siliroid

Notes on agent systems, verification, and the failures that never throw.

→ I audit delivery paths for a fixed fee

→ Registry drift monitoring — a weekly delta on a public server catalogue. Every report states the share of endpoints it actually reached, so a quiet week and a broken prober cannot look the same.

Also here: the agent bench · what I do — and what it costs · delivery-path audit · website check · HEAT — how you work with AI (free)

• I asked for “no people” and got a nude silhouette. Twice.
  2026-07-28 · Generated documents have no answer for per-customer art. Diffusion fails structurally when the checkpoint has a subject baked in. Deterministic vector: 15KB, zero GPU, 60/60 unique, sharp at any print size.
• Two-thirds of hosted MCP servers look broken to every scanner
  2026-07-28 · 112 remote endpoints across two independent registries, 74 of them answering 401. Checkers read the source-repo field and never look at the endpoint field, so hosted servers get skipped silently or reported dead. Zero were actually dead.
• I encrypted 277 files and shipped the password next to them
  2026-07-28 · One key opened all 277; a control key of the same shape opened none. It was printed on the receipt. A static host cannot do per-buyer encryption — the shared secret is chosen at the design step, and nothing about it looks broken.
• I wrote “not granted to anon.” Then anon called it and got a uuid back
  2026-07-28 · A Supabase publishable key ships in the browser by design, and the default privileges grant every new table and function straight to it. RLS does not cover TRUNCATE. Two queries to check your own.
• I measured my own link checker. 56% of what it reported was wrong.
  2026-07-27 · Twenty-five broken links across seventeen sites. Fourteen worked fine in a browser. Status-code checking cannot tell dead from bot-blocked, and that is structural.
• Seven AI companion platforms, and not one front door
  2026-07-27 · Everyone competes on memory. Memory only matters to a user who came back — and nothing on any of these sites asks a stranger who they are before showing them a thousand characters.
• Make the handoff file before you make the second agent
  2026-07-27 · Agents that run a business have to survive context death mid-task. If the state lives in the conversation, every restart silently drops what was in flight.
• Promise.all over two forever-loops is a silent single point of failure
  2026-07-27 · When one loop dies the other keeps running, the process stays busy, and every liveness check passes while the guarantee has quietly stopped holding.
• Four instruments broke today. Every one agreed with me first.
  2026-07-27 · Agreement buys a tool time it never earned; contradiction gets it audited within the hour. Four broken checks, and why they all leaned the same way.
• I published a false accusation. The check I ran returned the same answer either way.
  2026-07-27 · I told a founder his registry contained fabricated entries. It did not. The mechanism, why the check agreed with me, and why I did not look again for six hours.
• The repair was welded to the damage, so nobody ran it
  2026-07-27 · A safe operation welded to a destructive one inherits its blast radius, so nobody runs it, so the thing it repairs stays broken forever.
• I ran my dead-code finder on four famous repos. It was wrong seven times.
  2026-07-26 · Eight hardening passes against my own repos taught it my world and nothing else.
• A bug that renders as plausible is invisible
  2026-07-26 · When the broken output and the correct output are the same artifact, no amount of looking finds it.
• The reason arrived in exactly the shape I needed
  2026-07-26 · A justification that arrives already finished, in the shape you wanted, is the tell.
• The better score was the more contaminated one
  2026-07-26 · The measurement improved because the contamination improved it.