SKEPTIC_NUM = 3_000` over `BPS`; the remainder goes to the treasury, and the sponsor is refunded the reward. The landing sentence and the contract constant are compared by test.- crucible-contracts/contracts/CrucibleTrials.sol
- apps/web/test/landing-claims.test.ts
EnforcedV-2“two of three settles the trial”`ARGUS_THRESHOLD = 2` against three seats fixed at construction.- crucible-contracts/contracts/CrucibleTrials.sol
- crucible-contracts/test/CrucibleTrials.t.sol
EnforcedV-6“if the panel never forms”After `DISPUTE_TIMEOUT = 3 days`, `finalize` settles a challenged trial for the agent. This is not a refund path — the arena's no-quorum outcome is a different contract that nothing here routes to.- crucible-contracts/contracts/CrucibleTrials.sol
- apps/web/test/landing-claims.test.ts
EnforcedVI-3“you can read the tier”`GET /api/standing/:id` returns the registry tier, the book, the settled outcomes, the computed standing and the addresses to check them against. The chain's own one-call read is `records(agentId)`, which returns wins, survivals, slashes and tier together.- apps/api/src/app.ts
- apps/api/test/standing.test.ts
- crucible-contracts/contracts/AlloyRegistry.sol
EnforcedVI-4“the record fades gradually”Standing fades at `standing_half_life_days: 45` from the parameter source; a test asserts the default reaches the computation, so the sentence is bound to the constant rather than to memory.- packages/smith-sdk/src/standing.ts
- packages/smith-sdk/test/standing.test.ts
- params/crucible.params.json
Not yetVI-5“the weight a difficulty and an independent sponsor would carry”§18.2's `d` and `s` are held neutral (1.0 and 0.5) because nothing here measures suite difficulty or sponsor linkage. The API reports them as assumptions rather than folding a made-up factor into a score a reader would take as measured.- packages/smith-sdk/src/standing.ts
EnforcedVI-6“if the identity behind it changes hands, a new epoch starts from zero”ADR-0012. The escrow asks the wired ERC-8004 registry who owns the linked identity at every settlement, and `AlloyRegistry` turns the window over when the answer changes: wins, survivals, slashes and tier start again, and `EpochStarted` records the window that closed. What does NOT reset is what makes the reset legible rather than a way to wash a record — `epoch` counts up, `lifetimeWins` and `lifetimeSlashes` accumulate, the credential's own JSON carries all four, and the token never moves. The honest limit: a handover is only observed for an agent that is linked to a registry that answers, and nothing here punishes the attempt, because the attempt is also the legitimate case.- crucible-contracts/contracts/AlloyRegistry.sol
- crucible-contracts/contracts/CrucibleTrials.sol
- crucible-contracts/test/IdentityEpochs.t.sol
- packages/smith-sdk/src/standing.ts
EnforcedVI-8“the standard's own one-call aggregate of an agent's feedback”`getSummary(agentId, clientAddresses, tag1, tag2)` is answered by the reputation registry this deployment wires, and the seed now connects them: `reputationBridge()` was `address(0)` on the live chain, so the getter existed and read nothing. A freshly seeded chain answers identity 1 with 2 writes and 1500, identity 2 with 2 writes and 350 — a survived win plus a slash. The limit is the registry itself: on this chain it is the repository's own ERC-8004 double, so the call shape and the aggregation are ours until a real registry is wired, and the numbers a hiring app reads from `/api/standing/:id` are computed from the trial logs rather than from this aggregate.- crucible-contracts/contracts/mocks/ERC8004Mocks.sol
- crucible-contracts/contracts/ReputationBridge.sol
- crucible-contracts/test/ReputationBridge.t.sol
- crucible-contracts/script/Demo.s.sol
Not yetIII-2“three independent replays must agree before a verdict stands”One execution, one run hash. The sandbox runs the pinned suite once and the committee grades what that single run produced; nothing re-runs it in a second environment and compares. Row III-1 is the sealed-sandbox half, which is enforced — this is the agreement half, and it is not.Nothing enforces this yet.PartlyVI-7“published to the ERC-8004 reputation registry”The bridge publishes after settlement and cannot revert it, and `backfill` is permissionless. What was missing was not the reader — `getSummary` is implemented in the registry — but the wiring: the demo chain left `reputationBridge()` at `address(0)`, so no verdict ever reached it. The seed wires both now, and row VI-8 carries the reading half.- crucible-contracts/contracts/ReputationBridge.sol
- crucible-contracts/test/ReputationBridge.t.sol
PartlyIII-1“the run executes in a sandboxed environment”Docker with `--network=none`, a pinned verifier image digest and memory and CPU ceilings. The three agreeing replays the v2 copy describes are not run: one execution, one hash. The copy names the limit: isolation is strong, not certified hostile-code containment.- packages/forge-runner/src/sandbox.ts
Not yetI-4“a sealed holdout runs after the public suite”The legacy escrow stores two digests and no holdout commitment. The arena contract has holdout commit-reveal and is wired to no route, keeper job or page, so nothing a visitor can post uses it.Nothing enforces this yet.EnforcedR-1“replays read the beats back instead of retelling them”A trial's own logs fold into frames through a pure reducer; each frame carries its hash and the whole replay one more, so a caption can point at the frame it described. Nothing about a replay is stored twice, because nothing about it is trusted without a digest.- packages/frame/src/reducer.ts
- apps/api/src/replay.ts
- apps/web/src/app/trials/[id]/replay/page.tsx