vaultcrux.com · the technology behind cuecrux
Built to be checked.
This is the machine-room view of the CueCrux platform: how the engine, the event substrate, the daemon and the receipts actually work. Every claim on this site links to something you can read, run or verify offline. If you came to audit us, start here.
Structure mirrors architecture
Five subsystems, one spine
The platform is not a monolith with marketing names. It is five load-bearing subsystems, each with its own explainer page, each page mapping to real crates and real endpoints. Read them in spine order:
- crux engineretrieval + reasoning: lanes, fusion, token budgets
- corecruxthe event substrate: append-only, BLAKE3-sealed segments
- crux daemonone local binary: memory, retrieval, receipts, MCP
- receipts · rcxCROWN signing, capability tokens, signed refusals
- corpus · wikicruxreceipted knowledge with provenance kept at write time
crown · ed25519 · blake3
Receipts you can try to break
Everything above the substrate emits signed CROWN receipts, sealed into hash-chained segments. The signature binds to both the receipt id and the payload hash; ordering is carried by the seal chain. Flip the byte and watch verification refuse it, then run the same experiment for real: the tamper script ships in the repo and the export, offline-verify, tamper-reject cycle is a release-blocking CI gate.
corecruxctl verify-store --mode full --strict ok: true · 4 receipts · chain intact
Signatures bind to both the receipt id and the payload hash, so a receipt cannot be edited in place or transplanted elsewhere; ordering is carried by chained segment seals on the spine. You can run this exact experiment on your own machine: scripts/demo-receipt-tamper.sh in the Crux repo.
The verifier rejects malleable signatures and small-order keys; releases carry cosign keyless signatures, CycloneDX SBOMs and SLSA3 provenance. Details on the security page.
What a receipt proves, and what it does not. A CROWN receipt is a verifiable record: it binds the answer, its provenance and its retrieval metadata to a signature you can check offline, so tampering or substitution is detectable. It is not a proof that an agent behaved correctly or that a human read the result, identity and constraints live in the passport, conduct does not. Exactly which event classes are covered, and how, is enumerated in the assurance & coverage matrix.
Receipts are COSE_Sign1 envelopes over Ed25519, so you do not need our code to check them. The standalone crown-verify library has zero CueCrux dependencies, and the receipts follow the SCITT compatibility profile for transparency-service interoperability.
The proof spine
Numbers that carry their own receipts
A benchmark you cannot re-run is an anecdote. Every published figure here carries its corpus, lane flags and commit SHA, and the daemon numbers are regression-gated in CI against pinned baselines.
The one on the right is the economics claim: retrieval under a hard token_budget, enforced server-side, returning substantially fewer tokens than naive top-K stuffing at equivalent answer quality. We do not headline a single percentage: the size of the saving depends on corpus and query mix, so ScoreCrux publishes the methodology, negative controls and per-model runs for you to re-measure on your own workload.
Tokens returned for one query
lower is better
naive top-K · no budget
token_budget=4000 · server-enforced
token_budget=2000 · server-enforced
token_budget=500 · server-enforced
Budgets are enforced server-side: metadata first, content only when it earns its tokens. Same corpus, substantially fewer tokens than top-K stuffing. Reproducible from docs/benchmarks.md.
No dead ends
Where to go deep
- /architectureThe whole system in one map: planes, stores, trust boundaries.
- /securityThreat model, key handling, fail-closed behaviour, release attestation.
- /benchmarksThe proof spine: reproducible runs with corpus, flags and commit SHA.
- /guidesWorking documentation for operators, developers and agents.
- /for-agentsThe surface agents actually consume: MCP tools, budgets, receipts.
Sceptical is the intended audience.
Read the architecture, run the tamper test on your own machine, re-run the benchmarks. If you are wiring an agent instead, the credentialed path is still here.