Beta

Verified AI Runs

Last updated 2026-10-29

What it is

A signed, timestamped receipt for an AI model run. You tell us (or we observe) which model ran, on what prompt, with what parameters and what it returned. We hash the prompt and output, run the deterministic checks you attach, sign the result with a2ally's Ed25519 issuer key and write it to the public ledger, whose roots are anchored to Bitcoin. Prompts and outputs are never stored; only their SHA-256 hashes are.

What it proves, and what it does not

• Proves: the exact prompt, output and parameters (by hash), the model and provider named, when it was recorded, and which checks passed.
• Proves in agreement mode: a2ally itself sent the same prompt to several models from different providers and recorded each output hash and how far they agreed.
• Does not prove: that an output is true, safe or correct. Several models agreeing makes an answer more likely to be stable, not right.
• Does not prove in recorded mode: that the named model really produced the output. The caller reports it; the receipt says reported_by: caller.
• a2ally never uses an AI run to decide a grade, score or award. Those stay deterministic (https://www.a2ally.com/v1/judgement-provenance).

Deterministic checks you can attach

• json_schema: Output parses as JSON and matches {schema} (subset: type, required, properties, additionalProperties:false, enum, const, items, min/max length/items/value, pattern).
• regex: Output matches {pattern} (with optional {flags} from i, m, s, u). Set {must_match:false} to require no match.
• no_secrets: Output contains none of the common credential shapes (cloud keys, private key blocks, API tokens, JWTs).
• length: Output length in characters is within {min} and {max}.
• contains: Output contains every string in {all} and none in {none} (case-sensitive).

Checks are plain code, so anyone can re-run them on the same output and get the same result. Each check's exact spec is hashed into the receipt.

Cross-model agreement

POST https://www.a2ally.com/v1/ai-runs/agreement with {prompt, n} (2 to 5, default 3) or {prompt, models}. a2ally calls each model at temperature 0 through one router, records every output hash, and computes an agreement score with a published formula: round(100 x max(exact agreement, mean pairwise token overlap)). Outputs come back to you once and are not stored. Default models: openai/gpt-4o-mini, anthropic/claude-3.5-haiku, google/gemini-2.0-flash-001. Status right now: not enabled yet; recorded mode works today.

Use it

• Record: POST https://www.a2ally.com/v1/ai-runs with your API key and {model, provider, prompt | prompt_sha256, output | output_sha256, params, ran_at, checks}
• Fetch: GET https://www.a2ally.com/v1/ai-runs/{id}
• Verify (free, keyless): POST https://www.a2ally.com/v1/ai-runs/verify with {receipt, signature}, optionally the prompt or output text
• MCP: record_ai_run and verify_ai_run (https://www.a2ally.com/mcp?toolset=ai_runs)
• Offline: check the Ed25519 signature over signed_bytes with the key at https://www.a2ally.com/.well-known/a2ally-keys.json, then the ledger inclusion with the open verifier at https://www.a2ally.com/witness/verify

Where it helps

• Disputes: show exactly what an agent's model said before it acted.
• Compliance and audits: a tamper-evident log of model use, without handing over the prompts.
• Agent-to-agent deals: attach a receipt so the other side can check the output was screened for secrets or matched an agreed schema.
• Code review: record an AI reviewer's findings next to a blind audit receipt, as advice, while the decision stays with people and published rules.

Price

Verification is always free. Each signed receipt is planned at $0.01, with 100 free per month per account; that price line is announced at https://www.a2ally.com/transparency/changes and joins the pricing rulebook only after the 7-day notice. a2ally credits have no cash value during the preview, and nothing is charged today.