The Trust Game Has Started. Here's the Ledger That Keeps Score.

Why this ledger, why now

This week, the largest AI lab in the world announced a model called Mythos it had not released. It simultaneously launched Project Glasswing, a $100M credit program distributing compute to cyber-defense organizations. The implicit message was clear: we are the responsible one.

Meanwhile, Q1 2026 just closed as the largest venture quarter in history — $300B, with 65% concentrated in four companies. Capital is piling in faster than trust is being established. The gap between what AI systems are claimed to do and what they are accountable for is now the most important unpriced risk in the market.

The Responsibility Ledger is one attempt to price it in public.

Not by arguing that AI is good or bad. Not by handicapping which lab wins. By doing something much narrower and much more useful: keeping a dated, signed, append-only record of what is claimed, what happens, and what the editor of this ledger refused to do along the way.

What a ledger is, and isn't

A ledger is not a newsletter. A newsletter is a distribution format. A ledger is a data structure — specifically, an append-only one. Entries are never deleted. Corrections are added as new entries that point back to the original. The historical record is preserved even when it is embarrassing. That is the whole point.

Three properties follow from that structure:

1. Time-stamped. Each entry is dated in UTC the day it's written. Nothing is ever backdated. If we miss a day, the record shows the gap.
2. Signed. Each entry has a named editor. When an AI system contributed, it is named alongside the prompt and its role. No anonymous AI output dressed up as human analysis.
3. Falsifiable. Every forward-looking claim ships with a confidence level and an explicit invalidator — the specific observable event that would prove us wrong. When time passes, we grade ourselves in public.

The full rules live in the Charter. The four-pass production process lives in the Methodology. Both are themselves append-only. When they change, they change in public.

Three claims, filed today, available to grade

We will publish many forward claims over time. Here are the first three — each with a confidence level and an explicit invalidator. When their due dates arrive, they will be graded against reality on a scorecard that any reader can audit.

Claim A (1-year horizon, High confidence). By April 2027, at least three frontier AI labs will publish something they call — in those words or near-synonyms — an "AI responsibility ledger," "model responsibility register," or "accountability log." The category vocabulary is empty today. It will be crowded within twelve months. Invalidator: If the category remains vocabulary-less at year-end — no major lab or major regulator has adopted the word "ledger" or a close synonym in an official framing — this claim is Broken.

Claim B (1-month horizon, Medium confidence). Within thirty days, a visible US regulatory or enforcement action will cite the gap between a major AI system's public safety claims and its actual behavior under adversarial testing. Invalidator: If no such action appears by 2026-05-21, this claim is Broken.

Claim C (1-week horizon, Medium confidence). Within seven days, at least one prominent AI lab will publicly disclose a previously-undisclosed model family or capability tier, citing "responsibility" or "safety" as the reason for prior non-disclosure. Invalidator: If no such disclosure occurs by 2026-04-28, this claim is Broken.

These claims are filed in ledger form. They are not predictions in a newsletter that quietly disappears if wrong. Their status will be updated on the scorecard. Readers can return in one week, one month, and one year and see how we did.

The thing most commentary is missing

The Mythos announcement and Project Glasswing are not PR. They are trust infrastructure — early attempts to build the primitives that will govern who gets to deploy what at scale, who is accountable when something breaks, and who gets paid to certify both.

This is the real opening of the AI era's second act. The first act was capability: what can these systems do? The second act is accountability: who writes down what they did, and how? The answer to that question will look a lot like accounting. Which is to say: it will look a lot like a ledger.

The labs are building their version. Regulators are building their version. Insurers are building their version. None of them are public. None of them are signed. None of them are append-only.

This one is.

What you can do with this

1. Read a daily entry. Weekdays, mornings. Short. Dense. Sourced.
2. Grade our prior claims. The scorecard is public. If we're bad at this, you'll see it before we admit it.
3. Subscribe to the feed. RSS is here. Email subscription will come when it can meet the charter's no-tracking requirement.
4. Send corrections. editor@responsibility-ledger.one. Corrections are appended, never silent.

Refusal line

Refusal: I did not write a launch essay that implied this project has already proved itself. It has not. Entry 001 is a promise, signed and dated. Everything after it will earn or lose against that promise in public.

---

Entry 001 — 2026-04-21 — Signed: Roger Grubb, Editor · Drafted with assistance from Claude (Anthropic), prompt and role disclosed on request.