Failure terms aligned, three Claude Code guardrails read, and the products ledger rechecked (18 Oct 2026)

This dated report records the evidence behind the update of 18 October 2026. It compares the failure terms on 15 pages with the When Jev fails map, adds two Claude Code guardrails to that map and shows a third on the Claude Code and MCP page, and rechecks the products ledger. Nothing was run: no account, key, API call or third-party code.

Checks: 18 Oct 2026, 05:17–05:38 UTC (access 05:17–05:21; scan and drift 05:21–05:23; failure terms 05:24–05:31; Claude Code guardrails and products 05:32–05:36).

The research notes and the source ledger (R18-S… IDs) are held in the pool's private record. Key sources are linked below and on the pages each finding feeds. Checks on the sources: 155 cited IDs checked, 0 missing; 35 sentences that split a total were checked by script so that their parts add up and every category with 3 or more items (or 10% of the total) is named.

Rules kept: nothing was installed or run, so no Jev or other provider call was made, and the pool has no account with any of them. Failure cells come only from source at pinned commits. Vendor quality claims are Reported. Nothing in this report is Tested.

Previous report: 16 Oct research report. All dated reports: Research.

1. Failure terms: 15 pages compared with the map

A script matched every page row that shows a map row by repository and pinned commit, then compared each failure cell with the map's cell (R18-S63). Two rows pinned by npm version were compared by hand.

• No outcome differences: 0 cells on 15 pages give a different outcome from the map for the same condition and pin. Every difference is wording, colour or a column that groups outcomes.
• One pin difference, reported and not changed: the jkudish/jev-mcp row on the Claude Code page cites commit 86eae78; the map cites 53fe575 (5 cells).
• One correction, coding agents: the published key-finding split "On a Jev error the 13: 5 let the call run, 3 hold, 5 fall back" could not be rebuilt from the map or from the page's own column. The map's error cells for those 13 rows are 3 fail-closed, 9 fallback and 1 fail-open. The key finding now says so. A second line groups them by whether the tool call still runs: yes in 4 of 13, conditionally in 2, no in 3, and 4 are model-choice or routing rows (4 + 2 + 3 + 4 = 13). The split "24 read / 13 act on the answer / 10 do not / 1 unsettled" is kept as published.
• Coding agents, wording: 8 bypass cells now carry the term "recorded", as on the map, with their existing text as the detail. 4 hermes-jev-mana cells drop "(configured model)" from the term; it moves to the detail.
• Trading and fraud gates: 16 cells use the map term, with the rest moved to the detail line (for example "fallback-LLM, then fail-open" → fallback-LLM, "then fail-open if the LLM does not reject"). The 7 bypass cells say "recorded"; whether the order or payment goes ahead without Jev stays in the detail and in the last column. The key finding is kept. New secondary line: on a Jev error the 4 trading bots are 2 fail-closed and 2 fallback, and both fallbacks can still place a new order.
• Browser and computer use: 6 cells in the heat table use the map term. second-look is counted in the map's browser domain (8 rows); the browser page counts the 7 agents that click, type or tap and shows second-look separately because it only recommends. Both pages now say this; counts are unchanged.
• Left for a later update: wording-only cells on content moderation (11) and support-ticket triage (9), and a few on other pages. Their outcome words already match the map.

Term cells only; timeout-value cells are not compared. One more page, the dated 9 October evidence notes, holds two timeout values and was left as is (15 pages in all). Every row of the table sums: same + wording only (+ different pin) = cells.

2. Claude Code guardrails: three read, two join the map

Three Claude Code guardrails were read at pinned commits; their tool-call cells were read on 16 Oct 2026 and their bypass entry points on 18 Oct 2026. The latest commit on each repository equals the pin (18 Oct 2026, 05:20 UTC; R18-S200 to R18-S202). Documented.

• Why ohmyjev is not counted yet (editorial decision, 18 Oct): a bypass cell counts as recorded only when every entry point from one listing at the pin was read. For ohmyjev, every entry-point file was opened and the Claude Code module and Codex adapter were read through their tool-call paths, but its pi and omp adapters were read at their tool-call gates only. It is shown, with all its cells, on the Claude Code and MCP page.
• Map counts before → after (n 77 → 79; Claude Code / MCP guards 6 → 8):
  • No key: fail-closed 50 → 51, fail-open 7 → 8; now 51 / 17 / 8 / 3 / 0 (fail-closed / fallback / fail-open / advisory / not recorded).
  • Jev error: fail-closed 39 → 40, fail-open 8 → 9; now 40 / 22 / 9 / 5 / 3 (fail-closed / fallback / fail-open / no-decision / advisory).
  • Malformed answer: fail-closed 35 → 36, fail-open 11 → 12; now 36 / 24 / 12 / 4 / 3.
  • Below threshold: held 16 → 17, advisory 4 → 5; now 25 / 17 / 15 / 12 / 5 / 5 (no-threshold / held / acts-anyway / fallback / no-decision / advisory).
  • Bypass: 79 recorded, 0 not recorded; 13 "none found" (unchanged) and 66 with skip lists (64 → 66).
  • The 9 that let the action through on a Jev error: 4 moderation bots, 3 agent hooks (danna-zhou/jev-mcp's PreToolUse hook, hermes-jev jev-judge in enforce mode, claude-model-guard), a passage filter and an eval tool (4 + 3 + 1 + 1 = 9).
• New short answer on the Claude Code and MCP page: "We read ten Jev guardrails in their source code. Two hold or stop when Jev fails: LangChain AutoModeMiddleware stops the agent, and claude-jev-advisor's Windows delete guard keeps Claude Code's question for you. One, jev-permission-gate, falls back to Claude Code's built-in classifier. Two ship in a shadow or log-only mode, one is advisory by design, and three fail open: danna-zhou/jev-mcp, claude-model-guard and ohmyjev. One, gulbaki/jev-llm-guard, only returns a verdict: it is not counted as a decision. The official TypeSafe plugin makes no Jev calls at all." It replaces "only one of seven". The page lists ohmyjev among the three that fail open even though the map does not count it yet.
• Model IDs: ohmyjev pins jev-1.13.0, so it will not move when jev-latest is repointed; claude-model-guard and claude-jev-advisor send jev-latest (R18-S203 to R18-S230; cell reads R17-S236 to R17-S245).

3. Products ledger: 99 → 113 rows

A script first reproduced the published ledger (A 86, B 12, C 1 = 99). 14 rows were added, each with the Jev call read at a pinned commit (grade A): 3 Claude Code guardrails, 3 game and robot demos and 8 projects (3 + 3 + 8 = 14). Documented, 18 Oct 2026.

• Published claims rechecked 18 Oct 2026, 05:19 UTC Reported: 7 claims. Online with the same text: 5 of 7 (Brizz, Polylane, the Vercel CEO post, the Bryo AI post, FundaAI). Not rechecked: 2 of 7, the two figures relayed by VentureBeat (no direct link; VentureBeat returned HTTP 429). Changed: 0 of 7 (5 + 2 + 0 = 7) (R18-S103 to R18-S111).
• LangChain AutoModeMiddleware: the Python package is a counted row and its JS twin is not; kept as published, and the products page states the reason in one line.
• Directory counts (18 Oct 2026, 05:19 UTC): Made with Jev 855 builds; typesafeai.app 163 records ("Zero records have been independently reproduced"); JevCases 44 indexed cases; Jev in the Wild 2,170 (unchanged). These are listings, not adoption (R18-S112 to R18-S118).
• Two repositories read for this update did not qualify and stay leads: one had no Jev call at the pin and one did not mention Jev.

4. Access, aliases and routes

Checked 18 Oct 2026, 05:17–05:21 UTC: Documented. The access state is unchanged.

• Signups open; new accounts get no free credit; API operational. Price $0.042 per million input tokens for jev-1.14.0 and jev-1.13.0; rate limits "set per organization" (R18-S01 to R18-S05).
• Aliases: jev-latest → jev-1.13.0; jev-preview → jev-1.14.0. TypeSafe's models page is byte-identical to 16 Oct. The 14 Oct post still says jev-latest moves "in a week"; no move date and no 1.13 retirement date are published (R18-S05, R18-S09).
• Status page: no new event; the 90-day API uptime figure reads 99.826% (99.840% on 16 Oct). The page gives no cause (R18-S04).
• Home notices: 4 shown, none retired: no credit for new users, the Jev 1.14 preview, lookalike resellers, and the experimental features.
• Opper: now lists typesafe/jev-1.14.0 (alias typesafe/jev-preview, 64,000-token context, $0.042 per million input tokens). Its typesafe/jev-latest alias still points to 1.13.0 (05:20 UTC; R18-S32, R18-S33). Opper is the first of the five watched catalogs to list 1.14.
• Jev 1.14 was not listed on Cloudflare, Amazon Bedrock, Microsoft Foundry or Workers AI at 18 Oct 2026, 05:20 UTC (R18-S30, R18-S31, R18-S35, R18-S36). OpenRouter, Vercel and Netlify values unchanged.

5. Small items, scan and drift

• Vercel AI Gateway, 32,000: Vercel lists a 32,000-token context window for all three Jev IDs. This equals TypeSafe's per-question limit (32k; requests default to 64k), but Vercel does not define the field, so the match is not confirmed (R18-S25 to R18-S29).
• @typesafe-ai/sdk 0.7.0: checked against the 0.7.0 package files (not installed); the failure path is unchanged from 0.6.0 (error classes, 2 retries on 408, 429 and 5xx, 10,000 ms per attempt, a missing key throws in the constructor, jev-latest default). Three smaller changes: input-validation errors now reject the returned promise instead of throwing at the call; request bodies of 1 KiB or more are zstd-compressed by default on Node 22.15+; the package now needs Node 22 (R18-S68). Shown on TypeScript and errors and rate limits.
• Drift notes, not re-audited: patches ai 7.0.143, @ai-sdk/typesafe-ai 3.0.21, deepeval 4.3.1 (JevEval files not rechecked) and n8n 2.44.3 (Jev path last read at 2.44.1); pre-release @openclaw/typesafe 2026.10.10-beta.2. Unchanged: typesafe-sdk 0.7.5, @jkudish/jev-mcp 0.14.3, @effect/ai-typesafe 4.0.5, pydantic-ai-slim 2.57.0, litellm 1.105.1 and the rest of the list (R18-S57).
• Launched, not yet evaluated (72-hour look-back): Liquid AI d1-ultra, $0.042 per million input tokens on its launch page; Liquid says it outperforms Jev 1.14 "across all tasks" (vendor claim, Reported; R18-S50). Mouse-1, $0.10 per million input tokens on its launch page (R18-S51). Terms, data handling and API shape were not read. One dated line on alternatives; no rows or comparison columns.
• 72-hour look-back result (from 15 Oct 2026, 05:16 UTC): new: Liquid AI d1-ultra and a tazo labs audio decision model (API, price and terms not read; R18-S52); older but newly seen: Mouse-1 (post dated 13 Oct). Already recorded, no change found: the Decision Index, SubQ Snap, OpenAI's Decisions API (still public beta; R18-S55). Eevo's page was read (MCP, API and console; "around $0.20 per decision"); its terms and data handling were not found. None is a Home notice.
• Copies under many accounts: identical repositories (for example quicksilver, jev-browser, jevmem) were created under many accounts in the scan window, so GitHub counts for this window look inflated. Judged from names and descriptions only (R18-S42).

6. Evidence gaps

• ohmyjev: the pi and omp adapters' other handlers (output screen, done-check, routing, compaction) were not read line by line, so its bypass cell is not counted as recorded.
• jev-content-mod: the HTTP request is inside the third-party jevlang package, which was not read. The row is graded A on its call site in the repository.
• Vercel's 32,000 field: not defined by Vercel, so its match with the per-question limit is not confirmed.
• The two VentureBeat-relayed claims were not rechecked; newer X statements cannot be checked without X search. A DEV.to post claiming "35% of paid-for groups in Vercel" was not opened (Unverified).
• d1-ultra, Mouse-1, the tazo labs model and Eevo: launch pages only; terms, data handling and API shape unknown.
• The status page gives no cause for its uptime change.
• Wording-only cells on content moderation (11) and support-ticket triage (9) are left for a later update.
• Reddit returned HTTP 403 and X has no search access; YouTube, TikTok, GitHub trending and GitHub code search were not covered. GitHub result lists over 100 items cover the scan window only partly. Those surfaces are unknown, not empty.
• Subprocessors were not rechecked (the trust portal needs a browser render).