MNMatthew Nolan

Work / System

NALA

The governance system I run my work on.

Status
IN USE Running every day, privately. Not a public product.
My role
Designer and operator. I wrote the rules, built the pieces, and answer to it every day.
Built with
Scheduled agents · Local LLMs · Supabase · Cloudflare · Markdown
See it
Private. The architecture below is the evidence.

01The problem

I work alone, with a day job, across several products. Context gets lost between sessions. AI tools sound confident whether or not they’re right.

The risk isn’t a bad answer. It’s a bad answer that looks verified.

02The idea

Treat my own operation like something that needs governance: rules, evidence, escalation and memory. Then let agents work inside those rules.

03The build

Truth Layer. Claims carry evidence tags. “Deployed” means checked three separate times against the real URL, not a cached page.

Agent trust scoring. Agents earn latitude by being right, and lose it by being wrong.

Decision memory. Decisions are written down with the reason, so I don’t re-argue them at midnight.

Risk register and escalation tiers. Agents handle what they can. Money and strategy always come to me.

One shared state file. Every session reads it first. It’s the only bridge between my terminal, my phone and the scheduled jobs.

  1. MatthewOwner. Final say on money, strategy and truth.
  2. NALAGovernance layer
    • Truth Layer
    • Agent trust scores
    • Decision memory
    • Risk register
    • Escalation tiers
    • Shared state
  3. Agents19 active on schedules · 14 paused or retired
Work flows down. Evidence and escalations flow up. Anything financial or strategic always reaches me.

04The system

Hard rules live in plain text and get read on every startup. Example: never roll back a production deploy on one host, because it wipes environment variables. Retry the latest build instead. That rule exists because it happened.

05The decisions

Plain files over a database for the core state. Easy to read, easy to diff, easy to recover.

Cancelled paid AI subscriptions. The stack runs on local and free models, with fallbacks when a cloud route runs out of credits.

06The failures

A session once overwrote the whole shared state file with a one-line note. What happened →

07The result

In daily use. Private, so there’s nothing to click. The architecture is the evidence.

08What I learned

Verification is a feature you build, not a mood you’re in.

The most useful thing an agent can say is “I don’t know.”