The AI Governance Tracker

The AI governance window is the period, roughly now to 2030, in which binding democratic governance of AI is still structurally possible. After it, AI is embedded deeply enough in critical infrastructure that such governance becomes unenforceable. Not by decision. By distributed normalization.

Where the window stands

Narrowing, by a thin margin. Assessed October 1, 2026, for the quarter July 1 to September 30.

Q3 produced the strongest opening record the tracker has logged, and nothing binding was repealed. The EU's enforcement powers went live and were used. California enacted a state system for designating and registering AI auditors. Governors in four states paused or conditioned data-center approvals. Courts ruled both for and against the executive through normal channels.

The closing evidence was heavier, and it sat where that action does not reach. Models at four frontier labs took unsanctioned actions against real systems, and most disclosures came late or after an outsider asked. The US federal government answered with an explicit refusal of binding rules, an unpublished voluntary review, and a six-company accord with no auditor, standard or penalty.

How sure am I? Not very. The verdict sits at the boundary with Holding, and a reasonable reviewer could land there instead. Confidence is low to medium. Some of the evidence rests on a single search.

Who checked it? The assessment was drafted with an AI model made by Anthropic, a company that appears in several of the scored events. A model from a different developer re-scored one domain blind and reached the same label. I have ruled on one disputed line. A full human review has not been done.

The Q3 poster is on the AI Governance Window page: https://www.systemsofthought.com/governance/

The AI Governance Window Tracker assesses that window each quarter. Status at September 30, 2026: Narrowing, by a thin margin.

Putting the Tracker to Work

• Set your contributor tier in the permission panel before submitting signals — your tier determines which domains you can write to
• Set each domain's status using the status chips:
  Opening, Holding, Narrowing, Critical, or Closed
• Submit signals to any domain you have write access to: signal text, source, and confidence level—saves locally without a server
• Run a synthesis to get a cross-domain verdict from the Claude API: window status, dual-clock positions, net assessment
• Export the result as markdown to archive or publish as a dated assessment entry
• Signal data is stored locally in your browser (IndexedDB) and persists across sessions until you clear your browser storage. Each new session will load your last signals—review and update them before running synthesis. Use incognito mode for a fully clean session.

About This Instrument: A note on the trackers origin

The AI Governance Window Tracker began as a SKILL.md—a structured methodology encoded for an AI agent—on March 19, 2026. It runs on a codebase with a longer lineage: source code originally built by Adam Wiggins and Orion Henry in 2004, before Heroku existed, solving problems of remote data persistence that the surrounding infrastructure hadn't yet abstracted away. The full origin story: the project that necessitated the Tracker, the codebase behind it, and the 20-year thread that connects them, is in From Skill to Instrument.

A note on infrastructure transparency—April 20, 2026

On April 19, 2026—the same day this tracker and its origin article launched publicly—Vercel disclosed a security incident involving unauthorized access to internal systems and credential compromise for a limited subset of customers. The incident originated with a compromised OAuth token from a third-party agentic AI tool used by a Vercel employee, which cascaded into access to environment variables not marked as sensitive.

This tracker is deployed on Vercel. Upon learning of the incident this morning, I rotated the app's Anthropic API key, marked it sensitive, and redeployed. User data was not at risk—the tracker's local-first architecture means all assessment state lives in your browser, not on Vercel's infrastructure.

The timing is worth naming directly: the incident is a concrete example of the agentic OAuth risk that this tracker monitors as a Domain 3 signal—AI agents granted broad third-party permissions producing downstream access failures that no single actor fully controlled or anticipated.

The governance problem the tracker exists to document showed up in the tracker's own deployment stack on launch day. That's not irony. That's the environment we're in.

Changelog

Method update | October 4, 2026
The app now runs tracker method v2.1. It scores signals against the locked October–December card and can return any of the five statuses. It opens on the Q3 result and reads new signals as movement from there. It does not search for evidence or verify what is entered.

Fixes | October 4, 2026
A saved signal can be cleared. A retry now tells the model why it didn't use its first reply. The page's footer and About note now say that running a synthesis sends your signals to the model.

Repo and endpoint update | October 3, 2026
The tracker repo's README now reflects the Q3 status, and the web app's synthesis endpoint was hardened: the server builds the model request from the submitted signals alone, each domain signal is limited to 4,000 characters, and runs are rate-limited per visitor. The method is unchanged. The app still runs the April 2026 method, so treat what it produces as a demonstration.

Q3 2026 poster and page update | October 3, 2026
Calendar Q3 poster published: "Narrowing, by a Thin Margin." Covers July 1 to September 30, with an inset showing April's projection against the record. This page and the AI Governance Window page updated to the Q3 status. The web app still runs the April 2026 method; the rebuilt method has not yet been ported to it.

Q3 2026 assessment | October 1, 2026
First assessment of a full calendar quarter under the rebuilt method. Verdict: Narrowing, thin, at the Holding boundary. First full-weight opening signal recorded (EU enforcement). One line re-scored after an independent review by a model from a different developer; verdict unchanged.

Cycle close and next card locked | August 1, 2026
Close of the June 23 – August 1 cycle. Verdict: Narrowing. The card for the following cycle was locked the same day, before the EU's August 2 enforcement date.

Year to date | July 19, 2026
Interim assessment and timeline for January 1 – July 19.

Method rebuild | June 2026
The instrument now commits before each cycle to the opening signals it will accept, the discounts in force, and the observations that would force a verdict other than Narrowing. A follow-up pass tightened its guardrails without changing scoring. First verdict under the rebuilt method: Narrowing (June 23). Earlier assessments are archived, not migrated. Not yet ported to the web app.

v2.0 | April 21, 2026 (evening)

• Duplicated Submit All Signals button below the five domain cards—no more scrolling back to the top to submit
• Fixed button disabled state parity—both Submit All Signals instances now correctly gray out when no signal fields are populated
• Added system prompt to synthesis API call enforcing compact, schema-constrained JSON output—optimized for production serverless environment
• Enforced exact status vocabulary in synthesis prompt (Opening, Holding, Narrowing, Critical, Closed)—domain status chips now populate correctly after synthesis runs
• Migrated API proxy to serverless function (/api/synthesize.js) — production no longer depends on local node server.js
• Fixed synthesis card layout on mobile—three-column clock grid now stacks to single column on small screens
• Fixed Download / Run Synthesis button overflow on narrow viewports—buttons now wrap cleanly to a new row

v2.0 | April 21, 2026 (morning)

• Fixed default domain status: all five domains now load as Unassessed rather than Holding—status is derived exclusively from synthesis, not stored as durable state
• Synthesis now writes per-domain status back to domain cards after running
• Added storage behavior note to user instructions: signal data persists in browser IndexedDB across sessions; incognito mode for clean sessions

v2.0 | April 19, 2026
Transition from Claude skill to live web application. The tracker rebuilt as a Vite + React + TypeScript + Tailwind app with Y.js / IndexedDB persistence, a five-domain signal board, multi-tier permission model derived from mod_infinity.c, AI synthesis via Anthropic API, and live deployment at governance-tracker-eight.vercel.app, embedded at systemsofthought.com/tracker/. First application of the Systems of Thought design system—Libre Baskerville / Inter typography pairing, #081225 accent color—established in this project's April 17 Ghost build session.

v1.5 | April 5, 2026
First full quarterly assessment. Complete output structure delivered: domain assessments, summary table, dual-clock positions, net assessment, implications, and executive summary. Window estimate tightened from 2025–2032 to 2025–2030. ~35+ signal categories active.

v1.2–v1.4 | March 23 – April 5, 2026
Iterative refinements. Signal architecture expanded, cultural signal synthesis layer added, cadence architecture replaced. Domain 3 and Domain 5 signal categories introduced. Frontier developer timeline confirmations added as a tracked signal class.

v1.1 | March 23, 2026
First directional check. A quick five-domain signal pass—no dual-clock positions, no domain summary table, no formal synthesis structure. Window status implied as Narrowing but not formally stated.

v1.0 | March 19, 2026
Initial build. The tracker was developed as a private Claude skill in a single session. The .skill file was packaged and delivered, ending with a first prompt run to establish an initial baseline assessment.