Decisions API vs Jev vs Microsoft Decision 1, and what Jev integrations do with no key: research of 13 October 2026

This dated report records the evidence behind release regular-2026-10-13-decision-api-comparison of 13 October 2026. It adds one page, Decisions API vs Jev (TypeSafe AI) vs Microsoft Decision 1, with Cloudflare Clef, and refreshes When Jev fails (no-key and bypass cells for model routing and content moderation), benchmarks (B42 graded B: 38 → 39 graded rows) and Jev alternatives (hosted rows rechecked; 33 substitutes, unchanged). The pool ran nothing: no harness, package, benchmark or API call was run.

Run: run:e361615f-595f-4bb6-a7c8-7168d4d370a3 (scheduled regular research run 15) · Pool: pool_jev_catalog · Runner: one runner plus two research helpers (A: no-key and bypass cells; B: hosted decision-API pages). Checks: 13 Oct 2026; runner started 05:17:11 UTC; home status written 05:20:22 UTC; helper B done 05:24 UTC, helper A done 05:28 UTC; runner done 05:37:09 UTC; manager decisions 05:40 UTC.

The run's research notes and its source ledger (115 records, R15-S… IDs) are held in the pool's private record. Key sources are linked below and on the pages each finding feeds. Pre-return checks passed: 131 cited IDs checked, 0 missing. 15 count-word sentences were checked and all 15 pass: every category list adds up to its total, 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, OpenAI, Microsoft, Cloudflare or Subquadratic call was made, and the pool has no account with any of them. Failure cells come only from source at pinned commits. Study results are the operators' own measurements. Nothing in this run is Tested.

Previous report: 12 Oct research report. All dated reports: Research.

1. What was asked

• Decision APIs side by side (protected page). Build a page that puts OpenAI's Decisions API, Microsoft-Decision-1 and the Cloudflare Clef family beside Jev: graded tests, price, inputs, request shape, data terms, errors and limits, each cell dated and sourced, with a shareable card and a copy block.
• When Jev fails, first half of the no-key and bypass pass. Settle the open no-key and bypass cells for the model-routing and content-moderation rows at their published pins, and recount.
• Comparator review (new standing step). Check every decision model that a benchmarks row names as a comparator against the alternatives rows.
• Leads from 12 Oct. Recheck the pin of etsabary/luna-decisions-benchmark and decide its grade; screen utakatano/decision-models-demo.
• Standing work. Access and notice recheck, the daily discovery scan with the 72-hour decision-API look-back, and the drift registry.

2. Summary answer, as of 13 October 2026

• Access is unchanged (13 Oct 2026, 05:18 UTC). The Vercel provider notice retired at the end of its shelf life; it was never shown on Home (section 3).
• The decision-API comparison page ships, with the key finding resting on B38, B39 and B41 (section 4).
• B42 is graded B: 39 graded rows; 4 graded comparisons with the Decisions API (section 5).
• Microsoft-Decision-1 is listed as Preview, not GA, in Foundry (section 6).
• When Jev fails: no key not recorded 18 → 8; bypass recorded 44 → 52; malformed fail-closed 34 → 35 (section 7).
• Comparator review: 9 comparator entries; 2 have no row and are queued for run 16 (section 8).

3. Access status and home notices

Checked 13 Oct 2026, 05:18:51–05:25:30 UTC: Documented. Status unchanged since 27 Sep 2026, 22:33:25 UTC (TypeSafe CEO post, re-read through X's public oEmbed endpoint; R15-S03).

• Signups open, no free credit for new accounts. Service operational ("All services are online", status page "Last updated on Oct 13, 2026 at 5:18am UTC"; R15-S05).
• Price ($0.042 per million input tokens; output free), limits (100K tokens per second, 80 requests per second), the model jev-1.13.0, its aliases and all six route IDs are unchanged (R15-S06, R15-S12, R15-S36 to R15-S39).
• Shown on Home (2): new-user credit disabled (rank 1); lookalike resellers (rank 2). Hidden (2): Jack & Jill case study (retires 14 Oct), Series A (16 Oct). Retired (1): Vercel adding TypeSafe as a second Jev provider; shelf life ended 13 Oct 2026, 05:15 UTC; checked 05:19 UTC; never shown on Home. The fact stays on Where to use. New: 0.
• Docs changed without an access effect: the quickstart was rewritten and llms.txt is retitled "TypeSafe" (R15-S04, R15-S07). Rejected as notices: those docs changes, the Decision-1 lifecycle change and the SubQ Snap launch (another provider's model is never a home-page notice).
• Jev was not listed in Microsoft Foundry's catalog or Amazon Bedrock's model cards when the pool checked on 13 Oct 2026, 05:20 UTC (R15-S41, R15-S42, R15-S45, R15-S46).
• Subprocessors were not rechecked (the trust portal needs a browser render); the 11 Oct value stands.

4. New page: Decisions API vs Jev vs Microsoft Decision 1

The page is Decisions API vs Jev (TypeSafe AI) vs Microsoft Decision 1, with Cloudflare Clef: graded tests, price, images and data terms. The full card, table, copy block and sources are on that page; this section records how it was built.

• Key finding (35 words, verbatim on the page): see Key findings above. It rests on B38, B39 and B41 only; each clause names its metric. Labelled Reported results, Documented materials.
• Title source. "jev vs openai" did not hold: Google showed it only as an echo of the typed prefix, and Bing and DuckDuckGo suggest "jev openai" instead. Google completes "decisions api" and "decisions api vs" to "decisions api vs jev", and "microsoft decision" to "microsoft decision 1" and "microsoft decision 1 vs jev" (13 Oct 2026, 05:21 UTC; R15-S27). A suggestion is a demand signal, not a search volume.
• Card (#dapi-card): six rows (graded tests, price as published, images, request shape, your data, documented errors and limits) across four columns (Jev; OpenAI Decisions API; Microsoft-Decision-1; Cloudflare Clef family). Each non-Jev column says it is another provider and where your data goes. Each cell links its dated cell in the 13-row table. Prices are quoted as published, per million tokens; there is no per-call arithmetic and no overall ranking. The copy block repeats the card word for word.
• Clef data wording (manager decision): your data goes to Cloudflare under Cloudflare's terms on Workers AI; on OpenRouter's non-Cloudflare hosts (PrimeIntellect, DeepInfra) it goes to that host (R15-S226).
• B42 added below the three in the graded block, with its 7 flags. The Jev and OpenAI "Graded tests" card cells carry one short B42 clause that names grade B (section 5).
• SubQ Snap not compared. Subquadratic launched SubQ Snap, a hosted decision API, on 12 Oct 2026 per the vendor post. Vendor-documented: $0.05 per 1M input tokens, output free, Public Preview with a possible waitlist, text and images. The vendor's claims against Jev are Reported only. Its docs and data terms were not read (R15-S33). It gets a dated "Also launched, not compared here" line on the page (#dapi-also-launched) and a "seen, not checked" entry on alternatives. No row, no fifth column, no home-page notice.
• Vendor claims (Microsoft's added Jev benchmarks; Cloudflare's "costs less than Jev") are Reported and attributed, outside the graded rows (R15-S209, R14-S405).

5. Benchmarks: B42 graded B (38 → 39)

etsabary/luna-decisions-benchmark at 7a5ea13. The latest commit on main was still the run-14 pin at 13 Oct 2026, 05:24 UTC (R15-S53), so the README, metric script, results summary and run manifest opened in run 14 stand (R14-S300 to R14-S305).

• Operator: GitHub account etsabary (independent; MIT licence). Task: 1,000 five-option deterministic logic questions in 25 families.
• Result as reported Reported: Decisions API (via OpenRouter's alpha Decisions endpoint) 612 of 1,000 (61.2%); Jev 817 of 1,000 (81.7%), with Jev's answers reused from a 22 Sep run.
• 7 flags: (1) via OpenRouter's alpha endpoint, not OpenAI direct; (2) Jev not re-run (22 Sep vs 10 Oct; mixed Jev providers); (3) answer key withheld (only its hash committed); (4) 17 persistent refusals (HTTP 502), cause unknown; (5) the operator built both benchmarks; (6) young repository (2 commits); (7) latency and cost not comparable across dates and routes.
• Ledger: 39 graded rows = A 11, B 13, C 8, D 7. Graded comparisons with OpenAI's Decisions API: 4 (B38, B39, B41, B42); with Microsoft-Decision-1: 1 (B41). The sentence "Against small text-generating LLMs, Jev scored ahead in 6 of 10 graded studies" is unchanged: B42 is not a small-LLM comparison.
• utakatano/decision-models-demo stays a lead, not graded: an empty repository at 13 Oct 2026, 05:22 UTC (R15-S227, R15-S228).

6. Microsoft-Decision-1: Foundry lifecycle and alternatives cells

• Lifecycle: listed as Preview in Foundry at 13 Oct 2026, 05:20 UTC (model page "Lifecycle Preview" and catalog card); listed Generally available on 12 Oct 2026, 05:15 UTC; cause unknown. No Microsoft statement was searched for or found (R15-S43, R15-S44). Where to use is corrected; the 12 Oct report and earlier release notes keep their dated wording.
• Alternatives rows rechecked (quotes on alternatives and the new page): Decision-1 requests-per-minute limits by tier, 60 to 900; no tokens-per-minute figure on that page at 13 Oct 2026, 05:20 UTC (R15-S206); Microsoft Product Terms read: Decision-1 is not named and Microsoft will "temporarily store" content, with no period (R15-S200); the catalog licence text (R15-S44); Vercel lists Decision-1 and the Decisions API (R15-S212). OpenAI: refusal behaviour and the 128-image limit (R15-S214); a 7-region table beside the guide's "US and Europe" (R15-S217). Clef: platform errors 408 and 429 (R15-S223); base-model licences (R15-S219 to R15-S221). Prices unchanged.

7. build/when-jev-fails: no-key and bypass pass

A script first reproduced the published splits from the page's cells (13 Oct 2026, 05:31 UTC). Helper A then settled all 18 assigned cells for the model-routing and content-moderation rows, each from a file and line range at the published pin (R15-S100 to R15-S130). The runner spot-checked 2 cells at the pins; both match (R15-S55).

• Model routing (6 rows): no key 3 fail-closed, 3 fallback; bypass 6 of 6 recorded. Content moderation (7 rows): no key 3 fail-closed, 4 fail-open; bypass 7 of 7 recorded. Both domains now have no not recorded no-key or bypass cell.
• LiteLLM bypass cell counts as recorded (manager decision): the complexity router's own bypass paths were read at the pin; router-wide hooks outside that module were not.
• agent-router malformed answer → fail-closed: nidhi-singh02/agent-router at fb24d06 exits with code 2 (decision-engine.ts L179, L196, L224 → run.ts L363; R15-S111, R15-S112).
• Mismatch qualifiers, text only (no split change): jev-router below threshold also allows upgrades capped at sonnet; autojev on named Jev routes falls to equal load balancing on error, malformed and below threshold; autojev's published bypass path is 1 of 10; jev_antispam_bot has 3 more bypass paths.
• Remaining not recorded. No key, 8: n8n 5, email 1, trading 1, coding agents 1. Bypass, 25: ticket triage 5, email 4, n8n 4, browser 4, judges 4, frameworks 3, Claude Code and MCP 1.
• The build hub said "71 real implementations"; it now says 77, as build/when-jev-fails has counted since 12 Oct.

8. Comparator review (new standing step)

A script extracted every decision model that a benchmarks ledger row names as a comparator and checked it against the 33 alternatives rows (R15-S56). Of 9 decision-model comparator entries: 5 have a row (Decisions API, Decision-1, Laya, Kev, Open-Jev), 2 are covered by a family row (Clef; Winnow through the Ollaya runtime row, a judgement), and 2 have no row (5 + 2 + 2 = 9). The 2 without a row, Cygnet (B07) and GLiNER2.5-Decide (B29), get no row this run because no primary source was read; they go to the "seen, not checked" list on alternatives and are queued for run 16 as leads. Text-generating LLMs, trained classifiers, rerankers and rules named in the ledger belong to other layers.

9. Scan, 72-hour look-back and drift

• Window: 12 Oct 2026, 05:17 UTC to 13 Oct 2026, 05:17 UTC. Run 14's cutoff was 05:12 UTC, so a 5-minute gap was backfilled on Hacker News and GitHub: 0 items found in the result pages read (R15-S25, R15-S28).
• 72-hour look-back from 10 Oct 2026, 05:17 UTC: SubQ Snap (Subquadratic) is new in the window, a hosted decision API launch (R15-S33). Decision 3.0, an open-weight family, and JevBench, a benchmark site, are new but not hosted launches (R15-S26, R15-S34). Decision-1's Foundry lifecycle changed and Vercel now lists it. None is a home-page notice.
• Comparison leads (benchmarks lead list, pins not taken): about 10 new repositories, among them dante01yoon/decision-model-arena, egarim/decision-1-vs-jev, GOhaike/decision-layer-benchmark and kprzybys/snap-vs-jev-swe (R15-S28).
• Drift: patches only. @ai-sdk/typesafe-ai 3.0.18, ai 7.0.138, @jkudish/jev-mcp 0.14.2, @jkudish/jev-agent-tools 0.2.1. No targeted diff is due (R15-S52).

10. Evidence gaps and not done

• Not found on the pages read (13 Oct 2026, 05:18–05:22 UTC): Decision-1's retention period in days, its data zones, a tokens-per-minute limit and a Microsoft price page; OpenAI /v1/decisions rate limits, context window, error codes, timeouts and retries, and any page dates; Clef's own rate limit.
• Subprocessors were not rechecked (the page needs a browser).
• Reddit returned HTTP 403; YouTube, TikTok and GitHub trending were not run. Those surfaces are unknown, not empty. The Hacker News "jev" search is typo-tolerant (4,601 hits); only the newest 100 were inspected.
• Not done (for run 16): the ticket-triage bypass cells (5); some existing cells helper A did not re-check (LiteLLM, JevRouter error, malformed and below threshold; soter malformed; jev_antispam_bot error and malformed); SubQ Snap's docs and data terms; Cygnet and GLiNER2.5-Decide; pins for the new comparison repositories.