Prove how a figure was derived
Custody proves what arrived; reconciliation proves a join verdict. Nothing proved how an output was derived — until now. at1 databom seal emits a signed certificate over {input hashes, program hash, output hash, environment fingerprint}. An independent party runs at1 databom verify, which re-executesthe sealed program on the sealed inputs and confirms the output reproduces byte-exact. Certs compose into a hash-continuous pipeline DAG, so an entire lineage — raw inputs → intermediate artifacts → final figure — is re-derivable and tamper-evident.
- re-execute
- an independent party re-runs the sealed program and confirms the output hash
- tamper = 0
- edit an input, the program, or the output and verification fails — no silent pass
- pipeline DAG
- one stage's output hash == the next stage's input hash, checked across the graph
- refuse
- a non-deterministic step is refused, never rubber-stamped (fail-closed)
Seal a step, then re-derive the whole pipeline
at1 databom seal stage1.py --in raw.csv --out clean.bin -o s1.cert # step='stage1' 1 input(s) -> output f10ee900838c... (511,406B) sig=ed25519-tofu # A5 overhead: 1,644B / 511,406B = 0.321% (<=1% OK) # self-verify: VERIFIED — re-executed 'stage1' -> output reproduces byte-exact at1 databom verify s1.cert # re-executes; free, no account # VERIFIED — re-executed 'stage1' on 1 input(s) -> output f10ee900838c... reproduces at1 databom graph s1.cert s2.cert # compose into a hash-continuous pipeline # VERIFIED: 2 stage(s), 1 hash-continuous edge(s); acyclic + connected # edge stage1 -> stage2 (continuity f10ee900838c...)
The certificate is signed with an ed25519 trust-on-first-use key and filed in the codec-registry catalog. It carries the program source so anyone can re-execute it — and stays well under 1% of the artifact.
Verify by re-execution
A verifier doesn’t trust the sealed hash — it re-runs the program on the sealed inputs in the pinned environment and confirms the output reproduces. Tamper an input, the program, or the output hash and it fails; a silent pass is structurally impossible.
Pipelines, not just steps
Certs compose transitively: an edge exists only where one stage’s output hash equalsthe next stage’s input hash. The graph must be acyclic and connected — swap an intermediate artifact and the chain breaks, so a whole lineage re-derives from its root inputs or it doesn’t verify.
Honest about determinism
A step that doesn’t reproduce its own output on a second run is refused, at seal and at verify — the same fail-closed discipline as certified replay, implemented independently. DATABOM proves derivation; it does not claim the program is correct, nor hide the program (it’s carried so it can be re-run).
Part of the Trust & Regulated family alongside Custody Receipts, Reconciliation and AT-1 Ledger. The AI-Act / regulated data-lineage wedge: “show me how this figure was produced, and prove it.”