Connect a system
Nothing sensitive leaves your environment.
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
- Workspace
- Shared with underwriter
Harness deployment
Evidence sources
Connectors are not built. States below are illustrative.
| 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
Automated domains call your endpoint or run in your VPC; only results and signed attestations reach Perpendis. Policies and contracts, your team uploads. Weights, never.
Assessment run #R-0847
Seven sample domains, one live: computational integrity runs the real Perpendis referee in your browser.
Referee result live engine
Divergence detail — the disputed operation
- Location
- —
- Reference (SPEC)
- —
- Claimed (B)
- —
- Naive check
- —
- Witness roots
- —
Bisected to one operation, re-executed from Merkle-bound inputs, compared at tolerance 0.
Signed receipt — Ed25519, verify or tamper with it
Proves the run wasn't edited afterwards. Check it yourself, here.
—
Produced here, not fetched: one operator (RMSNorm), a real Qwen2.5-7B norm tensor, tolerance zero. The one thing we measure today.
Your operators, your hardware, your CI: custom work.
Talk to us about custom workDocumentary workflow (concept) — 11 of 15 complete
Training-data provenance gap · Vendor AI addendum gap · Supply-chain attestation missing
Gaps are recorded on the pack. Task assignment and pack deltas are not built.
Evidence pack EP-2026-0847
Read-only · point-in-time, model-version-pinned.
Concept demo. Five of the seven domains below are sample figures, not measured on any system. Only computational integrity is a live measurement; governance is documentary.
| Domain | Score | Method | Status |
|---|---|---|---|
| Performance | 94 sample | — | Not assessed |
| Hallucination | 91 sample | — | Not assessed |
| Robustness | 88 sample | — | Not assessed |
| Adversarial | 82 sample | — | Not assessed |
| Bias & fairness | 90 sample | — | Not assessed |
| Comp. integrity | — | Bit-exact re-verification | Sealed · live |
| Governance | — | Documentary | 1 gap |
Verify the attestation yourself
Real cryptography, running here — no trust in Perpendis required.
Technical detail
Ed25519 over domain-tagged canonical JSON. It covers these fields and nothing else: the tensor under test (a fixed fixture inside the engine, not the assessed system's own weights), the two targets compared, the spec version, which target the referee found correct, and — on divergence — row, reduction step, fault class, operand, accumulator before, reference and claimed after.
A valid signature is not a pass. It covers those receipt fields only. Outside it: witness roots, query counts, bond, timings, and the seven sample domain scores. It attests nothing about hardware, which samples ran, or who ran the engine — the second target is a scenario string the assessed party picked.
Reliance letter
- Relying party
- Meridian Specialty
- Purpose
- Policy AI-2026-114 underwriting only
- Liability cap
- Professional indemnity limits
- Scope
- Point-in-time · v2.4 only
Underwriting decision
Concept demo. Connectors, task assignment and e-signature are not built. All scores except computational integrity are sample data, not measured on any system. Everything tagged LIVE is the real Perpendis referee running in this page on real Qwen2.5-7B RMSNorm weights.