The control layer
The governed intelligence control layer.
KeepWise is a control layer around AI. Every AI response in KeepWise runs through it: checked before you see it, approved before anything consequential happens, and recorded so you can prove what happened. Law firms are its first market. The Verifier is one control built on it.
- Input
- Check
- Evidence
- Approval
- Record
Want to see the layer on your own work?
Request a demoWhat the layer does
Each control, in plain words, with where its evidence is.
Where a control maps to one of the 16 named controls, its current evidence status is shown beside it, exactly as on Security & Trust.
- 01
Checks before the model runs, and after it answers
Before the model is called, each chat request is validated, and a malformed one is refused (preflight). After the model answers, the response is held while it is checked (postflight): an answer that claims an action that never happened, discloses the system's internal configuration, or claims an approval it doesn't have is blocked, not shown. The case-citation checks run at the same point.
- 02
The action gateway and the approval rail
Every AI tool call that would change data is stopped at a single chokepoint and turned into a pending approval. It runs only after a person approves it through a separate, authenticated approval step. The model never sees a way to approve itself.
- 03
Fail-closed defaults
When a governance check cannot reach a clean answer, the action does not proceed. A response that fails postflight is blocked rather than passed with a warning; a Legal Hold check that errors stops the delete; a Legal Hold lift approval with no recorded requester is refused.
- 04
A hash-chained audit ledger, anchored outside the database
Audit events are written to a hash-chained ledger: each record's hash includes the one before it, so editing or deleting a past entry breaks the chain. The chain's starting value is kept outside the database, so it can't be quietly restarted from inside. Approval decisions are kept in a separate append-only record that standard database roles cannot edit or delete; that record is not hash-chained.
- 05
Tenant isolation, enforced by the database
Each firm's data is separated by row-level security in the database itself, not only by application code. A cross-tenant isolation test runs against a fresh database in CI on every proposed change.
- 06
Legal Hold, and roles within a firm
Owners and attorneys can see every client and matter in the firm, and roles decide what each person can do. A legal hold freezes a client or matter: database triggers refuse any delete, or move out of the matter, until the hold is lifted.
EvidenceSecurity & Trust - 07
Outside the model, across providers
The controls run on KeepWise's own servers, outside the model. Today Anthropic's Claude writes the chat answers, which pass through every check on this page, and Google's Gemini handles internal classification, summarization and the behavioral review. Both are on our Subprocessors list.
EvidenceSubprocessors - 08
Tested on every change
The layer is 16 named controls, each held to an evidence bar. Their tests run in CI on every proposed change, on every merge, and daily, against a fresh database, and that run is a required check before a change merges. Where a control's proof stops, the gap is written down, not hidden.
Same models, governed
Why not just use ChatGPT or Claude?
KeepWise doesn't replace the model. Today Anthropic's Claude writes its chat answers and Google's Gemini handles internal classification, summarization and review, both listed on our Subprocessors page. What KeepWise adds is the control layer around them.
Checked before you see it
Each response is held while automatic checks run: case existence and quoted language on AI-generated citations, and the postflight checks that block a response claiming an action it never took.
A person approves what's consequential
Every AI tool call that would change data becomes a pending approval and runs only after a person approves it.
Legal Hold
A client or matter can be frozen so nothing under it can be deleted or moved until the hold is lifted.
Roles within a firm
Owners and attorneys can see every client and matter in the firm, and roles decide what each person can do. An owner or attorney can give a paralegal or staff member access to a client and its current matters from the client's settings.
A tamper-evident record
Audit events are written to a hash-chained ledger, so an edited or deleted entry is detectable. Approval decisions are kept in a separate append-only record.
Controls are tested
16 named controls, with the gaps written down.
Every control is held to an evidence bar rather than asserted once. The full table, control by control, is on Security & Trust.
13 of 16 controls fully evidenced, 2 of 16 evidenced with a documented limitation, 1 of 16 partially evidenced.
Controls built on the layer
Domain controls use the same rails.
A domain control adds checks for one kind of work and inherits everything above: the checks before release, the approval rail and the record. The first is for legal work.
What the layer does not do
- It governs AI inside KeepWise. It does not monitor or control other AI tools your firm uses.
- KeepWise itself has not been independently audited for SOC 2; the controls are designed against SOC 2 criteria, and the evidence above is our own.
- It does not replace an attorney's review of AI output.