# Cascadian — notes for agents

## What this is

Cascadian is the agent-native causal decision substrate: one MCP server in front
of one Postgres, sitting underneath your agent at the moment of decision. Every
fact lives in tritemporal memory — event time, valid time, knowledge time — with
`available_at` as the anti-lookahead axis, so a reconstructed decision can only
see what was knowable when it was made. Causal structure lives in a governed
graph: you may propose nodes and edges, but only a human approval promotes them
to truth. No causal estimate ships without passing the identification gate;
anything unidentified is watermarked exploratory. `simulate_intervention` runs
do(X=x) on the approved graph and returns every downstream effect with per-edge
assumptions. Decisions, recommendations, and outcomes land in an append-only
ledger, where expectations are frozen before outcomes arrive — so calibration is
measurable and judgment compounds.

## Connect

One command:

```
claude mcp add cascadian --transport http https://mcp.cascadian.ai/alpha
```

Or in mcp.json:

```json
{
  "mcpServers": {
    "cascadian": {
      "url": "https://mcp.cascadian.ai/alpha"
    }
  }
}
```

## Tool surfaces

- **Memory** — what is known: `assert_fact`, `search_memory`, `get_context_pack`
- **Causal** — why: `propose_edge`, `identify`, `estimate`, `simulate_intervention`
- **Decisions** — decide and learn: `open_decision`, `record_decision`, `record_outcome`, `review_calibration`
- **Governance** — trust: `request_approval`, `reconstruct_decision`, `detect_leakage`

## The six invariants

1. History is never silently overwritten — ledger tables are append-only; corrections insert a new row and close the old one.
2. No future information in past decisions — every read takes `as_of`; open decisions pin theirs automatically.
3. No causal estimate without identification — unidentified estimates are exploratory and labeled as such.
4. No hidden arbitrary scores — recommendations carry rationale, explicit comparison, and assumptions instead of a scalar.
5. Agents never define truth — every agent write lands as a proposal; the approval token never crosses the MCP boundary.
6. No high-risk action without approval — the requirement is computed server-side from the decision, not from tool arguments.

## Where to look

- Machine-readable site map: https://cascadian.ai/llms.txt
- Human docs: https://cascadian.ai/mcp-api — quickstart at https://cascadian.ai/mcp-api/quickstart
- Access is pre-launch alpha, gated while we run paid pilots: briefings@cascadian.ai