How it works

Built like infrastructure, not like a dashboard.

This page is for the reader who wants to know why the receipts can be trusted. Short version: the provenance is not a feature layered on top. It is the write path.

The write path

Nothing writes a table directly. Ever.

There is exactly one way data becomes truth in Wavetable, and everything (connectors, packs, automations, agents, even your own inline edits) goes through it:

EventSomething happens. An email arrives, a form is filled, a human edits a fact. Appended to the log, encrypted per subject.

Claim (proposed)A statement about an entity, carrying source events, asserter, and confidence.

AdjudicationSerialized per tenant. Source trust and recency decide; low confidence goes to a human queue.

ProjectionThe tables you query. Caches of accepted claims, rebuildable from scratch at any time.

This one-way flow is why every answer can show a receipt, why a bad import is recoverable, and why erasure can actually erase: redact the claims, shred the keys, rebuild the projections.

The planes

Three planes, one substrate.

The architecture separates where intelligence lives from where your business runs, with a bounded conversational agent in between.

Design plane · AI, slow, human-gated

The Compiler

Agent sessions that author artifacts: schema migrations, views, workflows, policies. Never touches production state directly. Its output is always a diff a human approves, with a dry-run or replay report attached.

Run plane · deterministic, fast, cheap

Compiled programs

Approved artifacts execute as durable workflows with typed guards, budgets, and blast-radius limits enforced by the runtime. Every run leaves a reproducible trace with its cost.

Edge agent · bounded, interactive

Ask, drafts, proposals

The conversational surface. Read-only queries, cited memory reads, drafts, and proposals. No external sends, no claim acceptance, no projection writes. It proposes; you approve.

Underneath: one isolated substrate per workspace. Its own SQLite database, its own coordinating runtime object that serializes every write, its own file storage prefix, its own vector index namespace.

Replay

The counterfactual comes first.

Nothing recurring goes live on promises. A candidate workflow is replayed against your real event history in shadow mode: same guards, same projections as they stood at each moment, zero side effects. The approval you see says "this would have fired 14 times last quarter" and shows the drafts. Policy changes get the same treatment: yesterday's approvals under the current policy versus the candidate, side by side. You are never asked to trust a description of behavior when the system can show you the behavior.

Honest limits

Where the models are, and where they are not.

Expensive, capable models do the high-stakes design work, always behind a human approval. Mid-tier models write cited memory and answer Ask. Small models do high-volume extraction under strict output schemas. And the things that must never be creative (adjudication, projections, workflow execution, capability checks) use no model at all.

The constitution

Ten principles the system cannot violate.

These are build invariants, enforced in code review and architecture, not aspirations on a poster.

AI at design time, deterministic at runtime.The agent is a compiler, not the runtime. Anything recurring executes as a compiled, inspectable program.
Facts are claims, not rows.Every fact carries provenance. Tables are projections, which are caches, always rebuildable from accepted claims.
Everything is a versioned artifact.Schema, views, workflows, policies. Diffable, approvable, reversible. Your business, version controlled.
Trust is earned by replay.Before an artifact goes live, it is tested against your real history in shadow mode, and you see the report.
The edge agent proposes; policies grant.No agent takes external action without a capability minted by an approved policy. Blast-radius limits are structural.
Wrong merges are the cardinal sin.Identity resolution prefers precision over recall. Merges are soft and reversible.
No uncited sentences.Every sentence in memory maps to claims or events. Prose without provenance is not allowed in the system of record.
Isolation is physical.One database per tenant. No cross-tenant query path exists anywhere in the codebase.
Opinions are packages, not prisons.Vertical templates ship as forkable packs. The kernel stays unopinionated.
Be the home, not the competitor.Every workspace exposes a scoped MCP endpoint so any authorized agent can use the graph. Lock-in comes from where the data lives, not from walls.

Kick the tires on the physics.

Create a workspace, connect a source, and inspect the receipts yourself. The docs go deeper on every layer.

Get started free

No credit card. Connect your data, get receipts in minutes.