The Cadence Resumes

A note on the cadence

The weekly report has not shipped since Issue #1 on April 17. The v4.0 methodology whitepaper consumed the writing capacity, and the cadence broke. Issues #2 and #3 will not be backfilled because backfilling pretend-historical reports would be dishonest. This is Issue #4, published today, on the day the cadence is supposed to fire. Future issues ship every Friday.

Headline numbers

188,224 agents indexed on-chain across 14 blockchains. 186,321 of those are rated. 8 agents earn the top published grade of B- (numeric score 75.20). Zero earn a B or above. The bottom of the distribution continues to grow faster than the top.

The top tier is flat. Same 8 agents at B-, unchanged in count since Issue #1. The distinct identifier for each remains stable across the re-rate cycle that ran on May 5 and 6 of this week.

Cumulative deltas since Issue #1 (April 17)

Three weeks of growth, here are the totals.

The interesting line in this table is the flat one. Across three weeks of new agent registrations, total passing agents did not move. The new supply went into the D and F tiers. The credible tier population is constant in absolute terms and shrinking in proportional terms.

What we shipped this week

v4.0 methodology whitepaper. About 18,000 words across twelve sections plus appendix. Three structural contributions: seven signal categories, a non-substitutable safety floor, and two source-level corrections (source-conflict discounting and time decay). On-chain oracle specification. Honest limitations. Available at /methodology/whitepaper.

Companion announcement post. A 1500-word walkthrough of the methodology for journalists and analysts who want the argument without the math. Available at /blog/the-treebeard-methodology.

The 2008 question post. Why structural opacity and calibration opacity are not the same thing, and why the FICO architecture is the right precedent. Available at /blog/the-2008-question.

Three new /learn pages. What an agent rating is. How to improve a rating. Agent credit scoring vs traditional credit. Available under /learn.

Re-rate pipeline restored. The continuous re-rate, which had not run since April 15, fired on May 5 and 6. Top-tier ratings are now fresh as of this week.

Findings worth pulling out

The 60-point spread. Five competing rating services rated the same agent (Vexalith Obsidian, ERC-8004 ID 13684) this week and produced scores ranging from 18 to roughly 80. Treebeard sits in the middle of the distribution at 60.1 with a LOW confidence flag, honestly reflecting the coverage gaps on this agent. None of the other four publish their methodology. When five raters disagree by 62 points on a single agent, methodology transparency stops being academic.

Wallet age as an unsurfaced signal. One competitor (RNWY) scored the same agent 100 of 100 on operator wallet age (the wallet has been active for 4 years 10 months). Treebeard already stores the operator wallet but does not yet surface wallet age as either a rating input or a profile-visible fact. A Skinbark brief is queued to ship this signal in the coming week.

The dependency layer is still open work. The v4.0 whitepaper Section 11.7 names cross-agent dependency rating as an open problem. An agent that runs on top of another agent's runtime inherits that runtime's safety properties, and most current rating systems (including ours, in v4.0) do not propagate this. Real failure mode in production.

Coming next week

Wallet age signal ships into Security Posture as a sub-signal, with a methodology version bump and calibration-shift summary. The case-study post on the Vexalith 62-point spread goes live as a companion artifact. Issue #5 of this report fires next Friday morning.

Find an agent in the directory at /agents. Read what moves a score at /methodology/improve. Read the methodology in full at /methodology/whitepaper.

We rate every agent the same way. Methodology public. Calibration-shift summary on every version. No token. No payment from rated entities.