Decision provenance for consequential AI

Prove what the AI decided —
on what basis, and that the record is intact.

Provenant turns every consequential decision from physical systems and multi-step agents into a signed, hash-chained Decision Record. Your model and data never leave your environment. Only sealed commitments cross the boundary.

  • Edge-signed · Ed25519
  • Merkle transparency log
  • Why + evidence packages
  • Disconnected-first capture

Built for high-stakes AI · aerospace beachhead · horizontal core · agentic + physical

4 layersexec · capture · chain · verify
2 modalitiesone Decision Record
0 model IPleaves your boundary

The problem

When the decision mattered, the record did not hold up.

Whether an agent approved a transaction or a physical model cleared a vehicle, three failures show up under examination — and each is fatal to audit, liability, and certification.

Fragmented story

Reasoning, action, and human review live in separate systems. Reconstructing one decision under examination is slow to impossible.

Logs that can lie

Conventional logs sit inside writable storage. Auditors increasingly discount records that cannot prove they have not been altered.

Unanswerable “why”

The critical question is the same across money and machines: what did the system decide, on what basis, and can you prove this is the original record?

Product

One horizontal engine. Thin vertical profiles.

Provenant is a decision-provenance platform: capture at the edge, verify on the platform. Aerospace is the first vertical — not the product’s identity. The same core seals agent trajectories and physical control decisions.

L1

Deterministic / reproducible execution

Pin environment fingerprints, model identity, and enough context that a decision can be attested — not just logged.

L2

Reasoning trace capture

A single Decision Record schema — input, reasoning, confidence, OOD flags, output — fed by physical or agentic adapters.

L3

Cryptographic provenance chain

Per-source hash chain, Ed25519 signatures, Merkle transparency log, optional RFC 3161 timestamp anchors.

L4

Verification & compliance profiles

Rebuild any decision, answer why, export self-verifying evidence packages for DO-178C / EASA, EU AI Act Art. 12, and more.

How it works

Capture where the decision happens. Prove it anywhere.

1

Onboard

Register tenant, content-addressed Model ID, and source public keys. Private keys stay at the edge.

2

Capture

Adapter seals input · reasoning · output hashes, chains prev_record, signs with Ed25519, buffers if offline.

3

Ingest

Platform four-check pipeline: schema, signature, identity reconciliation, chain continuity — then ledger leaf + payloads.

4

Investigate

Why narrative, residual trails, inclusion proofs, and regulator-ready evidence packages — months later, still intact.

from provenant import CaptureSession, Signer, FileChainStore
from provenant import JsonlEmitter, ThresholdCapture, compute_model_id

model_id = compute_model_id(
    weights="checkpoints/aeromind_pinn.pt",
    config={"arch": "tilt_rotor"},
    quant="fp32-cpu",
)
session = CaptureSession(
    tenant_id="op-northstar",
    source_id="N-AM042",
    model_id=model_id,
    signer=Signer.load("keys/n-am042.key"),
    chain=FileChainStore("provenance/chain"),
    emitter=JsonlEmitter("provenance/buffer"),
    policy=ThresholdCapture(),  # FULL when flagged / OOD
)
sealed = session.capture(
    input_payload=telemetry,
    reasoning=health_result,   # model rationale ("why")
    output={"decision": "CONDITIONAL"},
    output_label="CONDITIONAL — battery limiting",
    confidence=0.88,
)
from provenant import CaptureSession, FullCapture, Signer
from provenant.integrations.agentic import AgenticCapture

session = CaptureSession(
    tenant_id="t-…", source_id="agent-1", model_id="MID-agent",
    signer=Signer.generate(), policy=FullCapture(),
)
cap = AgenticCapture(session, task={"instruction": "…"})

# Each step seals on the hash chain…
cap.step("plan", reasoning=plan, output=plan_out)
cap.step("tool_call", input_payload=tool_in, output=tool_out)
cap.step("llm", reasoning=thoughts, output=answer)

# Parent record binds the full trajectory
parent = cap.finalize()
# verify_trajectory(parent, steps) → ok
# Platform
GET /records/{decision_id}/why
# → narrative + sealed model_rationale + ledger proof

# Offline SDK
from provenant import build_decision_explanation
why = build_decision_explanation(
    sealed.record,
    payloads={"reasoning": reasoning, "input": inp, "output": out},
    public_key_hex=signer.public_key_hex,
)
print(why["narrative"])
print(why["why"]["model_rationale"])

Modalities

Same Decision Record. Different adapters.

It would be marketing fluff if agentic and physical were two products sharing a logo. They are not — the provenance primitives are identical. Only the capture adapter and compliance profile change.

Physical AI

Edge inference you can certify against

Instrument health scorers, control loops, and supervisory models. Content-addressed model identity, environment fingerprints, residual / rule trails in reasoning, episode windows for flight-recorder style capture.

  • AeroMind design-partner pattern (eVTOL health)
  • THRESHOLD fidelity at fleet scale
  • DO-178C / EASA-oriented evidence export
  • Disconnected-first buffers in flight / on-prem
Agentic AI

Trajectories you can replay and attest

Multi-step agents seal each plan, tool call, retrieval, and LLM step. A parent record binds the trajectory. Same chain, same ledger, same Why path for auditors and risk teams.

  • Step kinds: plan · tool · llm · retrieval · human
  • Trajectory verification (ordered chain + parent)
  • Framework-agnostic adapters (duck-typed)
  • EU AI Act Art. 12 logging posture

Architecture

The trust boundary is the product’s spine.

Provenant never ingests your model or raw training data. Capture is push-only. Payload and proof are deliberately split so the ledger stays lean and externally verifiable.

Never ingest the model

Only cryptographic commitments, manifests, and runtime fingerprints. Weights stay home.

Push, not pull

The platform never reaches into your runtime. Attack surface stays where you control it.

Payload / proof split

Sensitive bytes in object store; hashes on the ledger. Reconstruct when authorized — not by default.

Disconnected-first

Aircraft and air-gapped agents produce correct, chained records offline; sync later with drain.

Why now

Provability is becoming a legal obligation.

Regulation, agentic complexity, and emerging standards converged. Provenant is built toward the full stack auditors and certification bodies increasingly require — not another dashboard of editable logs.

EU AI Act · Art. 12

High-risk systems must keep logs that enable traceability of results with integrity — a horizontal mandate spanning physical and agentic systems.

EASA / aerospace

AI for aviation is expected to be explainable, predictable, and traceable. Decision trails must survive certification and investigation timelines.

Emerging standards

Industry drafts for cryptographically verifiable decision audit trails confirm the category — and that point solutions will not compose cleanly.

Profiles shipping aero-do178c-easa aero-change-do178c eu-ai-act-art12 + vertical profiles via adapter

Platform

Everything operators need after capture.

Investigator UI

Dashboard, decisions, Why, evidence, and integrity events. Deep-link any decision from your OEM tools.

Model registry

Content-addressed Model IDs. Identity changes are explicit registrations — never silent overwrites.

Merkle ledger

RFC 6962 roots, inclusion and consistency proofs, signed tree heads, dual file + RFC 3161 anchors.

Capture fidelity

FULL, HASH_ONLY, or THRESHOLD — store payload bytes only when reconstruction is worth the cost.

Evidence packages

Self-verifying exports: narrative, payloads, attestation, and machine-readable proof for third parties.

Prometheus metrics

Ingest, ledger, and anchor health for production SRE — not just a pretty status page.

Design partners

Instrument one high-stakes system. Prove the core. Replicate.

We partner with aerospace and agentic teams who need decisions to survive audit — not just demos. Beachhead: physical AI in aviation. Horizontal ambition: every vertical where “why” is a liability question.

  • SDK + platform already shipping (v0.6 · schema 1.2)
  • Capture adapters for physical and agentic paths
  • Operations & onboarding manual for your security team

Request a conversation

No spam. Submissions go to contact@aeromind-ai.com — we reply to serious design-partner and enterprise inquiries.

Prefer email? contact@aeromind-ai.com