User repository on GitHub
Notices your Jev decisions tilting before anything visibly breaks
- Forks
- 1
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Integrations & Adapters
Tiltmeter
Tiltmeter records the probabilities Jev returns for every question your app asks and alerts when one moves: a version change behind jev-latest, drift away from baseline, too many answers sitting next to your threshold, or an estimated accuracy drop computed without labels.
- Stars
- 0
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- abe75ch
- Last verified
- Oct 2026
About this project
Notices your Jev decisions tilting before anything visibly breaks. Jev, TypeSafe's System One model, answers with probabilities instead of text. Tiltmeter records those probabilities for every question your app asks and alerts when one of them moves: Each alert names the likely cause: a model change, an edit to the question, or a change in your inputs. It fires once, stays quiet while it persists, and is reported again when it resolves.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Four alerts, version, drift, edge and accuracy, each naming the likely cause as a model change, an edit to the question, or a change in your inputs.
- It stores probabilities, the model name and a fingerprint of each question, and never the text you send to Jev.
- In Pydantic AI you pass tiltmeter.instrument() as the provider's http_client; answers come back unchanged and recording problems are logged, never raised.
- As a proxy it accepts only JSON and never adds an API key, so other programs and web pages cannot spend your quota through it.
Watch out
Four limits matter before you page anyone. The accuracy estimate assumes Jev's probabilities stay calibrated and is an early warning, not a substitute for labelled examples. Each window is compared with the one immediately before it, so a very slow steady slide can stay under the limits, and comparison against a pinned reference window is not built yet. A question needs at least 50 answers in each of two windows of up to 200 before it is checked, so alerts arrive late on low-traffic paths. The in-process wrapper only covers the async client that Pydantic AI and AsyncTypeSafeClient use, so the synchronous TypeSafeClient needs the proxy. Score questions get no accuracy estimate, storage is a single local SQLite file, and the demo and self-check run on simulated answers. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
It reads the confidence Jev already reports, using the probability of the option it chose and converting that into an expected-correctness figure. That is only meaningful if those probabilities are calibrated, which the README flags as an assumption, and it is why accuracy drops are framed as an early warning rather than a measurement. The estimate deliberately avoids TypeSafe's confidence field, which measures how concentrated the answers are, and score questions get none because a score is not right-or-wrong.
Not quickly. A question needs at least 50 answers in each of two windows of up to 200 before it is checked at all, and drift or accuracy alerts additionally require a real effect size and a p-value below 0.001, which also keeps questions with many options quiet. The upside is low noise: in simulation, unchanging traffic raised any alert in at most 0.3% of checks, while a moderate real shift was caught every time. The self-check is part of the install.
Probabilities, the model name and a fingerprint of each question, never the text you send to Jev. In Pydantic AI you hand TypeSafe's provider a Tiltmeter instrumented http_client and the agent receives the same answers, with any recording problem logged rather than raised into your app. As a proxy it passes status codes, headers and bodies through unchanged, returns a clear 502 or 504 when TypeSafe is unreachable, and callers keep sending their own API key.
That is the known blind spot. Each window is compared against the one immediately before it, so a very slow, steady slide can stay inside the limits indefinitely, and comparing against a pinned reference window is not built yet. The edge alert is the partial answer, since it fires when too many answers sit right next to the threshold your code acts on, catching the narrow band where a small shift actually changes behaviour. Thresholds are set per question or per project and question.
pip install the git URL for the core, which is standard library only, or the pydantic-ai extra for the wrapper. tiltmeter demo runs a simulated Jev that switches versions and fires the alerts, tiltmeter selftest is the offline check, and tiltmeter check gives a report against a live setup at any time. Data lives in one local SQLite file, which suits a single app or machine rather than a fleet.
version fires when the model behind jev-latest changed, drift when the spread of answers moved away from its baseline, edge when too many answers sit right next to the threshold your code acts on, and accuracy when the estimate fell. Each names its likely cause as a model change, an edit to the question, or a change in your inputs. Each fires once, stays quiet while it persists, and is reported again when it resolves.
It is an independent project, not made, endorsed or supported by TypeSafe AI. The README says it was written with an AI coding assistant and then reviewed, tested against real Jev calls, and covered by tiltmeter selftest. The licence is MIT, and the demo and self-check deliberately run on simulated answers rather than real Jev results.
More Jev resources like this
41GitHub stars
Typed Python planner runtime that turns batched Jev action distributions into executable probability trees and graphs for finite decision tasks.
Jev Integrations & Adapters#community#jevbest#catalog-import
Community
Chuf-H/jev-treePython
Chuf-H/jev-treePython25GitHub stars
Agent runtime with a terminal UI and a messaging gateway that keeps memory, self-built skills, session search, scheduling, and subagents across Telegram, Discord, Slack, WhatsApp, Signal, and CLI.
Jev Integrations & Adapters#community#jevbest#catalog-import
Community
wundercorp/lokiPython
wundercorp/lokiPython51GitHub stars
GraphRAG layer that runs a 4-phase ingest-to-answer pipeline over Neo4j, Memgraph, AGE or Kuzu using Laya or Jev decisions instead of per-hop LLM calls.
Jev Integrations & Adapters#community#jevbest#catalog-import
23GitHub stars
Memory-maintenance library that checks every stored memory against each new event with Jev yes-or-no votes, then marks stale facts superseded or contradicted instead of editing them.
Jev Integrations & Adapters#community#jevbest#catalog-import
Community
chopratejas/invalidatePython
chopratejas/invalidatePython14GitHub stars
Python library for schema-enforced JSON extraction into Pydantic models, with provider-native JSON mode fallback, auto-repair, token and cost tracking, plus agents, RAG, and eval helpers.
Jev Integrations & Adapters#community#jevbest#catalog-import
Community
jhd3197/PrompturePython
jhd3197/PrompturePython30GitHub stars
Experimental language where questions and methods are composable values, with standalone .jpp sources run by a native Rust kernel and a Python implementation kept as reference.
Jev Integrations & Adapters#community#jevbest#catalog-import
Community
Towow-ai/jppRust
Towow-ai/jppRustFrom the Jev community
Posts from builders shipping with Jev right now.