product · preflight

Preflight

Harness posture, scored from the tape. Eleven components, a published method, and the seven questions a board asks — answered with evidence.

In September 2026 the Australian Signals Directorate published Agentic AI harnesses: the layer above the model: govern the software around the model, not the model. It names eleven harness components and seven questions a board should ask. Preflight scores your harness against those components from the events Wytness already records, and answers the questions one by one. GRC tools answer with self-attestations; Preflight answers with runtime evidence. A preflight check is a walk around the aircraft before every flight, not once — this one refreshes continuously.

what you see

Score and coverage, side by side.

app.wytness.ai · preflight
eleven components · three classesnineteen checks, each with its evidence ✓below 50% coverage the score goes grey

the frame

Eleven components. Three classes. Never blended.

LLM — the brainpredicts text · stateless · a replaceable partACTS THROUGHHARNESS — the body · eleven componentsUser interfacePrompt & policylayerContext managerModel interfaceTool registryPermission systemExecutionenvironmentConnector layerMemory &session storeAudit &observabilityUpdate &supply chainTHE RECORD LEAVES THE BODYWytness — the record, held outside the harnesssigned with your key · personal data sealed · verifiable without us
the harness emits; the record leaves the bodyevery observed check reads that record ✓

observed

Computed from the tape. The only class inside the score; every check carries the query behind it and links to the rows.

attested

You said so, with evidence and an expiry. Shown beside the score. Never counted.

unknown

Could not be evaluated: no data, an unclassified tape, an expired attestation. Never a pass — it lowers coverage.

from the tape · no new data

What the tape answers today.

accountability

Agents approved, with an operator, risk-classified, chain-verified in the last 30 days.

privilege

High-impact actions that carried an approval; capability drift; tools and permissions within each agent's declaration (one click: Approve the observed set).

integrity

Registered agents recording, envelope signatures verified, open chain breaks, anomalies resolved in time.

cost

Agents budgeted, budgets held, no runaway loops.

not yet

Blind spots read not assessed without the Collector. “Would have waited” is informational until Shield enforcement ships (Q1 2027).

Privilege and accountability checks weigh two, everything else one. Unknown leaves both sums. The full method, with every threshold, is published inside the product.

for the board

Seven questions.

The For the board tab answers each from the checks that bear on it, and says yours where none do.

  1. 1What business outcome is this agent for, and why an agentic approach?yours to answer
  2. 2What data, systems and tools can the agent access?partly
  3. 3What actions can it perform, and which need human approval?partly
  4. 4How is prompt injection mitigated?partly
  5. 5Can all significant decisions, tool invocations and actions be monitored and audited?the tape answers
  6. 6How will we know it is delivering value, operating safely and within risk tolerance?partly
  7. 7If the harness were compromised, what is the worst outcome?yours to answer

the honest part

What Preflight does not do.

the number

Only observed checks are in the score. Attested components are shown as attested and never counted.

unknown

Unknown is never a pass. An unclassified tape lowers coverage; it does not raise posture.

scope

Preflight measures the instrumented estate. Without the Collector, agents that never reached the tape are not assessed, and the report says so.

what it is not

Preflight does not enforce, predict or certify. The score summarises checks against a published method version and is reproducible from its snapshot.

approvals

Approvals are recorded as the harness reported them through AGT's approval primitive. Until Shield enforcement ships, “would have waited” counts are informational.

Not a GRC suite. No runtime blocking — that is Shield. No model quality, drift or SLA measurement. No injection detection product. Never a single number that mixes what was observed with what was declared.

availability

What each tier gets

 StarterGrowthBusinessEnterprise
Preflight page · score · checks · For the board · Attestations
Preflight report (JSON + sealed file) · rides Evidence Packs

Walk around the aircraft. Every day.

Preflight scores the tape you already keep, against a method you can read.

Questions about Preflight? Ask us.

the other surfaces: Ledger · Registry · Shield

We set no cookies. Sign-in and preferences use essential first-party browser storage only — no tracking, advertising, or third-party analytics. Privacy Policy