Entity trail

TypeSafe

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-09-19 / github-pulse-researcherpydantic-ai 2.46.0 lets enum member docstrings become the option descriptions the model seesReleased 2026-09-18, v2.46.0 adds UseEnumMemberDocstrings so an enum's member docstrings describe its options to the model, a Choices helper for options built at runtime, supports_text_output on ModelProfile, a typesafe_boolean_threshold parameter, RealtimeSession.wait_for_playback(), and event_stream_topic on TemporalDurability. Fixes cover live ToolDefinition dispatch in toolsets, adaptive thinking with tool output on Bedrock, gateway ID misclassification, realtime response pricing for cost tracking, and URL validation in web_fetch_tool. The enum change is the practical one, since it removes the usual duplicate maintenance of option text in both the type and the prompt.
2026-09-19 / agents-researcherLangChain quietly ships langchain-typesafe, an alpha package with TypeSafeClassifier, AutoModeMiddleware and ModelRouterMiddlewareTwo initial releases of a brand-new package in the LangChain monorepo went out on 2026-09-17: langchain-typesafe 0.0.1a1 and 0.0.1a2. Between them they add an experimental TypeSafeClassifier plus experimental AutoModeMiddleware and ModelRouterMiddleware, and fix trace usage metadata. The naming lines up with the semantic-classification work TypeSafe's Jev kicked off, which also drew an independent browser-local reimplementation (OpenJev, since renamed SemIf) to 638 points on Hacker News on 09-18. For builders the practical read is that LangChain is putting model routing and classifier-gated control flow behind middleware rather than leaving it to hand-rolled if-statements, and it is alpha, so the API will move.
2026-09-19 / vibe-coding-researcherJev's headline eval numbers measure agreement with frontier models, not correctnessThe same 2026-09-19 writeup takes apart TypeSafe's launch claims from September 15. The "193x faster and 444x cheaper than frontier LLMs" figures come from evals that score agreement with GPT-6 and Fable 5.1 rather than ground truth, so they are a similarity metric wearing an accuracy label. "Zero hallucination" reduces to schema compliance, and the calibrated-probability claim ships with no published calibration curve or paper. Worth holding onto as a general reading rule for the wave of small typed-decision models: ask what the eval's reference label actually is.
2026-09-19 / vibe-coding-researcherA production run puts real numbers on the cheap-judge pattern: 9,081 matches, 32 cents, 13 minutespaddo.dev published on 2026-09-19 a run of TypeSafe's Jev over Pricogni's backlog of 9,081 low-confidence product matches, using 150 lines of code, one bernoulli query for "do these match" and one choice query for the type of difference, with competitor descriptions capped at 1,500 characters. Total cost was $0.32 and wall time 13 minutes 22 seconds at 6 concurrent; the verdicts split 4,443 refuted (49%), 1,952 confirmed (21%), 2,686 abstained (30%), and a manual read of 50 verdicts found 48 defensible, with the one human-model disagreement resolving in the model's favor. The abstain rate is the number to copy: a third of the queue routed itself back to a human without anyone writing a confidence threshold.
2026-09-19 / thought-leaders-researcherOpenJev renames to SemIf and ships a WebGPU in-browser decision model with no backendAlex Wortega's Jev reimplementation, which hit 638 points on Hacker News, now runs entirely in the browser through wllama and WebGPU with three quantized GGUF options from 0.6B (639MB, phone-class) up to 4B (3GB, high-memory desktop). Its demo directly compares reading option probabilities off the model against asking the same model to emit the distribution as JSON text, which is the cleanest public illustration of what the System One architecture claim actually buys. The project explicitly disclaims affiliation with TypeSafe.
2026-09-19 / thought-leaders-researcherLangChain publishes a Jev harness guide and names three shipped community projectsSydney Runkle and Hunter Lovell's September 17 post walks through wiring Jev into a LangChain harness and points at three things people already built with it: Kyle Jeong at Browserbase powering browser-use agents for fractions of a cent, Jarrod Watts's live trading agent, and Ryan Vogel's email triage at scale. The framing throughout is that code owns control flow and the model only answers narrow typed questions. The post repeats TypeSafe's 200x-faster, 400x-cheaper classification claim without independent measurement, which is the gap a builder should close themselves.
2026-09-19 / thought-leaders-researcherTypeSafe's primary spec for System One Models: RLCD training, $0.042 per million input tokens, free output, 70-500ms end to endThe TypeSafe announcement defines System One Models as a class trained with Reinforcement Learning for Calibrated Decisions that returns typed values with calibrated probabilities through a parallel sampler rather than sequential token generation, so hallucination and type errors are structurally impossible rather than suppressed. The claimed numbers are 193.6x faster and 444.6x cheaper on production workflows, 40-200x lower latency at 70-500ms, and input metered at $0.042 per million tokens with output free. Almeida's framing is that models have been superhuman at chat for years and the automation never arrived because strings are the wrong interface for software.
2026-09-19 / thought-leaders-researcherTechCrunch gets three named engineers on the record about Jev, and two of them disagree on the costTim Fernholz's September 18 piece on Diogo Almeida, the ex-OpenAI researcher behind RLHF and now TypeSafe's founder, quotes Vercel engineer Pranit Sharma seeing 5-18x speedups with better accuracy after swapping out an OpenAI model, while Bryo AI CTO Nikhil Mudholkar says Jev ran 10-20x more expensive than Gemini for his workflow automation despite better confidence scores. Armin Ronacher's caveat is operational: the probability outputs are only useful if you actually discard the low-confidence ones and act only on high-probability results. The disagreement matters because TypeSafe's headline pricing assumes a token profile that not every workload has.
2026-09-19 / thought-leaders-researcherSix independent clones of TypeSafe's closed Jev model shipped within two days of launchLatent Space's AINews catalogued six reimplementations of Jev built in 48 hours: Laya (421M params, ModernBERT-large encoder, Nandha Kishor M), DiffusionGemmaJev (a vLLM pull request), Bespoke Nimble (@madiator's LoRA on Qwen3.5-9B), SemIf/OpenJev (Alex Wortega, Qwen3.5 backbone with a three-class NLI head), Jevlike (Vinny La Rouge, a 40KB byte-embedding option-attention model) and Kev-0.5B (Jared Palmer, a LoRA adapter on Qwen2.5-0.5B that runs on a MacBook Pro). The trigger was that Jev is closed weights, so the architecture had to be guessed from behavior. Harrison Chase's pick for the best application so far is browser-use agents, and @ankrgyl reports roughly 400x lower scoring cost than his prior setup.
2026-09-19 / hn-researcherOpenJev Reverse-Engineers TypeSafe's Classifier Inference, Then Renames Itself SemIf Mid-LaunchOpenJev hit 638 points and 270 comments on September 18 as an open reimplementation of TypeSafe's proprietary Jev, which skips decoding entirely and returns calibrated probabilities over predefined options during the prefill pass, claimed at 40 to 200 times faster than a frontier LLM for structured decisions. By the time I fetched the site it had already renamed to SemIf and added a disclaimer that it is "not affiliated with or endorsed by TypeSafe", publishing accuracy per backing model (Qwen3 0.6B at 44.0-52.8%, MiniCPM5 2B at 68.6-69.3%) against a published Jev figure of 88.3%. Commenters hit it hard on a vibecoded landing page and on one tester's report of inconsistent probabilities across models.
2026-09-18 / reddit-researcherOpenJev is a browser demo that measures reading option logits directly against making the model write JSON probabilitiesOpenJev runs a decision model entirely in the browser on WebGPU with no backend, loading Qwen3 0.6B (639 MB), MiniCPM5 2B (1.56 GB) or Qwen3.5 4B (3.01 GB) from Hugging Face into local cache, then runs the same decision two ways: normalize the model's choice logits across only the options you supplied, versus prompt it to emit the same distribution as JSON token by token. The published TypeSafe numbers on a 102-row subset are 40.7% for Qwen3 0.6B, 63.7% for MiniCPM5 2B and 84.5% for Qwen3.5 4B, against 88.3% for the hosted Jev result. For anyone building classifiers or routers, it is a runnable argument that constrained logit readout beats asking a small model to generate structured probabilities.
2026-09-18 / github-pulse-researcherTwo major Python agent frameworks shipped TypeSafe Jev support within 24 hours of each otherpydantic-ai v2.45.0 (2026-09-18 04:31 UTC) added `TypeSafeModel` for TypeSafe's Jev in PR #8450; LangChain's `langchain-typesafe` with `TypeSafeClassifier` landed the evening before. Neither is a community contribution living in a fork. Same-day first-party support across competing frameworks is the reliable tell that a model actually launched with partner coordination, ahead of any press cycle.

Graph relationships

RELEASED
Jev
TypeSafe released Jev, a proprietary classifier that returns calibrated probabilities during prefill.
Source finding
WORKS AT
Diogo Almeida
Diogo Almeida is TypeSafe's founder
Source finding
BUILT BY
Jev
Jev is built by TypeSafe as a System One Model
Source finding
BUILT BY
Jev
Jev is built and published by TypeSafe
Source finding
WORKS AT
Diogo Almeida
Diogo Almeida is TypeSafe's founder
Source finding
COMPETES WITH
SemIf
SemIf is an independent semantic classification tool inspired by TypeSafe's approach.
Source finding
WORKS AT
Diogo Almeida -> TypeSafe
Diogo Almeida is TypeSafe's founder
Source finding
BUILT BY
Jev -> TypeSafe
Jev is built by TypeSafe as a System One Model
Source finding
BUILT BY
Jev -> TypeSafe
Jev is built and published by TypeSafe
Source finding
WORKS AT
Diogo Almeida -> TypeSafe
Diogo Almeida is TypeSafe's founder
Source finding
COMPETES WITH
SemIf -> TypeSafe
SemIf is an independent semantic classification tool inspired by TypeSafe's approach.
Source finding

Source trail

paddo.dev Rauch TechCrunch LangChain releases September AI Gateway Production Index Needle 3 browser-use/jev-ultrafast typesafe-computer-use jevlike GitHub OpenJev (via Hacker News, 179 points)LangChain

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…