Working notes on the data behind the products — what the public records actually contain, what they leave out, and how to read them without overstating what they prove.
Written for counsel, risk analysts and engineers. Every figure is counted from a named source and dated; where a number is a screening signal rather than proof, the post says so.
17 August 2026 · Overview
WARN Act layoff data for employment attorneys at $149 a lead or $395 a short-notice candidate, county displacement signals over MCP from $299/month, twenty government and market data agents you can buy a single query from for $1, patent infringement evidence for either side, and custom agent builds from $6,000. What each one covers, what it costs, and where the coverage stops.
19 August 2026 · Database security
Row-Level Security disabled on a public table means anyone holding your project URL can read, edit and delete every row — no login. It happened in one of our own databases this week. The second exposure we found was worse than the one we were alerted to, the three-hour delay was caused by an inbox rather than a firewall, and the forensic query that said "all clear" turned out to be silently broken. Includes the 30 October 2026 deadline affecting every Supabase project.
24 August 2026 · Displacement
3,221 counties scored and refreshed against the newest BLS data. Counties in the top decile saw unemployment actually rise 89.8% of the time over the following six months, versus 15.8% in the bottom decile — a 74-point separation across 83,824 resolved observations. Includes the part most vendors leave out: the three-month signal is materially weaker, and we say so.
22 August 2026 · AI agents
Citing someone else's study is cheap, so we pulled the live registry: 39,339 feedback records over 30 days. On Base, 65.5% of rated agents have 90%+ of their reputation written by a single address, and 11 addresses produce two-thirds of all feedback — one of them rated the same agent 6,688 times. Full code published so you can check every number.
19 August 2026 · AI agents
ERC-8004 has 170,000+ registered agents and 150,000+ feedback records. A peer-reviewed study measured coordinated Sybil behaviour among 73.5%, 59.2% and 90.6% of reviewers on Ethereum, BSC and Base. The standard anticipated this and expected someone to build the missing layer — the measurements are what it looks like when nobody has. We have no product here; we posted it as an open question.
17 August 2026 · AI security
Grey-market LLM access sells at 70–93% below list, and brokers deliver it by routing your requests through their own endpoint rather than handing over provider keys — where prompts, tool calls and code context can be logged and resold as training data. It takes one environment variable, and nothing in a normal review pipeline is looking at it. What we scan for, and the three things the scan cannot see.
15 August 2026 · WARN
The statute requires 60 days. In the last 12 months, 121 filings on the California and New York registers record less than that at sites of 50 or more workers — 14 of them record zero, 41 record 14 days or fewer, and 14,073 workers sit across the set. What the feed contains, what it costs, and the statutory exceptions that mean a short interval is a candidate rather than a conclusion.
15 August 2026 · Agent commerce
Our tenth agent registered at 17:18:29. Our eleventh, eleven seconds later, returned
Agent limit of 10 reached — a per-account ceiling we could not find documented
anywhere. Plus the one-word logging bug that hid the real error, cost a registration slot and
forced us to retract a forecast the same day.
14 August 2026 · Agent commerce
They are not competitors. MCP connects a model to your tool and gives you no identity, no metering and no payment; ACP moves money between two agents and puts escrow at the centre. Written from running both against the same data — including the ten jobs that settled for zero, and the buyer parameter that silently defaulted for weeks.
14 August 2026 · Agent commerce
Building the agent is the easy half. Registering it, proving it is genuinely sellable, taking money, and keeping it alive is the other half — told with real numbers from twenty agents, including the four that are live, the three that are registered with nothing to buy, and the credential that expired on a three-day timer nobody knew about.
14 August 2026 · WARN
A state WARN download is a notice log, not a lead list. Three properties of the raw data hide the filings worth reading: the notice-to-layoff gap, layoffs that only cross a threshold once aggregated across a 90-day window, and employers split across dozens of name strings. With counts from 3,997 California filings — and the reasons none of them are evidence on their own.
The posts here point at pages that are free and public.
California rebuilt to April 2023 and New York, with employer names resolved and county unemployment attached.
What this data answers that a state download does not, and how the paid delivery works.
Sources, refresh cadence, counting rules, and the coverage limits stated rather than hidden.