What your auditor
actually receives.
A step-by-step walkthrough of a real evidence pack generated by ThetaZero. Every governed workflow produces this automatically — no manual write-up required.
Who ran what, when, and where.
Every ThetaZero execution begins with a cryptographically sealed metadata record. This tells auditors exactly which agent ran, which model was used, when it started, how long it took, and which compute node handled it — with no possibility of retroactive modification.
The metadata record is sealed at execution start using HMAC-SHA256. Any alteration — including retroactive changes to the timestamp, agent version, or model — invalidates the seal and will be detected by the Merkle root verification in Step 7.
Auditors can verify: this specific agent, running this exact model, executed at this timestamp, on this compute node — with no reliance on ThetaZero's live systems to confirm it.
What the agent received.
ThetaZero records the structure and fingerprint of every input — but never retains the content. The evidence pack shows what was submitted (document type, page count, SHA-256 hash) without including the document itself. Your data never persists after execution.
The input SHA-256 hash lets auditors verify that this specific document was submitted without ThetaZero ever needing to store or expose the document contents. The hash is a fingerprint: it proves what was processed without revealing what was in it.
This satisfies GDPR's data minimization principle — you can prove data was processed correctly without retaining a copy of the data.
Every step. Exact timing.
ThetaZero captures a timestamped execution log for every step the agent performs — tool calls, model calls, data retrievals, and decision points. This is the full decision log, not a summary. Auditors can trace every action back to its exact timestamp and verify it matches the sealed record.
The processing trace is an immutable, timestamped record of every decision and action. Auditors can verify not just what the agent did, but when each step occurred. The trace is included in the Merkle hash — any alteration would break the seal.
Timestamps are generated inside the TEE sandbox, not the application layer. An attacker with access to application logs cannot retroactively alter timestamps — they're sealed before the TEE exits. This makes the trace usable as legal-grade evidence in regulated industries.
What the agent produced.
The evidence pack includes a redacted output summary — structurally complete, with content removed where it contains information specific to the organization. The structure itself proves the agent performed the task as instructed: it extracted metrics, flagged risks, and produced compliance disclosures.
Revenue: [REDACTED] · EBITDA margin: [REDACTED] · Net income: [REDACTED] · Debt/equity: [REDACTED]
YoY delta: [REDACTED] · 12 line items extracted across 5 financial statements.
3 risk flags identified: [REDACTED — nature disclosed to requestor]. Risk severity: 1× HIGH, 2× MEDIUM. Each flag includes page reference and supporting excerpt.
External auditor opinion: [REDACTED]. 2 findings noted. No material weaknesses identified. Going concern language: absent.
4 regulatory disclosure sections extracted. Jurisdictions: [REDACTED]. All disclosures conform to applicable standards per agent review.
The output summary proves the agent performed its assigned task — not just that it ran. The structure of the output (4 categories, correct fields, risk severity levels) proves the agent behaved as expected under its policy constraints.
The SHA-256 hash of the full, unredacted output is included in the evidence pack. Authorized parties can verify the full output against the hash — the redaction here is for public display only.
Which policies ran. What they found.
ThetaZero enforces your policy set at execution time — not post-hoc. Each check runs inside the TEE sandbox and produces a verifiable result. The evidence pack shows every check, its result, and the timestamp. A policy violation stops the execution and surfaces a finding instead of silently passing.
8 of 8 policy checks passed. The Human Review check shows a HITL review was triggered (high-severity risk flag) and completed before delivery — this is expected behavior, not a failure.
Policy results are cryptographically bound to the execution record. You can't forge a "PASS" result after the fact — the policy evaluation runs inside the TEE and its output is included in the Merkle hash chain.
Most AI platforms run policy checks in application code — software that can be bypassed, patched, or disabled. ThetaZero enforces policies inside the hardware TEE. The policy runtime is sealed into the compute context and verified at boot. It cannot be disabled by anyone, including ThetaZero.
A tamper-evident record of everything that happened.
The audit trail is a series of immutable event records — each one hashed to the previous entry using a SHA-256 Merkle chain. Deleting or altering any entry breaks the chain and is immediately detectable. The audit trail covers the full lifecycle: trigger, execution, review, delivery.
The audit trail is a complete chain-of-custody record. Every event is hashed to the previous one. Auditors — including external auditors with no ThetaZero account — can verify the Merkle root hash independently and confirm no events were inserted, deleted, or reordered.
The human review event (entry 6–7) is part of the immutable audit trail. Auditors can see that a human reviewed the high-severity risk flag, who approved delivery, and when. This satisfies EU AI Act Article 14 human oversight requirements — and it's cryptographically proven, not a checkbox in a spreadsheet.
Cryptographic proof that nothing changed.
The final artifact is the evidence package manifest — a SHA-256 Merkle root hash of all 6 artifacts, signed with ThetaZero's RSA-4096 key. Any alteration to any artifact — including a single character change in any file — produces a completely different hash. This is verifiable by external auditors with no dependency on ThetaZero's live systems.
a5c817b2d4e6f0123456789abcdef0123456789ab
The Merkle root is the "fingerprint of fingerprints" — a single hash that represents all 6 artifacts. If any file is modified — even one byte — the root hash changes. Any auditor can re-run the hash calculation against the downloaded ZIP and confirm it matches.
The RSA-4096 signature ties the hash to ThetaZero's public signing key, which is published and independently verifiable. The signature proves ThetaZero produced this exact package — no forgery, no tampering, no substitution.
No live system dependency: auditors don't need to log in to ThetaZero, call an API, or trust any ThetaZero response. The math is the proof.
Without ThetaZero vs. with ThetaZero.
This is what a compliance review looks like for a team running AI workflows without governance infrastructure.
Which controls this satisfies.
Every field in the evidence pack maps to specific compliance framework controls. These aren't general claims — each field in the evidence pack above satisfies the specific control listed here.
Trust Services Criteria Coverage
The evidence pack above satisfies these SOC 2 controls directly — no supplemental documentation required.
Data Processing Documentation
GDPR Article 35 (DPIA) requires organizations to document how personal data is processed by AI systems. This evidence pack is that documentation.
High-Risk AI System Requirements
Financial review workflows may qualify as high-risk AI under EU AI Act Annex III. This evidence pack satisfies the documentation requirements for high-risk systems.
AI Management System Evidence
ISO/IEC 42001 (AI Management Systems) and NIST AI RMF both require documented evidence of AI governance practices. This pack satisfies both.
Every workflow. Every run. Every proof.
Start free. Run your first governed workflow in under 5 minutes. The evidence pack above is generated automatically — no setup, no configuration, no manual write-up.