Paradigm.
A policy-governed dataset construction and verification instrument. It validates, fingerprints, signs, and documents every dataset it builds. Each claim it makes is checkable with tools Static Signal does not maintain.
Why AI systems fail.
Most failures in production AI trace back to one cause: unreliable data.
Schemas drift. Shards disappear. Remote storage falters. Teams retrain models without knowing exactly which data the model learned from.
The result is unpredictable behaviour and brittle systems.
Paradigm removes this uncertainty by making dataset quality an explicit, enforced stage of the pipeline.
Capabilities.
Validation
- Six tiers: structural checks, subject gate, data quality, perceptual quality, categorical assessment, consistency
- Pre-flight checks on schemas, compute targets, storage, and credentials before GPU time is spent
- Drift detection and corruption checks
- Failed validation blocks the build and records the refusal
Provenance
- SHA-256 content and schema fingerprints
- Branch-aware Merkle trees and hash splits
- C2PA v2.3 content credentials
- SLSA v1.2 build attestations
Evidence & signing
- OMS signatures — Sigstore keyless or Static Signal PKI, with rotation and revocation
- Signed evidence bundle per run: what ran, where, when, on what data
- Append-only audit trail (JSONL)
- Runs tracked with an audited lifecycle: queued, dispatched, running, completed
Integrity
- Drift detection between dataset versions
- Corruption checks on every shard
- Health semantics distinguish accessibility from throughput, healthy from slow, broken from degraded
- Deterministic streaming from the catalogue, not ad-hoc paths
Custody
- mTLS, PKI with revocation
- Role-based access, API keys, audit trails
- Multi-tenant isolation with cross-tenant enforcement
- Raw inputs deleted within 30 days of acceptance; deletion on request within 7 working days, with signed attestation
Compliance export
- Exports EU AI Act Articles 9–15 evidence
- Maps dataset artefacts to technical documentation requirements
- One evidence pack per engagement, versions recorded cryptographically
A single source of truth for data and execution.
The operator console.
What it shows:
Stack.
Security. In production, not on the roadmap:
Authentication and role-based access
Multi-tenant mode (orgs, teams, projects)
API keys and audit trails
mTLS agent trust with PKI
OMS-compatible evidence signing with key rotation
Kill-switch via certificate revocation
GPU compute with CUDA enforcement
C2PA + SLSA fingerprint stack
EU AI Act evidence export
Pre-flight validation
Chain-of-custody lineage graph
Paradigm runs inside COSMOS, a control plane that gates every mutation, keeps a hash-chained event log, and composes attestations for each engagement. ARCHON, a sealed execution engine for heavy compute, is in development. Neither is sold.
This site does not list our release numbers. Every engagement scope pins the exact instrument versions used, and every evidence bundle records them cryptographically. If you need a version claim, it will be in a document we signed.
In development: transparency-log inclusion proofs, erasure attestations, and canary tokens for custody; contamination and poisoning scans with measured false-positive rates for integrity. Expected summer 2026. We ship when it’s real.