Connect a system
Point Perpendis at the system you want covered. Nothing sensitive leaves your environment until you choose how the harness runs.
Support Agent v2.4
- Base model
- Qwen2.5-7B-Instruct fine-tune
- Use case
- Customer-support agent, tool calls
- Target
- AIUC-1 pack Annex IV mapping — planned
Coverage goal
- Application
- AI-liability policy (Lloyd's paper)
- Workspace
- Shared with underwriter after compile
Harness deployment
Evidence sources
Connector states below are illustrative — connectors are not built.
| Model endpoint | api.acme.ai/v2/agent | Illustrative |
| Eval samples | s3://acme-evals/support-v24 | Illustrative |
| Cloud config | read-only | Illustrative |
| Policy documents | 4 of 7 | |
| Vendor contracts | Not connected |
What gets collected, and what never leaves your environment
Automated domains call your endpoint or run inside your VPC; only results and signed attestations reach Perpendis. Documentary evidence (policies, contracts) is uploaded by your team and flagged into remediation tasks when gaps exist. Weights are never uploaded.
Assessment run #R-0847
Seven sample domains plus one live domain: computational integrity runs the actual Perpendis referee — compiled to WebAssembly, executing in your browser right now.
Referee result live engine
Divergence detail — the exact disputed operation
- Location
- —
- Reference (SPEC)
- —
- Claimed (B)
- —
- Naive trace check
- —
- Witness roots
- —
The referee bisected to a single arithmetic operation, re-executed it from Merkle-bound inputs, and compared at tolerance 0 (bit identity).
Signed receipt — real Ed25519, verify or tamper with it
In plain terms: this proves the run wasn't edited after the fact — and you can check that right here, without trusting us.
—
That verdict was produced here, not fetched: one operator (RMSNorm) on a real Qwen2.5-7B norm tensor, tolerance zero. It is the one thing we measure today.
Pointing it at your own operators, your hardware matrix or your CI is contract work.
Tell us what divergedDocumentary workflow (concept) — 11 of 15 complete
Owner & governance ✓ · Incident-response plans ✓ · Oversight gates ✓ · Training-data provenance gap · Vendor AI addendum gap · Supply-chain attestation missing
Gaps are recorded on the compiled pack. Task assignment and pack deltas are not built. Keeping evidence current between renewals is the intended subscription — that loop is designed, not shipped.
Evidence pack EP-2026-0847
Read-only relying-party workspace · point-in-time, model-version-pinned · free for underwriters, auditors and reinsurers.
Concept demo. Five of the seven domains below are fixed sample figures — they were not measured on any system. Only computational integrity is a live measurement, and governance is documentary.
| Domain | Score | Method | Status |
|---|---|---|---|
| Performance | 94 sample | Not assessed | Not assessed |
| Hallucination | 91 sample | Not assessed | Not assessed |
| Robustness | 88 sample | Not assessed | Not assessed |
| Adversarial | 82 sample | Not assessed | Not assessed |
| Bias & fairness | 90 sample | Not assessed | Not assessed |
| Comp. integrity | — | Bit-exact re-verification | Sealed · live |
| Governance | — | Documentary | 1 gap |
Verify the attestation yourself
Proves the assessment run wasn't edited after the fact. Real cryptography, running here — no trust in Perpendis required.
Technical detail
Ed25519 over domain-tagged canonical JSON. The signature covers exactly these fields and nothing else: the tensor under test (a fixed weight fixture that ships inside the engine, not the assessed system's own weights), the two execution targets compared, the specification version, which target the referee found correct, and — when they diverge — the first divergence: row, reduction step, fault class, operand, and the accumulator before, the reference result after and the claimed result after.
Everything else on this page is outside the signature: the witness commitment roots, the query counts and bond, the timings, and the seven sample domain scores. The seal does not attest to hardware, to which samples were used, or to who ran the engine — the second target is a scenario string the assessed party selected before the run. Run the fabricated-trace scenario first if the receipt is empty.
Reliance letter
- Relying party
- Meridian Specialty
- Purpose
- Policy AI-2026-114 underwriting only
- Liability cap
- Assessment fee / PI limits
- Scope
- Point-in-time · v2.4 only
- Fee
- $2–5K one-time · paid by applicant
A litigable instrument with a named relying party and a real cap — not a marketing PDF.
Underwriting decision
Accepting this pack replaces your internal technical assessment for this submission. Evidence verified, capped, version-pinned above.
Concept demo. Connectors, task assignment and e-signature are not built. All scores except computational integrity are illustrative and were not measured on any system; everything tagged LIVE is the real Perpendis referee (100-test Rust crate, bit-identical native ↔ wasm) executing in this page: real witnesses over real Qwen2.5-7B RMSNorm weights, real bisection, real Ed25519 receipts.