LLM Observability
Continuous assurance,
not a periodic audit.
Every governed call writes one row to an immutable ledger. That ledger is the single source of truth for traceability, metering, and compliance evidence — because it is a record of what actually executed, not a reconstruction after the fact.
The Shift
Point-in-time audit describes a system that no longer exists
Traditional AI oversight is commissioned, not continuous. Evidence is gathered by hand, a report is produced, and it lands months later describing a configuration that has since changed.
The gap between what was audited and what is running is where risk accumulates. A quarterly review cannot tell you what happened on a Tuesday afternoon six weeks ago, or whether a policy change last month altered behaviour in a way nobody logged.
A ledger written at execution time removes the gap. There is no reconstruction step, because the record is a byproduct of the call itself. Evidence is not gathered; it already exists.
The Record
What one row contains
Each governed call produces a single ledger entry, written non-blocking so audit never becomes a latency tax on the request path.
Identity and scope
User, persona, workspace, and organization — who invoked what, in which context, under which membership.
Interaction summaries
Input and output summaries, retained according to the organization's content-logging mode: full, fingerprint, or metadata-only.
Model and provider
Which model handled the call and through which provider — the basis for usage control and provider reconciliation.
Token estimate
Consumption for the call, aggregated upward into budgets and cost estimates without being stored separately.
Governance verdict
Whether input validation found threats, whether output validation passed, and whether the call was flagged or blocked.
Content-logging mode used
The mode in effect at write time, so a later policy change never makes historical rows ambiguous.
Derivation
The meter cannot drift from the record
Usage is never stored as its own counter. Every figure on the meter is aggregated from ledger rows for the period.
This is a deliberate architectural constraint rather than an implementation detail. A separate usage counter can diverge from reality through a missed increment, a retry, or a failed write. A derived meter cannot, because there is nothing to diverge from — the number is a query over the record.
The same property makes the meter defensible under scrutiny. When a figure is questioned, the answer is not a reconciliation exercise; it is the set of rows that produced it.
Access
Who can see what, and how it leaves
Row-level access
Users see their own rows through row-level security. Aggregation across an organization happens server-side under an explicit role check.
The traceability feed
Recent governed interactions with a flagged-only filter — timestamp, use case, persona, action, model, tokens, and verdict.
Export
Full audit-trail export as JSON or CSV with fixed column order, optionally scoped since a date. Admin and owner only.
Certified deletion
An owner can permanently delete the trail and receive a tamper-evident SHA-256 certificate covering org, requester, timestamp, row count, scope, and cutoff.
Evidence that already exists when you need it
See the ledger, the traceability feed, and the export path that turns governance into compliance evidence.