Guide · By the Ingenious AI team · Published 14 November 2026 · Reviewed 14 November 2026
Importing CompanionGuard Evidence Into an Audit Room
Short answer
CompanionGuard evidence — the safety and guardrail record for companion-style systems — enters the room like every suite source: export from CompanionGuard, import unchanged, and the room seals it at intake and badges it 'CompanionGuard'. File on the rhythm the safety work itself follows, and treat configuration changes as first-class events: when guardrails change, the exports before and after the change are both evidence, and the room's thread should say when the change happened. Safety evidence persuades by continuity; a room that shows an unbroken, honestly-versioned safety record is the strongest exhibit a deployer can file.
What is safety evidence?
Safety evidence is the guardrail and safety record produced by CompanionGuard for companion-style AI systems and imported into AI Audit Exchange rooms: imported read-only and unchanged, SHA-256 sealed at intake, badged with the CompanionGuard source label in the manifest, and read by auditors as the system's own account of how it was constrained and monitored during the audit period.
What safety evidence has to show
A companion-style system's audit turns on whether its guardrails were real during the period: configured, active and monitored — not merely designed. CompanionGuard's exports are the system's own testimony on those points, produced in the ordinary course of operation. Imported into the room unchanged, they let an auditor test the safety story against the safety record rather than against a slide deck about the safety record. That distinction is the whole ballgame for safety-flavoured evidence, because safety claims are easy to assert and hard to fake retrospectively across a year of system output. A monthly-imported run of CompanionGuard artifacts, sealed as produced, is what 'the guardrails were on' looks like when it is true and provable.
Cadence: follow the safety work, file monthly
Safety work has its own tempo — reviews, evaluations, configuration adjustments — and the import rhythm should be monthly at minimum so the manifest draws that tempo honestly. Where a safety evaluation cycle completes, its export belongs in the room that month, not in a quarterly bundle. Bundling has a specific cost for safety evidence: it blurs the sequence. An auditor reconstructing 'what did the deployer know, and when?' reads import dates as part of the evidence — a record filed in the month it was produced testifies to a living process; the same record filed in December testifies to a filing project. The difference is visible in one column of the manifest, and safety examiners are trained to look at that column early.
Configuration changes are evidence events
Guardrails change: thresholds adjusted, categories added, escalation paths revised. In the room, a configuration change creates a natural before-and-after in the CompanionGuard record — and both sides of the change are evidence. File the exports that document the state before, the change itself (as the product records it, or as a dated direct upload where the change decision was organisational), and the state after. Narrate the change once in the room's thread: what changed, when, why, and which artifacts bound the change. Suppression is the failure mode here — a record showing only the final configuration invites the auditor to wonder what the earlier one was hiding, usually at the least convenient moment. Versioned honestly, changes read as governance. The pattern is the platform-wide one in evidence version control.
Freshness and the safety record
Validation's freshness rule — artifacts recorded more than 365 days before the run flagged stale — has a particular bite for safety evidence. A safety posture is a claim about the present system; last year's evaluation of a since-changed configuration testifies weakly about today's guardrails. Work CompanionGuard flags by filing the current configuration's evidence: the latest evaluation cycle, the current-state exports. Old safety records stay — they are the period's history — but the room should always contain safety evidence young enough to describe the system as it now stands, especially at the moment an auditor arrives. A room whose newest safety artifact is 14 months old is telling a story no commentary can out-argue.
Reading safety evidence in the two regimes
In an SB 1119 room, CompanionGuard artifacts form the safety chapter of the system's accountability record, sampled and read alongside decision and provenance evidence. In a CPPA room, the same exports can speak to the security posture of a system handling sensitive interactions — a different reading of the same sealed facts. File the exports into both rooms where both examinations apply, each room sealing its own copy. And where the organisation runs no companion-style system, the source is simply absent from the manifest — honestly, with the scope noted in the thread if questions could arise. Absence explained is scope; absence disguised is a finding waiting for its moment.
The rehearsal questions for safety evidence
Drill these in your readiness review, in the room's thread, before the auditor asks them in earnest. Which artifact evidences the guardrail configuration as of a chosen mid-period date? Where is the record of the period's most significant safety configuration change, and does the thread narrate it? What is the newest safety artifact, and how old is it? Can five sampled CompanionGuard seals be verified in five minutes? A room that answers all four from its own surfaces is a room whose safety story will survive examination. A room that cannot has been given, by friends, the exact worklist its next month should clear.
Five actions to take
- Import CompanionGuard exports monthly, untouched, matched to regime and period.
- Treat every guardrail configuration change as a filing event: before, change, after — plus one thread narration.
- Keep the newest safety evidence young: file each evaluation cycle as it completes, never in year-end bundles.
- File the same exports into both regime rooms where both examinations read safety.
- Drill the four rehearsal questions quarterly and convert failures into the next month's filing tasks.
Worked example
A fictional companionship-app firm, Willow Companion Labs, files CompanionGuard exports monthly into its SB 1119 room: 26 artifacts across the period, including three that bound a significant October guardrail revision — the last pre-change evaluation, the change record as a dated direct upload with the decision rationale, and the first post-change evaluation. A thread note of four lines ties the three together with dates.
The auditor's safety sampling draws 7 artifacts — 7 seals verified — and lands, as samplers do, on the revision. The room answers without a meeting: both configurations' evidence present, the change narrated contemporaneously, post-change evaluations running monthly since. Her note characterises the revision as documented governance. The same evidence, filed as only the final configuration in a December bundle, would have read as a gap with a story attached. Filing rhythm and one honest note made the entire difference.
Frequently asked questions
Does AI Audit Exchange evaluate our companion system's safety?
No. Safety evidence is produced by CompanionGuard and imported unchanged. The Exchange hosts, seals and badges it; the evaluation of what it shows belongs to your auditor.
Our guardrail config changed mid-period. Which version do we file?
Both, with the change itself documented and narrated in the thread. Before-and-after evidence with an honest change record reads as governance; a final-state-only record reads as a question.
How fresh must safety evidence be?
Validation flags artifacts recorded more than 365 days before the run. For safety, the operative standard is stricter in spirit: the room should always hold evidence describing the current configuration, filed as each evaluation cycle completes.
Can safety evidence serve both regimes?
Yes — import the same exports into each room separately. SB 1119 reads them as accountability evidence; a CPPA examination may read the same sealed facts as security posture.
What if we do not run a companion-style system?
Then CompanionGuard evidence is honestly absent. Note the scope in the room's thread if the absence could be misread, and file the sources you do operate completely.
Keep reading
All AI Audit Exchange guides · Auditor directory · About AI Audit Exchange · Security · Pricing · AI Audit Exchange home · Previous: Importing ReviewDesk — ADMT Decision Review Evidence Into an Audit Room · Next: Importing CRAClock Evidence Into an Audit Room
How to Answer Auditor Evidence Requests Faster · Common AI Audit Evidence Mistakes (and Their Fixes) · AI Audit Evidence for Small Companies: The Minimum Viable Discipline