v30.5, 2026-10-14: SITE QA, MONTHLY CHECKS, QUIETER ALERTS.

A 7-site sweep (TRG, Fancy Pet Salon, JMF, MySEODesk, Sarasota IV, Dr. Knopp, Apollo Office) became a skill and a schedule. (1) site-qa-sweep skill: a real Chromium check of every sitemap page at 320, 360, 390 and 414 plus a 300px desktop window, for heading skips, unsized images, redirecting sitemap URLs, broken images, sideways scroll and text spilling out of its box. Fixes ship with desktop pixel-identical to the clean build and pass Fable plus kimi_k3 plus Astra. (2) audit-headings-live.sh now runs after every deploy on 4 sites and is a written step on the 2 manual ones: 206 pages had skipped heading levels, all fixed. (3) Monthly site checks on Contabo 194, 4th of the month: 805 pages x 3 widths in about 17 minutes, one ntfy push only when something NEW breaks, report in Reports/site-checks. (4) Uptime alerts rebuilt after a Resend quota burn: state change plus a 6 hour reminder, 12 a day cap, idempotency key, ntfy push to the phone, and Uptime Kuma also on ntfy. (5) Monthly AI visibility replaces the Sun and Wed probes: one DataForSEO run on the 2nd for the 7 paid clients, about $3.70 a month, with a trend chart. (6) Jev: run against Claude Haiku 5.5 on 154 real CRM leads (Haiku caught 8 of 10 confirmed junk, Jev 3; Ryan labels the rest before Jev leaves shadow). The jev-seo audit tool was assessed: useful for content judgments, wrong on local keywords, competitors and grades, so prospect audits only. (7) Apollo CMS (EmDash 1.2) on its own subdomain with an MCP server for drafts. (8) Lesson written into the gate: never put the skip tag on a command that runs a deploy; it skipped Astra on 4 ships before the gate caught the fifth.

v30.4, 2026-09-29: CUTTING THE CHATGPT PLUS BILL.

Plus hit 70% with three days left, and a week of logs showed where it went: Astra reviews about 30%, codex_terra on the every-turn panel about 18%, redesign-live builds about 14%. Four changes. (1) The ship gate now REQUIRES gpt-6-astra only on live or client ships over 50 changed code lines or touching logins, payments, data scripts, secrets or deploy, and on every Life Gateway ship, and blocks those without it; everything else ships on Fable plus one cross-arch lane (kimi_k3 by default). (2) ollama_glm takes the second automatic seat back from codex_terra, which keeps every gated review. (3) Agentic repo coding moves to Sonnet 5.5 high; Codex stays for second opinions and stuck bugs. (4) The redesign judge runs Luna at low effort. Tested and rejected the same day: gpt-6.1-sol for the Astra seat (same catches as Astra at about twice the tokens on Plus) and gpt-6-luna / gpt-6-sol (fewer catches than 5.6).

v30.2, 2026-09-19: JEV, CORRECTED AGAINST WHAT WE ACTUALLY MEASURED.

Six fixes to the v30.1 Jev entry. (1) The 67.8% vs 67.9% figure is TypeSafe's own published benchmark against GPT-5.6 Terra; v30.1 called it "measured 09-18" and it was not measured here. (2) "Two free Ollama lanes tied it 400/400" was row-count parity on a fixture later found unable to rank lanes (identical wrong answers, jaccard 0.875), not an accuracy tie. (3) Latency is now measured, not vendor-stated: p50 0.61s per call, 120 decisions in 2.6s, 64 leads x 10 questions in 6.3s, and about 27x wall time on many questions against one record (21 questions, 3 records: 1.7s in 3 calls vs 45.8s in 63 calls on the local 14B). (4) "Not an accuracy gain over the free lanes" is half right. On single questions both it and a local 14B scored 108/108 on the held-out set; on the authored set 139/144 vs 133/144, no significant difference (exact McNemar p = 0.146). On typed CRM flags against the generative GPT extractor it was right where the profile was wrong 14 times and the reverse twice, across 20 contacts x 4 flags hand-labeled before reading either model, exact McNemar p = 0.0042 (flags from one contact are correlated, so read the p loosely). Both are now stated. (5) "No client content goes near it" is no longer the policy: scrubbed client text may go, with a zero-leak assertion over the whole batch before egress and the client told in writing that the scrubber, not a DPA, is the protection. (6) Added what the week taught: in both CRMs tested, form-template labels were read as facts (JMField "Monthly Volume" field, Life Gateway Calendly boilerplate) until the template was stripped; two real "SS" mentions scored 0.35 and 0.26 on a "Social Security" question, so the abbreviation goes in the question text; and TypeSafe's own system-one-adapter hides the question inside the JSON schema when structured_outputs=True on Ollama, dropping a 14B to 34/108. What Jev does daily today: nothing. The 09-19 audit of the repos, crons and hooks in this dashboard found no production path that calls it; the review-canary lane is opt-in and off, the ca-breach-monitor lane is committed and not deployed. Two integrations are proposed for shadow evaluation, not yet built: gating the Life Gateway miner's typed flags, and a junk lane on JMField intake (6 of 6 vendor pitches identified with no tuning, each confirmed by reading the note; false-positive rate unmeasured, no labels yet). Both are proposals until a shadow period produces labels.

v30.1, 2026-09-18: JEV ON THE ROSTER, A BROWSER LADDER WITH A CANDIDATE RUNG.

TypeSafe's Jev joins the model roster as a metered row, 35 models now. It is a classifier, not a generator: 0 for every generative capability, 3 for speed. Measured 09-18, it scored 67.8% against codex_terra's 67.9% and two free Ollama lanes tied it 400/400 at $0.0043, so it is not an accuracy gain over the free lanes (corrected in v30.2: the 67.8% is TypeSafe's own benchmark, not measured here, and the 400/400 fixture cannot rank lanes). What it offers instead is row-count integrity, which we did measure, and latency the vendor states at 70 to 500ms, which we did not. Vendor states it cannot count, do arithmetic, or compare dates, and TypeSafe has published no DPA, so no client content goes near it. A BROWSER_LANES block now shows the bot-wall ladder. browser-use is listed as rung 4 CANDIDATE only: its SOC 2, DPA and 81% anti-bot reach are vendor statements, unmeasured here, and its own jev-ultrafast agent went 0/3 on an unambiguous navigation target while reporting DONE. Patchright stays the locked default until rung 4 is measured. Residential proxies are $5/GB and on by default, which dominates the $0.02/hour browser time. Also fixed a hero count that read 33 models while the roster held 34; both pages now read 35, checked against the yml by hand (nothing computes the hero string).

v30, 2026-09-16: TWO ACCOUNTS, ASTRA ON PLUS, A MEASURED PANEL.

Claude Code now has a real profile switcher: claude = jesse Max 20x (own config dir ~/.claude-jesse, all MCPs symlinked), claude-jmf = admin@jmfield.com Max 5x; Orca panes stay jmf. Codex CLI 0.154 removed mcp-server, so mcp__codex__codex is dead: Codex work rides the codex plugin (/codex:rescue) or codex exec, and luna / terra / sol / gpt-6-astra all answer on the Plus plan at $0. Astra moved off the funded key and is now the FINAL cross-arch pass AFTER Fable on the ship gate (Fable on the 20x plan is the abundant one, Astra on Plus base is scarce, so it runs last, once). A 27-call gauntlet (3 diff sizes x 3 plants + decoy) reordered the panel: gemma4 9/9 at 4.5s takes seat 1, terra 2, glm 3, minimax 4 (9/9 but 48s on 13.5KB), mistral 5, qwen3.5 6 (new: Alibaba family, domain verify seat). luna off the panel (two OpenAI seats is an echo), Fable and Opus have cards, a full model roster page went up.

v29

BULK HTML MOVED TO GROQ. Eight lanes benched on one real landing-page spec (four Groq, four Ollama) with eight machine gates: Groq gpt-oss-20b 4.7s and gpt-oss-120b 8.8s, both 8/8, free tier; deepseek-v4.1-flash 12.5s on the flat Ollama plan; Kimi K2.7 passes only with thinking forced off and once returned nothing; glm-5.3-flash hit the token cap. Groq retired every Llama model plus Kimi K2 and qwen3-32b, so groq-mcp was repointed. Because gpt-oss is OpenAI-family, codex_terra is no longer cross-arch to HTML drafts; kimi_k3 on the ship gate is. kimi_k3 left the AUTOMATIC panel earlier (too slow for a 30s slice) and stays the manual cross-arch lane Fable consults; the panel table below reflects that. Cerebras key added, 402 until billing clears.

v28

THE GATED REVIEW CHAIN IS WIRED: the Ollama Pro lanes draft, then kimi_k3 → ollama_glm → ollama_mistral → ollama_minimax → codex_terra review, then Fable signs off. Five families, no per-call spend, same panel at internal, client and prod. The automatic hook runs seats 1 and 2; the rest fire on gated reviews and the ship gate. gpt-6-astra is manual only and never sits on an automatic panel. Moonshot, OpenCode paid, Gemini Pro and DeepSeek-direct are all unwired. Plan tiers differ: Claude Max 20x, Ollama basic Pro, ChatGPT basic Plus.