Connect every portfolio decision to evidence.

Provider and spending boundary

Each organization brings its own AI provider and its own key. The key is held encrypted, and every call is counted against a monthly spending ceiling before it runs.

Proposal before action

Generated plans and updates retain their inputs, output and usage record. A person can accept or reject the proposal with a recorded reason.

Acceptance names a person

Who accepted comes from the signed-in session, never from the AI output or from anything the browser sends. An agent can propose; it cannot sign.

Append-only decision trail

Decisions on proposals, and every other governance act, are appended to the trail. Deletion is blocked in the database, and every release proves it by attempting the deletion.

Which models are approved, and where they run

An approval names the model, its hosting posture — hardware the organization controls, a tenant-isolated deployment under contract, or a public API — and whether work carrying regulated or member-identifying content may reach it. That last decision is separate and starts at no.

The approver states the basis in their own words, and an approval without one fails the write. Withdrawing an approval records the withdrawal rather than removing the row, so a run that happened under it stays explainable afterward.

What a run did, stage by stage and lane by lane

A run records what it was asked to do, which approved model it ran under, and its outcome at each stage: plan, build, review, verify, ship.

Where the work fanned out in parallel, each lane records the slice of the tree it owned, the lanes it waited on, the harness that ran it and the commit it landed. Two lanes claiming one path is surfaced before the merge instead of discovered after it.

Which rules the run was held to

A workspace adopts operating policies from a written catalog and sets each one advisory or blocking — an agent may not grant itself authority; a result is a proposal until a person accepts it; regulated content stays inside its approved boundary; a waiver names a person.

What a policy returned for a run is sealed with a digest over the bytes it covered. The table has no update path, so a sealed verdict cannot be rewritten; a correction is a new run. A policy with no verdict against a run reads as not evaluated, never as passed.

Where credentials are not

The registers record where a key lives, in words — the environment variable, the vault entry. Never the key. A constraint refuses key-shaped text at write, so a paste accident is a failed write rather than a secret sitting in the database an estate points its reviewers at.

What a workspace may configure

Its own name, wordmark and accent color, applied across the product for the people who work in it. Everything that governs how evidence is judged stays the same in every workspace.

Recorded scenarios preserve their assumptions

A named person can author a scenario with project stances, funding references and notes. Comparisons show funding, capacity, benefits and dependencies with the basis and denominator for each figure; the product does not select a preferred scenario.

Shared control definitions

The evidence ladder and its order. The finding classes. The status colors, checked for contrast and for color-vision safety. Who may accept a claim. These are governed product definitions, not labels a workspace can rename.

Enterprise identity

SAML 2.0 supports Microsoft Entra ID and other standards-based identity providers, routed by your verified email domain and enabled per deployment. Your identity provider authenticates the person; SprintLoop membership still decides which workspace they reach and what they may do in it.

Tenant isolation

Every record carries the organization that owns it, and the database refuses a read or write that crosses that line — not only the screens in front of it. The same boundary applies on every plan.