Delta Atlas
Loading…

Delta Atlas

Simplification is systems maintenance; the substrate is human cost.

My theory of agentic AI governance and the instruments I built from it, shown end to end with receipts.

Three things you can use today

Complexity is technical debt, and the debt is paid in human cost somewhere down the line. Each card below says what the thing is, how to run it now, and where its receipt lives. Nothing here needs an account, an API key, or a model call.

Framework

BespokeNode

A governance framework for agentic AI. Nothing an agent does counts until it has been defined, permitted, gated, verified and accepted, in that order, by the right party at each step. Two safety states, HOLD and TRIP, each with its own recovery path. Four invariants that never bend, among them: approve exact output, never just the plan; and no automatic restart.

Run it now

pilot-001 is the first clause run through the framework end to end: sealed expectations, blind two-track execution, a contradiction found and resolved on the record, an independent audit that attacked the checker, and public receipts a stranger can re-run. MIT licensed.

The same discipline governs this site's own changes: every proposed change is a state-delta in a register, sorted by a human with a written reason. Read how the register works.

Receipt

pilot-001 was minted gold by owner decision on 2026-09-23. The public specification is a v0.1 candidate. Implementation, acceptance and independent replication are pending and stated as such.

Walk the architecture, top to bottom →

Skill

Pin-Check

Every precise number in a technical document came from somewhere. Pin-Check reads a document, pulls out every value precise enough that it must have a source, and looks for that source under a directory you name. Each value comes back traced, rounded, or currently undefined.

Run it now 
python skills/pin-check/pin_check.py <document> <root>

Python 3, standard library only, no install, no network. Exit code 0 when everything traces, 1 when anything does not. The README explains each result, including the copy-detection step to read before trusting a clean run.

Receipt

This home page and the architecture page were run through Pin-Check against this repository before publication. The calibration log records the run. The tool finds broken links between prose and artifacts, not bad math.

Read the source on GitHub →

Method

OpenMirror

A method for turning a metaphor into policy one correspondence at a time. The analogy is treated as a one-to-one mirror: verification is correspondence by correspondence, and a break at any joint is a precise, located failure rather than a reason to throw the analogy away.

Run it now

The Bridge Card is the one-minute form of the method. Six fields: SOURCE, CLAIM, OPEN TERMS, PARALLEL, HINGE, STATUS. Copy it into any assistant as a custom instruction or fill it by hand. It ends in accept, not yet, or a named uncertainty.

Receipt

The worked example is the V=IR crosswalk: authority as voltage, effects as current, the gate as resistance. I red-teamed it against a first draft; those notes are not yet published. It is a crosswalk with a stated claim frame, not an empirically validated law.

Read the crosswalk →

What is proven, and what is not

Proven

  • The probe machinery works end to end on synthetic cases: declarations, sealed expectations, blind execution, contradiction detection, audit hardening, public receipts.
  • An audit leg with an opposing brief catches errors in the advisor's own record, not just the executor's. Roles with opposing briefs produce friction by construction.
  • The mechanism probes detect their targets on deterministic synthetic cases.

Not proven

  • Anything about real substrates. No evidence on true concurrency, scheduler preemption, power loss, crash durability, or production behavior.
  • The V=IR mapping and the simplification thesis are crosswalks with a stated claim frame, not validated laws.
  • Nothing has been independently replicated.

The instruments

Deterministic browser tools built along the way. Fixed rules, versioned word lists, no model at runtime. They read structure, not meaning.

  • Paste a plan or policy. It looks for a risk named with no control beside it, and for autonomy with no oversight named.
  • Paste an agent's action log. It flags wording where a goal climbs over a rule, or a check is bypassed.
  • Paste a framework or plan. It scores each part on whether it says what to do, whether you could tell if it worked, and what happens on failure.
  • Answer five short questions about a task and its backup. It reports what is missing when the usual person is away.

Experimental checks using fixed rules. Results are questions to review; a clean result does not certify correctness or safety.

Preview Gap Check in place

A fixed example running in the real tool. See which phrases it recognizes and what it asks you to review. It does not verify the plan.

Search the glossary

Recorded definitions and relationships for the language of agentic AI governance, defined by function. A fixed lookup, not a chatbot.

This answer comes from the fixed function-first engine, not a chatbot.