QuantovaQstampby Quantova
EN
EnglishEN日本語JAFilipinoFIL한국어KO简体中文ZHFrançaisFREspañolESУкраїнськаUK
GitHub Start building →
Qstamp / SI agents
Primary audience

Audit trails and accountability for superintelligence (SI) agents.

Agents now act on behalf of people and institutions. Qstamp gives every action a durable, verifiable record of what was done, by which system and under which model and policy, and the latest time by which it was done. The evidence survives the transition to quantum computing and does not depend on trusting the operator that produced it.

AGENT FLEET OPERATOR EVIDENCE SERVICE · QSTAMP SDK QVM · QUANTA CONTRACTS Payment agentapprovals · transfers Research agenttool calls · retrievals Customer agentmessages · outputs Model registryweights · data manifests Policy enginelimits · mandates · versions Capturecanonical JSON of each action · SHA3 digest Batch windowevery N seconds or M actions · salted leaves Stampone commitment per batch · kind ai_agent_action Receipt storereceipt beside each action log entry QStamp contractopen anchoring · caller bound QStampIssuerissuer signed orders · revocation QStampCouncilthree of five approvals FinalityML DSA 65 committee · about 0.2 s AGENT FLEET OPERATOR EVIDENCE SERVICE · QSTAMP SDK QVM · QUANTA CONTRACTS Payment agentapprovals · transfers Research agenttool calls · retrievals Customer agentmessages · outputs Model registryweights · data manifests Policy enginelimits · mandates · versions Capturecanonical JSON of each action · SHA3 digest Batch windowevery N seconds or M actions · salted leaves Stampone commitment per batch · kind ai_agent_action Receipt storereceipt beside each action log entry QStamp contractopen anchoring · caller bound QStampIssuerissuer signed orders · revocation QStampCouncilthree of five approvals FinalityML DSA 65 committee · about 0.2 s
Figure 4 · Agent evidence architecture. Actions are captured and batched inside the operator boundary. Institutions that require authorised issuers or multi party approval use the issuer or council templates.
What to stamp

The evidence an agent should leave behind.

kind 6 · ai_agent_action

Actions and tool calls

The canonical record of each tool invocation, approval or transfer, including inputs, outputs and the identity of the acting agent.

kind 4 · ai_model

Model passports

The digest of every released model artefact, so the exact version behind any later decision can be established.

kind 5 · ai_dataset

Dataset manifests

Training and evaluation data manifests fixed at the moment of collection, supporting provenance and licensing claims.

kind 7 · ai_output

Generated content

Text, images and audio with their provenance labels, so an original can be distinguished from an altered copy.

kind 0 · record

Policies and mandates

The version of every spending limit, tool permission and approval rule in force when an agent acted.

kind 0 · record

Evaluation results

Safety and performance evaluations fixed before release, so results cannot be revised after the fact.

Integration patterns

Match the cadence of the agent.

Interval batching

Collect actions for a fixed window, such as every ten seconds, and anchor them in one transaction. This suits high volume agents and keeps the cost per action negligible.

Decision checkpoints

Stamp immediately before or after material actions such as payments above a threshold, so the record is final before the action completes.

Release gates

Stamp model weights, data manifests and evaluation reports as part of the release pipeline, producing a model passport per version.

Institutional issuance

Deploy the issuer template under the institution's key so that only authorised issuers can anchor, with revocation for withdrawn records.

Reference implementation

An evidence service in a few lines.

evidence-service.jsinterval batching
const qstamp = require('@quantovainc/qstamp'); const queue = []; function capture(action) { const bytes = Buffer.from(JSON.stringify(action)); queue.push({ action, digest: qstamp.digestBytes(bytes) }); } setInterval(async () => { if (!queue.length) return; const batch = queue.splice(0, queue.length); const receipts = await qstamp.stamp({ seed, index: 0, kind: 'ai_agent_action', records: batch.map((b) => ({ digest: b.digest })), onPending: savePendingSafely, }); receipts.forEach((r, i) => store(batch[i].action, r)); }, 10000);

The onPending hook persists each submitted batch before the SDK waits for finality, so no receipt is lost if the process stops.

Why this matters

Logging obligations that agent operators already face.

RequirementSourceHow Qstamp supports it
Automatic logging of eventsEU AI Act Article 12Every logged event can be anchored, making later alteration of the log detectable.
Retention of logs and documentationEU AI Act Articles 18, 19 and 26Receipts remain verifiable for the full retention period, including periods that extend past the expected arrival of quantum computers.
Marking of generated contentEU AI Act Article 50 · California AI Transparency ActProvenance labels and originals can be anchored, so stripped or altered labels can be detected.
Records of algorithmic decisionsColorado AI law · financial supervisionDecision records and the model version used are fixed at the time of the decision.
Governance and traceabilityNIST AI RMF · ISO IEC 42001Provides the traceability evidence that governance frameworks ask operators to keep.

Qstamp supplies technical evidence. Operators remain responsible for their wider compliance programme. See the compliance page and the Responsible SI Policy.