You are looking at a read-only demo. Everything here is fictional and the AI is simulated. You can read every page and ask the assistant questions, but nothing can be changed.Sign in with an access code to try the full workflow →
Trust centre

What the AI did, what stayed inside, and how you can check.

Every AI call is listed with the model that made it and what a person did next. The safeguards below are checked against the running system when you open this page, not just described.

16
AI calls, all recorded
0
calls to anything outside the environment
7
cleared by the AI under the rules
7
sent to a person, 0 decided so far
16
AI outputs checked by the Assurance Layer, 0 stopped
0
questions the assistant refused rather than guess

Data boundary

Where it runs
Validated environment (simulated for this demonstration)
Data leaving the environment
None
AI engine
Simulated assistant (simulated-engine v1.0)
Hosting
In-environment. In production: a self-hosted open-weight model inside the customer boundary.

In this demonstration the engine is a placeholder, so there is no external model to call. The production design keeps it that way: regulated data is only ever sent to a model hosted inside the customer's own boundary.

Audit trail integrity

Intact. All 68 records re-verified just now.

Each record carries the fingerprint of the one before it. Changing any record breaks every fingerprint after it, so tampering is detected rather than trusted away.

Download the audit trail (CSV)Browse and re-verify

Safeguards

  • Every AI proposal and answer is checked before it is usedVerified
    A fixed, versioned rule set (v1.0.0) checks each AI proposal and each assistant answer. A failed blocking rule stops it: a proposal goes to a person, an answer is withheld.
    Checked live: the 17 simulated bad outputs on the Assurance page were just run through the rules and 17 of 17 were handled as expected.
  • Audit trail cannot be edited or deletedVerified
    The database itself rejects any UPDATE or DELETE on audit records.
    Checked live: the protecting triggers are installed in the running database.
  • Audit trail is tamper-evidentVerified
    Every record is hash-chained to the one before it, so any edit is detectable.
    Checked live: the whole chain was re-verified just now (68 records).
  • Every change has a recordVerified
    A change and its audit record are saved together or not at all, so nothing changes without being written down first.
    Checked live: every assessment's status was rebuilt from the audit trail alone and compared with the stored status (14 of 14 match).
  • The AI cannot lower the barVerified
    Routing is decided by fixed rules. The AI can add reasons for a human to look, never remove them.
    Covered by automated tests on the routing rules.
  • Roles are enforced by the serverVerified
    Reviewers, the QA head and team members each have separate permissions. Nobody signs off their own escalation.
    Covered by automated tests; identity is simulated until real login.
  • Retries never duplicate a recordVerified
    Every action carries a one-time key, so a repeated click or retried request is recorded once.
    Covered by automated tests.
  • The assistant refuses rather than guessesVerified
    It answers only from live data and approved, effective documents, and cites the section it relied on.
    Covered by automated tests, including questions the library cannot answer.

Recent AI calls

#TimeKindSubjectModelWhat the AI didWhat a person did
68Oct 26, 12:55 PMChatWhy was ASM-0005 flagged?simulated-engine v1.0Answered from the controlled documents (2 citations)—
64Oct 26, 12:55 PMAssessmentASM-0014 — Grade B cleanroom: two non-viable particle excursionssimulated-engine v1.0Proposed grade D4, routed to a human (Level 3)Waiting for the auditor
62Oct 26, 12:55 PMChatWhat needs my review?simulated-engine v1.0Answered from live platform data—
58Oct 26, 12:55 PMAssessmentASM-0013 — Like-for-like HVAC filter replacementsimulated-engine v1.0Proposed grade A1, cleared at Level 1—
53Oct 26, 11:10 AMAssessmentASM-0012 — Cold-chain shipping lane not yet qualified for summersimulated-engine v1.0Proposed grade B2, routed to a human (Level 3)Waiting for the auditor
49Oct 26, 11:10 AMAssessmentASM-0011 — Aseptic gowning requalification current for all staffsimulated-engine v1.0Proposed grade A1, cleared at Level 1—
45Oct 26, 11:10 AMAssessmentASM-0010 — Autoclave IQ executed with one calibration deviationsimulated-engine v1.0Proposed grade C3, routed to a human (Level 3)Waiting for the auditor
41Oct 26, 11:10 AMAssessmentASM-0009 — Batch record missing a second-person verificationsimulated-engine v1.0Proposed grade C3, routed to a human (Level 3)Waiting for the auditor
37Oct 26, 11:10 AMAssessmentASM-0008 — Stability study 23-114, 12-month timepoint within limitssimulated-engine v1.0Proposed grade B2, cleared at Level 2—
33Oct 26, 11:10 AMAssessmentASM-0007 — CAPA-2026-031 effectiveness check closedsimulated-engine v1.0Proposed grade B2, cleared at Level 2—
28Oct 26, 11:10 AMAssessmentASM-0006 — New excipient supplier introduced without qualification on filesimulated-engine v1.0Proposed grade D4, routed to a human (Level 3)Waiting on the team
24Oct 26, 11:10 AMAssessmentASM-0005 — Packaging line 2 label reconciliation variancesimulated-engine v1.0Proposed grade D4, routed to a human (Level 3)Waiting for the auditor
20Oct 26, 11:10 AMAssessmentASM-0004 — GMP training current for packaging line 2 operatorssimulated-engine v1.0Proposed grade B2, cleared at Level 2—
16Oct 26, 11:10 AMAssessmentASM-0003 — Batch 24-0871 assay results within specificationsimulated-engine v1.0Proposed grade A1, cleared at Level 1—
12Oct 26, 11:10 AMAssessmentASM-0002 — SOP-MFG-014 rev 7 periodic review completed on timesimulated-engine v1.0Proposed grade A1, cleared at Level 1—
8Oct 26, 11:10 AMAssessmentASM-0001 — Temperature excursion in cold storage room 3simulated-engine v1.0Proposed grade C3, routed to a human (Level 3)Waiting for the auditor