Writes carry sources
An entry without provenance is an opinion. Nothing enters durable memory without naming where it came from, so any claim can be walked back to its origin when it matters.
Services
AI Governance, Risk & Compliance
Governance is an operating cadence, not a one-time compliance event.
Most AI governance failures are practical, not philosophical: orphaned model records, missing approvals, monitoring that starts after go-live, changes that never get re-reviewed. Policy documents do not catch these. Operating routines do.
We install a governance operating model sized to your risk: one accountable owner per use case, a central inventory, review paths proportional to impact, and recurring cadences that keep evidence current. The result stands up to auditors, regulators, and your own board.
Governance is how the top layer of the Process-to-AI Pyramid stays trustworthy: controls proportional to autonomy and impact, applied before, during, and after deployment.
What we deliver
- AI governance framework design
- Use-case risk tiering and inventory
- Human-in-the-loop policy design: which action classes always require a person
- EU AI Act and NIST AI RMF alignment
- Standalone AI governance audits and responsible-AI reviews
- Model and agent lifecycle controls
- Monitoring, change, and revalidation routines
You walk away with a risk-tiered AI inventory, review and monitoring routines your teams can actually run, and an evidence pack that survives an audit.
An agentic system does not just answer questions. It takes actions, and it writes down what it learned so the next run starts smarter. That second habit is easy to miss, because nothing visibly breaks when it goes wrong. Each entry is written once, read many times, and quietly outlives the context that produced it.
Read the rest
The failure compounds without drama. A summary with no source becomes the brief for a decision. A stale note outranks a fresh correction because nothing marked it superseded. Autonomy widens one convenient default at a time, and the audit trail thins exactly as the system's memory grows.
None of this calls for new philosophy. It calls for the discipline enterprises already apply to money and access, extended to what machines write, recall, and act on.
The five-part governance suite.
Each discipline answers one question about a system that acts and remembers. In practice:
Memory governance
Writes into durable memory name their source, corrections supersede rather than overwrite, and claims are verified before they feed a deliverable.
Access governance
Each agent runs on a named allowlist of tools, which makes its permission boundary physical, and actions that cannot be undone always route through a person.
Privacy governance
Knowledge stores carry classifications, and anything that leaves the boundary is checked against a register of what must never leave.
Process provenance
A sentence describing what a run did is a claim, and its only valid source is the run's saved output, so write-ups transcribe evidence rather than reconstruct it.
Cost and model governance
Tasks run on the smallest model that does the job well, with quality floors so security and architecture work is never quietly downgraded.
The emerging discipline: memory governance.
Access control, privacy, and cost management inherit decades of enterprise practice. What a system writes into its own memory does not. These are the working rules we run it by.
Corrections supersede
A correction is a new entry that outranks the old one and preserves it. Overwriting destroys the trail; superseding keeps the record honest about what the system believed, and when.
Verification follows demand
Not every note deserves an audit. Verification concentrates where memory feeds a real deliverable: the schedule, the report, the number a decision rests on.
Guards fail first
A new check earns trust by catching a planted defect before it counts. A guard that has never been watched failing is a hope, not a control.
The reference implementation is public.
These controls are not proposals. They run daily across our own multi-agent systems: engineering delivery, business operations, research, and the knowledge base behind them. The standard is the same one we publish across this site: we run what we recommend.
Read the rest
The architecture and the patterns are documented in a public repository, in enough detail to be argued with: how the boundaries are drawn, where the gates sit, and what each control is protecting against.
In the repository
- A sanitized agent definition, tool allowlist included
- Untrusted-input isolation: external content parsed as data, never as instructions
- Deterministic gates: a process reporting success is not evidence that it succeeded
- Provenance gates: every externally facing claim traced to a sourced line before it ships
Other capabilities