Launch Command Center · Confidential client · 2026
An interpretable AI layer for a multi-market healthcare launch
An interpretable AI layer over a multi-market healthcare launch. It reads, drafts, and flags, but never decides, and you can always open three files to see why it said what it said.
Snapshot board
Asana → timestamped JSON
Agents draft
blocked, at-risk, recommended action
Human gate
Telegram, approve or reject
Email + logged
approval logs a name and a date
- 4
- workflows, one gate each
- 4
- human gates, at the judgment
- 0
- irreversible machine actions
01 · Problem
A multi-market clinic launch is a hundred clocks running at once
Credentialing timelines, licensing windows, payer enrollment, permits, buildout. A launch never dies in the status meeting; it slips in the gaps between them. A permit quietly breaches, a licensing clock runs while everyone watches a different market. By the time it surfaces, the cheap fix is gone and the open date is at risk.
The client is anonymized; the board and its data stay private. What’s public here is the architecture.
02 · Solution
Reject the black box: a layer that reads, drafts, and flags, but never decides
Everything an agent knows lives in one of three layers, each a set of readable files in Git. Ask “why did the agent say that?” and you open three files.
Instructions
what it was told
Context
what it saw
Memory
what it remembered
It runs on two frameworks. WAT(Workflows, Agents, Tools): probabilistic AI reasons, deterministic code executes. The script that pulls the board doesn’t think; the agent that reads it doesn’t touch anything. ICM (Interpretable Context Methodology): the three layers above, versioned in Git.
03 · The constitution
The gate sits at the judgment, not after the action
Most automation
The system does the thing, a person reviews the wreckage. The gate is after the action.
Here
The machine runs up to the irreversible call, then stops one tap short. The gate is on the judgment.
Hard gates, written into every agent’s charter: no agent writes to the board except flags and comments; nothing external is ever machine-sent, draft only; patient data, licensing filings, and credentialing documents are read-only, forever. In a regulated environment the expensive failure is never a missed status update. It’s an AI asserting something untrue about a licensing deadline, in an email the whole leadership team read.
04 · Status
What works today, what's designed
Working
The Monday spine: the board puller produces real timestamped snapshots, two agents run as wired subagents, the Telegram gate is connected and confirmed delivering.
Designed
Three more workflows on the same gate pattern (threshold breach, CEO pre-read, post-launch retro) plus a review dashboard, on paper.
Every output ends with two blocks: SOURCES, the task IDs the agent actually read, and CONFIDENCE, anything it inferred rather than saw. An uncited claim is treated as a bug, not a feature.
05 · The tradeoff
It trades fluency for auditability, on purpose
A system that refuses to infer is only as good as the board it reads. When the board is stale, the brief doesn’t paper over it with plausible guesses; it surfaces the staleness as a low-confidence flag, and board hygiene becomes a visible, fixable problem. That’s the correct trade everywhere a regulator, a payer, or a parent might one day read the paper trail.
Why did the agent say that? Open three files. Instructions are what it was told. Context is what it saw. Memory is what it remembered.
Stack: Claude Code for agents and orchestration, Asana via MCP and API, TypeScript tools, n8n on Railway for scheduling and routing, Telegram for the human gate, Git for versioned instructions, context, and memory.
Want the code, or a version tuned for your team? The build is on request.
Request the code