Trust & Provenance
Pangura's evidence layer makes analytical outputs independently checkable. This page explains what is recorded, why history cannot be silently rewritten, and how you — or your auditor — can verify it without trusting us.
Wording discipline: our systems make tampering detectable and locatable — we do not claim tampering is impossible, and we characterize model behaviour rather than certify it.
Content identifiers (CIDs)
Every persistable artifact — a source paper's PDF, an operator's implementation, a dataset, a result — carries a content identifier of the form sha256:<64-hex>. Structured objects are first canonicalized under RFC 8785 (JSON Canonicalization Scheme), so logically identical objects always hash identically. Change one bit anywhere and the CID changes; every downstream reference breaks loudly instead of silently.
The hash-chained ledger
Evidence events append to a ledger where each entry's hash is computed from its own content plus the previous entry's hash. Editing any historical line therefore breaks every hash after it. Pre-existing history from before the chain was introduced is preserved byte-for-byte in a frozen file whose whole-file CID is anchored in the chain's first (genesis) entry — so even pre-chain history cannot be quietly altered.
How verification works
Verification is a deterministic recomputation: walk the ledger from genesis, recompute each entry's hash from its predecessor and body, and compare. It needs no trust in Pangura, no proprietary tooling, and no network access — the procedure itself is the proof. Operators (admins) can run this from the Engine Cockpit's Provenance panel, which also exports the full evidence bundle as a JSON file suitable for handing to an auditor or regulator.
Operator fingerprints
Each analytical operator's identity is pinned by a CID computed from its declared source files — meaning any third party holding the code can recompute the same fingerprint. Golden tests and attestations bind to these fingerprints: change the implementation and the fingerprint changes, automatically invalidating stale evidence. Behavioural (LLM-in-the-loop) operators additionally pin their prompts and model versions, so attestations can never silently outlive the thing they attested.
What this layer does — and does not — claim
It claims: content integrity (bit-level change detection), append-only history (tamper location), reproducible operator identity, and time ordering. It does not claim: correctness of any analysis, completeness of coverage, or safety of any model. The substrate makes claims checkable; it never replaces checking.