The trust layer

Proof, at every layer.

“Trustworthy” is easy to write and hard to show. This page lists what we can demonstrate — the mechanism behind each claim, and the product it lives in.

01

Model verification

TokenSeq

Why. An upstream can swap the model, lower the tier, or serve from a pool of consumer accounts. The request returns 200. Nothing at the protocol level tells you.

How.

  • Blind capability testsScience, mathematics, code, instruction following, tool use, abstract reasoning; fixed seeds, identical questions to every channel; public evaluation sets; wrong answers count as wrong, never as “too small a sample”.
  • Upstream provenanceOfficial direct, cloud-hosted, relay, or consumer subscription pool — identified by independent probes, not by what the upstream declares.
  • Independent axes of evidenceRelay, downgrade and origin are judged separately; weak evidence cannot launder a strong finding. Not tested ≠ tested clean. Where no conclusion can be reached, the report says so.
  • Difficulty that never saturatesWhen a frontier model scores full marks, the set is too easy; it is hardened until the strongest model still has headroom.
  • Scored and acted onA weighted score across capability, performance, reliability and verification; untested dimensions are removed rather than assumed. Channels that fail scheduled inspection are taken out of service automatically.
THE PROBLEM What an upstream can do silently Swap the model behind the name Serve a lower tier or a quantised build Route through a relay you never agreed to Serve from a pool of consumer subscriptions Under-report or over-report token usage HTTP 200 · the response tells you nothing THE CHECKS · RUN CONTINUOUSLY Blind capability tests Same questions, fixed seed, every channel. Hardened until the best model still has headroom. Upstream provenance Official · cloud-hosted · relay · consumer pool — identified by our probes, not by what it declares. Behaviour & usage fingerprints Latency and throughput baselines per model; token-count patterns; refusal and style. THE VERDICT Independent axes of evidence Relay, downgrade and origin are judged separately. Weak evidence cannot launder a strong finding. Three possible outcomes Verified — evidence on every axis Flagged — a finding that is acted on No conclusion — reported as such not tested ≠ tested clean THE ACTION Scored, scheduled, enforced A weighted score across capability, performance, reliability and verification. Untested dimensions are removed, never assumed. onboard evaluate list inspect scheduled · repeats A channel that fails inspection is taken out of service automatically — and can be restored automatically when it recovers. every step recorded · one path for agents and people WHAT A REPORT LOOKS LIKE · illustrative Verification report channel upstream-02 · model frontier-text-01 · 17 checks · run on request ● verified UPSTREAM IDENTIFIED AS Cloud-hosted deployment Consistent with the contracted source. Probes agree on all axes. origin ✓ relay ✓ downgrade ✓ SCORES Channel quality99 Performance88 Model capability93 untested dimensions removed before scoring PER-ITEM mathematics12 / 12✓ code11 / 12✓ instruction following12 / 12✓ tool use10 / 12✓ abstract reasoning11 / 12✓ latency baselinewithin band✓ each item carries its actual cost · wrong = wrong Report content and figures are illustrative. Real reports are generated per channel and per run. key used once · not stored · not logged
Verification report · redrawn · upstream identified as cloud-hosted; quality 99 / performance 88 / capability 93 · illustrative
02

Data provenance

DataSeq

Every knowledge unit carries its origin — page and block coordinates for documents, timestamps for audio and video. When an agent answers, the source can be opened at exactly that place.

03

Retention control

TokenSeq

  • Text: zero storage.Pass-through in memory, released on completion. Metadata only — timestamp, model, token counts, latency. Auditable by a third party of your choosing; written into the service agreement.
  • Generated media.Stored once, in our own store, for delivery. Retention defaults to a fixed period you can change; the expiry is written at storage time and enforced — an expired asset is reported as expired, never as “not found”.
04

Metering integrity

TokenSeq

Quote → deduct → then call upstream. One generation, one ledger line. Prices are snapshotted into the line, so a later price change never alters history. A failed generation is refunded in full. The price shown is the price charged.

05

Governance & audit

AgentSeq

Every task records who started it, under which organisation and authorisation, which capabilities it called, what it consumed, and how it ended. Governance objects — people and organisation, enterprise capabilities, knowledge and agents, resources and security — each have configuration, authorisation, observation and audit views.

Service levels

Tiered SLA, committed in contract.