Verification and trust boundaries
Purpose
This page describes how SIGMA combines evidence, review, economics, and publication into a coherent verification model: what is deterministic, what is model-mediated, how certificates are anchored, and how deployments differ. It is written as system architecture, not as commentary on external reviews.
Evidence before and during review
- Sandbox and heuristic analysis produce structured findings on submissions before validator rounds complete. These feed the same audit process as model outputs.
- Declared-surface checks (especially for API-shaped work) yield signals about endpoints, coverage, and fit between declarations and observed behavior.
- Two-phase certificate assemblies and, when applicable, monitoring assemblies ensure multiple validators contribute before a round closes.
Together, these layers mean issuance is driven by assembly outcomes that integrate non-model evidence and consensus policy, not by a single isolated completion.
Consensus and model-mediated review
Each assigned validator produces a verdict through a managed audit model in a supervised pipeline, augmented by validator plugin XML where slotted (see Validator plugin concepts).
The consensus layer evaluates phase verdicts together with sandbox source signals (for example injection and manifest-mismatch indicators). Split SAFE/UNSAFE patterns and high-risk combinations route to escalation rather than silent certification.
Approval follows the round’s assembly outcome, which incorporates those model verdicts. SIGMA does not present itself as a complete formal proof of arbitrary programs; probabilistic review remains an explicit part of the design.
Validator plugins in the trust model
SIGMA validator plugins are typed, slot-bound, screened, versioned, and marketplace-governed. They are composed into a fixed audit XML envelope so reviews stay structured.
Usage, eligibility, and governance bonuses may tie plugin economics to validator identity where policy enables it.
Today, plugins ship as structured advisory content composed into audit context. They are not a general isolated executable plugin runtime for arbitrary validator-supplied code. The product’s leverage is distribution, screening, slot governance, and consistent packaging of reviewer knowledge.
Certificates and public verification
- A successful approval seals a certificate snapshot as the durable product artifact.
- The protocol records a cryptographic hash of a canonical certificate payload, binding verification to that snapshot (algorithm metadata is available on verify surfaces).
- Commit history allows multiple issuance generations without overwriting prior records.
- On deployments where it is enabled, a certificate registry on the target network records on-chain commitments linked to the payload hash. Certificate and verify experiences expose the hash, commitment lifecycle, and explorer reference when a commitment transaction has settled.
Consumer verification should rely on published verify APIs, payload hashes, and registry posture for the specific deployment under evaluation. A compact on-chain commitment does not, by itself, bundle every intermediate artifact of every historical audit for off-the-shelf replay by arbitrary third parties without supported tooling.
Economics and validator standing
- Validators meet stake expectations (for example USDC and token stake where enabled) that feed governance weight.
- After scored rounds, consensus alignment updates an audit-accuracy ledger; misalignment lowers the derived accuracy signal and therefore selection weight and odds.
- SIGMA reputation is a longer-horizon input alongside accuracy in governance composition.
- Round payouts and monitoring assemblies allocate value per protocol rules.
Magnitude of adjustments, floors, and handling of escalation rounds are deployment-tuned. The core product does not depend on external prediction markets or third-party slash oracles.
Operator posture and on-chain anchors
- Operator-maintained catalogs, pricing, and escalation paths are part of the product definition.
- On-chain modules—network- and release-dependent—may include treasuries, auditor registries, marketplace integration, and a certificate registry for commitments.
- Validator weights combine stake, derived accuracy, model tier, SIGMA reputation, and related signals; deployments may mirror these for wallets and explorers.
Evaluators should treat SIGMA as progressive trust infrastructure: reliance on concrete network, contract, and verify behavior for the environment in question.
End-to-end pipeline (summary)
- Intake and sandbox — structural and heuristic evidence.
- Declaration and behavior checks — especially for APIs.
- Two-phase assembly — managed-model reviews under consensus policy.
- Outcome — approve, reject, escalate, or monitoring assembly branch.
- Certificate — sealed snapshot, canonical payload hash, optional registry commitment, public verify as the consumer surface.