The Toolchain
The vetted build kit, mapped to the engagement phases — every tool reads from the same record.
Confidential · June 2026. The toolchain below is the kit we standardize on; specific deployments vary per engagement. The Aperture engine is built and tested today against synthetic data — live connectors are the next build, so treat everything downstream of Aperture as modeled until connected. Grace Hill references model a plausible internal estate inferred from a mature mid-market B2B SaaS profile; they are assumed/illustrative, not confirmed facts.
How to read this
govrn is a standard plus an engine. The standard tells you what must be true — three lenses (Technology, Security, AI), a product-AI review, independent attestation, and crosswalks to six frameworks. The engine, Aperture, is the spine: a deterministic, metadata-only system of record that holds the one record both the build side and the attest side read from. Every other tool in the kit either feeds Aperture (telemetry, lineage, audit, policy decisions) or consumes its record (the crosswalk library, the attestation pack). No model sits in the measurement path; Aperture moves pointers, not payloads.
We distinguish ours from vendor-neutral deliberately. The standard, the crosswalk library, and Aperture are ours — the things that make a govrn record a govrn record. Everything else is an open or commodity tool chosen because it emits machine-readable evidence we can carry forward. We do not lock a client into a tool we own where a neutral one will do.
The kit, by capability and phase
| Capability | Tool | What it does | Phase it lights up |
|---|---|---|---|
| Policy-as-code | OPA / Rego (neutral) | Encodes governance rules — access, data-residency, model-use guardrails — as evaluable code, so a control is a decision you can replay, not a PDF | Design (write the policies), Operate (enforce + log decisions) |
| Compliance evidence | OSCAL (neutral) | Renders controls, assessment results, and the system security plan as machine-readable artifacts — the attestation pack is generated, not hand-typed | Attest (emit the evidence), Re-baseline (diff against prior) |
| LLM audit / observability | Langfuse (neutral) | Traces prompts, completions, tool-calls, and cost for the AI a company ships — for Grace Hill, this is where "Gracie" gets instrumented so fair-housing-sensitive outputs are observable | Implement (wire the product), Operate (continuous review) |
| Telemetry | OpenTelemetry (neutral) | Standard spans/metrics/logs from infrastructure and services — the substrate signal Aperture reads to attribute usage and cost | Assess (baseline the estate), Operate (steady-state) |
| Data lineage | Apache Atlas / Marquez (neutral) | Maps where data originates, flows, and lands — establishes which systems touch regulated data (tenant-screening records, in the assumed Grace Hill estate) | Assess (map the estate), Design (scope controls to real flows) |
| System of record | Aperture engine (ours) | The deterministic, metadata-only spine — surfaces shadow AI with a dollar figure, attributes cost, holds the one record both lenses read. No model in the measurement path | All phases (the record every phase writes to / reads from) |
| Security event routing | SIEM (neutral) | Receives security-relevant events (incl. OPA denials, anomalous model access) and correlates them — the Security lens's live nervous system | Implement (route), Operate (detect + respond) |
| Identity / provisioning | IdP + SCIM (neutral) | Authoritative identity and automated provisioning/de-provisioning — answers who and what can reach which AI and data | Design (define access), Operate (enforce least privilege) |
| Secrets management | Secrets vault (neutral) | Centralizes keys, tokens, and model-API credentials so AI access is governed, rotated, and never embedded in shadow tooling | Implement (consolidate), Operate (rotate + audit) |
| Framework crosswalk | Crosswalk library (ours) | Maps every control and finding to NIST AI RMF, ISO/IEC 42001, EU AI Act, OWASP, MITRE ATLAS, and CSF 2.0 — one record satisfies six frameworks | Attest (prove coverage), Design (target the right controls) |
Why the spine matters
The failure mode in governance work is a pile of tools that each tell a partial story and never reconcile. Our discipline is the opposite: every tool above emits evidence into a single, deterministic record. OPA writes its decisions, Langfuse writes its traces, OpenTelemetry and the lineage graph write the estate, the SIEM writes the events — Aperture holds them as metadata, and OSCAL plus the crosswalk library render that record into an attestation a separate body can sign. One record, two lenses: the team that builds and the team that attests read the same source, which is what makes the attestation defensible.
That is also why independence is structural rather than rhetorical. Because the record is deterministic and the engine keeps no payloads — pointers only — the attesting body can verify against the same record without trusting the build team's narrative.
Honest staging
Aperture exists and runs today against synthetic estates; the connectors that pull live OpenTelemetry, lineage, Langfuse traces, and SIEM events into a real record are the next build. Until those connectors run against a given client, any figure Aperture produces — including shadow-AI dollar attribution — is modeled, and we say so. For Grace Hill specifically, the internal estate (the systems behind "Gracie," the tenant-screening data flows, the IdP topology) is an assumed mid-market SaaS shape used to exercise the kit, not a measured one. We earn the word "measured" only after the connectors are live.
See also: The Standard · The Aperture Engine · Crosswalks