{"path":"research/jev-and-the-judgement-layer.md","content":"# Jev, and which parts of the pipeline are decisions rather than writing\n\n**Date**: 2026-09-22, widened 2026-09-26 (§ 5) and 2026-09-29 (§ 6, the relevance judge), first measured against the model itself 2026-10-04 (§ 3f, § 6e) · **Type**: research, on founder instruction — *\"look into which parts of the\npipeline could use Jev, the new super quick and cheap RLCD-trained AI. Research extensively online\nand document findings.\"* · **Status**: findings researched and cited; every fit proposal below is\n**unruled**.\n\n---\n\n## 1. What it is, from the sources rather than from memory\n\n**Jev** is a proprietary model from **TypeSafe AI**, limited early access from **15 September 2026**\n— six days before this was written. It is not a language model in the usual sense: **it does not\ngenerate natural-language text at all.** Its outputs are meant to be read by software, not people.\n\n**Three primitives, and they are the whole surface:**\n\n| Primitive | Returns |\n|---|---|\n| **Choice** | one option from a defined set, **with a probability per option** |\n| **Score** | a rating against ordered levels, with a probability per level |\n| yes/no | a probability between 0 and 1 |\n\n**RLCD** is *Reinforcement Learning for Calibrated Decisions*: probabilities optimised **against\noutcomes** rather than against human rater preference. The claim is that a higher-confidence\ndecision is more likely to be right — which, if it held, is a usable signal for when to automate.\n\n**The numbers it claims**: 70–500 ms; 40–200× faster and 40–400× cheaper than frontier models on\ncomparable tasks; **$0.042 per million input tokens with output free**, about **$0.0004 per\ndecision**. TypeSafe itself says reported gains *\"likely sit at the high end of real-world results\"*.\n\n**What is not disclosed**: the context window, the architecture, the weights. There is no independent\ntechnical paper.\n\n---\n\n## 2. The independent results, which matter more than the claims\n\n### The finding that is this project's own thesis, measured on a model\n\nAn independent benchmark ran Jev against Claude Haiku 4.5 on 2,000 synthetic phishing emails.\n\n- **Asked once** — *\"is this phishing?\"* — **Jev 62.6%, Haiku 81.3%** (significant, p < 0.0001).\n- **Split into five atomic questions** about links and sender patterns, then combined — **Jev 95.0%,\n  Haiku 93.2%**, and the gap between them no longer significant.\n\n**Splitting helped only Jev.** Haiku's best single signal (94.2%) beat its own five-question\ncomposite. So this is not \"decomposition helps everyone\"; it is **decomposition converting a weak\ndecision-model into a strong one**, which is a sharper claim and the one closer to ours.\n\n**Deliberus should notice this rather than merely enjoy it.** It is external, independent evidence\nthat breaking a judgement into atomic sub-judgements improves the judgement — the same operation\nthis project claims for human reasoning, measured on a machine, by someone with no stake in our\nthesis.\n\n### The calibration results, which are worse than the pitch\n\n- Out of distribution, **expected calibration error 0.107, against a 0.024 noise floor** — 4.4×.\n- The error **reverses by question type**: yes/no answers underconfident, choice and score answers\n  overconfident.\n- On **unknowable** tasks it was right 44.7% of the time while assigning its answers an average\n  probability of **0.74**. A model confidently wrong on the unanswerable is precisely the failure\n  this corpus watches for.\n- A pre-registered evaluation found accuracy **flat from 0.50 to 0.95 confidence**, then jumping to\n  100% at 0.99 — and that top bin was 60.2% of traffic.\n\n**The operative conclusion from the independent write-up is one line: *calibrate per question, not\nper model.* No single threshold transfers across decision types.**\n\n---\n\n## 3. Where it would fit here, and where it cannot\n\n**Read the pipeline's own shape first.** Measured 2026-09-22, the extraction and judgement layer\nmakes 23 model calls through `llm_call`, and the majority return a **closed enum with a\nconfidence** — which is Jev's Choice primitive exactly:\n\nscheme classification (~60 Walton options) · the relation menu in auto-connect · the seven-key\nsameness check · the decomposition-kind judgement · horizon reading · concept sense picking · the\nstance field · claim type · epistemic status. Two more are Score-shaped: decomposition readiness and\nthe looks-bundled ranking.\n\n**What it can never do here**: every pass that *writes* — splitting a claim, proposing a premise,\nrendering a question, synthesis. Those need text, and Jev produces none. So it is a candidate for\nroughly half the judgement layer and none of the authoring layer.\n\n### The sharpest fit, and it arrived tonight by coincidence\n\n**The relatedness gate ruled 2026-09-22** — *a machine's no-answer belongs in a gate, never on the\nmenu* — creates exactly Jev's third primitive: **a yes/no with a calibrated probability, asked\nbefore an expensive call.** Every gate of that shape is a filter that could run at 70–500 ms and\n$0.0004 in front of the model that does the real work.\n\nThat shape is also the **cheapest possible trial**, because it needs no labelled data to start: run\nthe gate in shadow against the current pass and measure agreement. The corpus already has the\nmethodology from the independent write-up — pin the version, shadow 1,000–2,000 records, fit\ncalibration on half and score the other half, compare cost per correct decision.\n\n### Four honest blockers\n\n1. **Calibration must be fitted per question on labelled data, and we have almost none.** The\n   sameness sitting is built and unsat; the weighing ground truth is 72 sentences. **The labelled\n   data this needs is the same data three other open items need**, which is an argument for the\n   sitting rather than against Jev.\n2. **The context window is undisclosed.** Our pair judgements carry two full claims plus their\n   context, and the decomposition judgement needs both texts in full. Unknown until tested.\n3. **Confidently wrong on the unanswerable** is the measured failure, and this corpus deliberately\n   asks unanswerable questions — a horizon reading, a does-not-fit. A model that returns 0.74 on\n   *nobody can tell* would corrupt exactly the confession channels built to be honest.\n4. **The trust ladder governs it like any other pass.** It would enter as proposing, not writing.\n\n### The one gain that is not about speed\n\nDeliberus's honesty checkers run on the same model family they audit — **same-hand bias**, named in\nthe residual-error taxonomy as one of the least-defended classes, and the corpus wants a **two-family\njury** as the cure. A decision model trained on a different objective entirely is a more independent\nsecond opinion than another frontier chat model would be, and jury disagreement across families is\nthe best cheap fragility predictor the literature offers.\n\n---\n\n## 3b. Verified 2026-09-22 by trying, not by assuming\n\n**Reachability.** `api.typesafe.ai` answers **403** — real, live, gated on a key.\nNo `TYPESAFE_API_KEY` or `JEV_API_KEY` in this environment.\n\n⚠ **CORRECTED 2026-09-22 — the first version of this line said \"OpenRouter lists 443 models and\nzero TypeSafe, so the OpenRouter route does not exist today\". That was wrong, and the way it was\nwrong is worth keeping**: the check read `/api/v1/models`, which is the **text-generation**\ncatalogue, and Jev generates no text. Its modality is `text->decisions` and it lives behind a\nseparate **Decisions API**. Verified live the same day:\n\n- `GET /api/v1/models/typesafe/jev-1.13/endpoints` returns the real model, served by TypeSafe at\n  $0.042/M input and free output, listed **18 September 2026**.\n- `POST https://openrouter.ai/api/alpha/decisions` answers **401** without a key — live and gated.\n- `~typesafe/jev-latest` is the rolling alias.\n- **Vercel AI Gateway** added the same three paths on 21 September (`typesafe-ai/jev`, with a plain\n  `POST /v1/evaluate`).\n\n**So the blocker is an OpenRouter key rather than a TypeSafe account**, which is self-serve prepaid\nwith no waitlist and bills alongside everything else. The lesson generalises past this model: a\nmodel whose output is not text will not appear in a text-model listing, and an absence found in the\nwrong catalogue reads exactly like an absence in the world.\n\n**Access is self-serve and needs an account**: *\"Get your API key from the dashboard\"*\n(`console.typesafe.ai/keys`). There is **no waitlist and no signup form on the marketing site** —\nthe call to action links to job postings, and the only other route is `hello@typesafe.ai`. The\nfounder authorised account creation 2026-09-22. **Since 2026-10-04 a key exists**, through OpenRouter, and the model has been measured on nine real uses outside this corpus; the method note in § 3e carries what transfers. Before that date the text read: creating it remains his single step, after which\neverything below runs.\n\n**Still undisclosed after reading the full public docs** (`docs.typesafe.ai/llms.txt`, ~40 pages):\n**the context window, the rate limits, and whether inputs are retained or trained on.** The first is\nthe gate on our pair judgements; the third matters before any claim text leaves the machine.\n\n**Their own named patterns are our builds, which is worth noticing rather than enjoying.** Three of\nthe four architectural patterns they publish are things this project arrived at independently:\n**confidence-gated routing** (*\"the answer tells you what; confidence tells you whether to act\"*) is\nthe relatedness gate ruled hours earlier; **composite scoring** (*\"break a complex judgment into\natomic scores, combine with weights you control in code\"*) is the atomic-versus-compound shape; and\n**speculative fan-out** is asking several questions in one call, which is what the concept backfill\nand the decomposition judgement both do. Convergence on architecture by a vendor with no stake in\nour thesis is mild evidence the shape is right, and no evidence at all that their model is.\n\n## 3c. The three probes, run against our own backend first\n\n`scripts/decision_model_bench.py`, fixtures copied out of the live graph with the query that\nproduced each one. Run 2026-09-22 on the Claude backend, 3.8 minutes:\n\n| Probe | Result |\n|---|---|\n| **The unanswerable** | abstained on **3 of 5**; where it committed, mean confidence **0.70** — against the published Jev figure of 0.74 while being right 44.7%. It committed on one typed `fittingness` claim and on one genuinely unknowable statement. |\n| **The relatedness gate** | **4 of 4** correct, mean confidence 0.81 — including the three pairs the old menu-shaped question got wrong. |\n| **Atomic versus compound** | compound **3 of 4**, atomic **4 of 4**. n is four; this is a shape check, not a measurement. |\n\n**The first row is a finding about us, not about Jev**, and nobody had asked it: our own classifier\ncommits an answer on undecidable claims about two times in five, at a confidence indistinguishable\nfrom the failure mode we were worried about in somebody else's model. The probe was built to\nscreen a vendor and screened the incumbent on the way past.\n\n⚠ **Corrected 2026-10-04 (§ 3f): that reading does not stand.** The five items were not fit for the\npurpose. Both models committed on exactly the items that have a defensible answer, and on eight\ncontested value claims copied from the graph our own model abstained every time.\n\n**And the gate result matters more than it looks.** Those three pairs are the exact shape that put\n84 bad decomposition edges in the graph, and the gate-shaped question gets all three right where the\nmenu-shaped one got all three wrong — which is the ruling of 2026-09-22 measured rather than argued.\n\n## 3d. Negation — the founder's hypothesis, tested against the sources (2026-09-22)\n\nFounder: *\"Isn't this the perfect judgment call to use Jev for? It should be able to handle\nnegation?\"* The sameness sitting's proposal is pure embedding cosine, and embeddings are\nstructurally polarity-blind (NevIR: nearly all retrieval systems at or below random on negation\npairs; ColBERT's max-similarity operator never attends to the negation token), so a decision model\nthat actually *reads* the two claims looks like the obvious patch.\n\n**The answer is no as posed, and TypeSafe says so itself.** Negation is on its own published\njaggedness page for `jev-1.13`:\n\n- *\"Double negatives and indirection degrade it\"* — reduce hops, point at the relevant state.\n- *\"Scoping words, negations and implied conditions are read literally.\"*\n- Its limitations doc lists literal interpretation among the named risks, and one independent\n  write-up states plainly that Jev *\"can still put an invoice in the wrong category, misread a\n  negation, or be influenced by adversarial content.\"*\n- The recommended eval set is told to include negations explicitly — i.e. this is a known\n  risk area you are expected to measure rather than assume.\n\n**But the two failure modes are different, and that is the useful part.** An embedding is\n*structurally* blind — the token contributes nothing, so a claim and its direct negation land on\ntop of each other roughly 100% of the time. Jev reads the text and can get a negation wrong. The\nfirst cannot be fixed by tuning; the second is a measurable error rate.\n\n**So the fit is a narrow yes/no in front of the cosine, not a seven-key replacement.** This is the\ndecomposition lesson from § 2 applied: the phishing bench went 62.6% → 95.0% when one fuzzy\njudgement became five atomic ones. *\"Do these two claims contradict each other?\"* is short input\nwith a crisp label, which is Jev's measured best regime (Enron spam 98.7, SST-2 95.7, AG News 91.3,\neach ahead of gpt-5.4-mini). A seven-way semantic-relation pick is the *other* regime — nearest to\nBanking77 at 76.0, its worst public set and the only one where both comparison models beat it.\n\n**And the standing verdict from every independent test is filter, not replacement**: *\"trust the\nends, and hand the middle to a model that can actually read the thing.\"* Measured, confident\nanswers (≥0.9, or ≤0.1) are right 90–99.6%, while the middle band is close to a coin flip.\n\n**One blocker from § 3b is now answered**: `jev-1.13.0` has a **64k token request budget, of which\n32k covers the state plus the longest question**, text only, 1,200 requests/minute. Our pair\njudgement is two claims — far inside that. **The account blocker is gone as of 2026-10-04** (a key exists); the other one stands: the calibration must be fitted per question on labelled\ndata, which is the sameness sitting's own output.\n\n## 3e. What nine measurements elsewhere say about using it here (2026-10-04)\n\nThe founder's other systems tried Jev on nine real uses the same day the key arrived. The numbers\nlive in a private document; what transfers to this corpus is the shape of the result:\n\n- **It worked wherever one short item met one concrete question**: are these two concerns the\n  same subject, is this message a correction, is this release the film that was asked for, is\n  this sentence a first-person factual claim. Accuracy 92 to 100 % on hand-labelled real cases,\n  Swedish included in four of them.\n- **It failed where the question was a judgment over a long transcript** (\"is anything here\n  worth saying?\"), in four shapes including the vendor's recommended one. The vendor names one\n  cause itself: *\"Accuracy falls as the state grows with content unrelated to the decision.\"*\n  A second cause showed when the input was varied instead of the question (2026-10-04): a small\n  state failed as much as a large one, and richer input did not help, because the reference\n  was a model's verdict that agrees with its own earlier verdict on the same text only about\n  three times in four. A decision model can be measured only against labels that are stable.\n- **Text similarity cannot tell \"the same item again\" from \"another item of the same kind\" when\n  the items share a template**, and a decision model given both texts can (2026-10-04, 158\n  pairs of short status sentences: similarity at 0.65 called 19 of 51 different items the same;\n  the decision model called none of them the same and still caught 106 of 107 true repeats).\n  The parallel here is claims that share a sentence frame: similarity nominates, a question\n  with both texts decides.\n- **Context that a person would need, it needs too.** Two names alone were judged no better than\n  an embedding; the same two names with their texts were judged well. For the sameness and\n  relatedness questions in this pipeline that means: send both claims in full, never the labels.\n- **A widened candidate set mattered more than a better question.** Letting it see pairs the\n  embedding had discarded found real matches; rewording the question found none.\n- **Two test rules**: label before the first call, and never put a test case in the criteria's\n  examples.\n\nNone of this is a ruling for the graph. It says the relatedness gate and the contradiction check\n(§ 3c, § 3d) are the right shape to try first, and that § 5's whole-document placements are not.\n\n## 3f. Asked of Jev itself (2026-10-04)\n\n**Founder**: *\"Go ahead with the Jev benchmark.\"* A key reached the model that day, and the script\ncould ask only our own backends, so a small client came first: `deliberus/decision_api.py` (three\nquestion shapes, several questions in one request, a refusal inside tests, and no answer rather\nthan a guess when the account is empty). **Nothing in the pipeline imports it.**\n`scripts/decision_model_bench.py --backend jev` and `scripts/decision_model_shadow_relevance.py`\ndo. The raw answers are in [`jev-and-the-judgement-layer/`](jev-and-the-judgement-layer/). In all,\nabout 490 requests, a third of a second each, about one cent.\n\n*Confident* below means a probability at or beyond 0.9 either way, the band in which § 3d and § 3e\nfound the model right almost every time.\n\n| Probe | Items | Result |\n|---|---|---|\n| The unanswerable, first five | the September items | these do not test what they were written for: see the correction below |\n| **Contested value claims** (new) | the eight value claims with the most live supports and attacks in the graph | **Jev took a side on 4 of 8** when offered *cannot be determined* (at 0.53 to 0.68), the same four in six runs and in both option orders. Asked a bare yes/no it sat at or beyond 0.7 on 3 of 8. **Claude, asked the same eight, abstained on all eight.** |\n| The relatedness gate | the September pairs: three unrelated, one related | right on three; on the fourth it answered 0.48 to 0.52 across six runs, so right in three runs of six. Never beyond 0.16 or 0.72, so never confident. Claude in September: 4 of 4. |\n| Atomic versus compound | the same four pairs | atomic 4 of 4 in all six runs; the one compound question 3 of 4 in five of them. The September shape, repeated. |\n| **The labelled decomposition links** (new) | all 190 links the judgement pass has labelled | confident on 52 (51 *related*, 1 *not*): right on 41 by the pass's own label, and on the other 11 by hand reading, one of them marginal (below). Its probability that the child bears on the parent orders the links much as the pass's relevance does (rank correlation +0.64). |\n| **Random pairs across sources** (new control) | 100 pairs, fixed seed | **confidently *not related* on 83**; 96 below 0.3; two at or above one half, and both are real pairs from the two sides of one debate. No confident *related*. |\n\n**⚠ A correction to § 3c: the first probe's five items were not fit for their purpose, so the\nSeptember reading of them does not stand.** That reading was *our own classifier commits an answer\non undecidable claims about two times in five*. Three things were wrong with the items, found when\nJev was first asked them:\n\n- The two typed claims are not the graph's wording (the live texts read differently), and each is a\n  claim *about* an open question or an assumption. *\"… is an open question … that no behavioural\n  test settles\"* is a sentence a careful reader may call true.\n- Of the three statements written as unknowable, one has a base rate (born on a Tuesday: one in\n  seven, so a firm *no* is calibrated) and one is arguably true (an unwritten integer whose digits\n  spell a haiku: integers are endless and almost none are written down).\n- Both models committed on exactly the items with a defensible answer. Claude, asked again on\n  2026-10-04, changed its answer on three of the five, and answered the Tuesday item *no* at 0.86,\n  which is the base rate.\n\nSo neither model was caught being confidently wrong there. The items were written for the probe\nand not copied from the graph, which is the failure the script's own header warned against. They\nare kept, with a note beside them, so that the September run stays comparable.\n\n**The contested value claims are the items that test what the probe was for.** They are copied\nfrom the graph by a stated rule: value claims with at least one live support and one live attack,\nthe eight with the most. No fact settles them, and this project maps the reasoning about them\nwithout ruling, so the answer its confession channels need from a machine is *cannot be\ndetermined*. The four Jev took a side on: that unequal application is no valid ground for\nabolishing the death penalty (*false*, 0.68); that stronger custody is the right response once a\nmodel crosses a danger threshold (*true*, 0.60); that Israel has a right of self-defence against\nHamas (*true*, 0.53 to 0.65); that arguments from maldistribution are fallacious (*false*, 0.58).\n**None of those is confident, so the rule *act only on the ends* would have passed none of them.**\nWhat the probe shows is narrower than the published failure and still decides the role: asked for\na verdict on contested ground, the decision model gives one, and a reasoned model declines.\n\n**The eleven confident disagreements, read by hand.** In each, the judgement pass had labelled the\nlink `does_not_fit`, and Jev answered *related* at 0.91 or more. Ten of the eleven pairs are\nplainly related: a restatement of the parent or of a piece of it (five), a broader version of it\n(two), the opposite claim (one), a consequence (one), a premise running the other way (one). The\neleventh is marginal: a programme, and a trend that cites it as an example. The pass's label does\nnot deny any of that. `does_not_fit` means *neither a part nor evidence for*, and its stored\nreasons say what each pair is instead. So the two were answering different questions, and on its\nown question the decision model was right.\n\n**What this says about the gate, which § 5d ranked first.** Two kinds of pair, two answers:\n\n- **Clearly unrelated pairs it discards confidently and cheaply**: 83 of 100, with no confident\n  mistake.\n- **Same-topic pairs are where the gate's work is, and there it is not confident.** The three pairs\n  that wrote bad decomposition links got 0.17 to 0.52, and one of them lands on either side of one\n  half from run to run.\n\nCandidate pairs reach `auto_connect` through similarity, so they are the second kind, and a gate in\nfront of the keeper would discard little. The first kind matters only if the candidate set is\nwidened past what similarity proposes. § 3e found exactly that elsewhere (*a widened candidate set\nmattered more than a better question*), and the two real pairs among the hundred random ones point\nthe same way here.\n\n**How steady it is.** Across six identical runs no probability moved by more than 0.06, and every\npick-one answer was the same in all six runs and in both option orders. Asking it twice is\ntherefore not a second opinion. A different question is.\n\n## 4. What would settle it, in order\n\n1. ~~**Ask what the context window is.**~~ **Answered** (§ 3d, § 5a): 64k tokens a request, 32k for state plus the longest question.\n2. ~~**Shadow the relatedness gate**~~ **Run 2026-10-04 on what is stored** (§ 3f): 190 labelled\n   decomposition links and 100 random pairs. Still unmeasured: the gate on the keeper's own\n   candidate pairs, because an *unrelated* verdict is never stored.\n3. ~~**Run the unanswerable probe**~~ **Run 2026-10-04** (§ 3f), on eight contested value claims\n   after the first five items proved unfit. It takes a side on four where Claude abstains on all\n   eight; none of its answers there is confident.\n4. Only then consider the enum-shaped judgements, per question, calibrated per question.\n\n**Nothing here is ruled.** Availability is limited early access, pricing is stated as possibly\nsubsidised, and the whole model is six days old at the time of writing.\n\n## 5. Every place it could go, by phase — the full sweep (2026-09-26)\n\n**Founder instruction**: *\"evaluate all opportunities to apply Jev to different places and phases in\ndifferent contexts and for different purposes.\"* Sections 3–3d looked at the extraction pipeline's\nenum-shaped calls. This section widens to the whole system — ingest, linking, housekeeping, reading,\ninstruments, the live session — and grades each place against one set of tests. **Every verdict\nbelow is a proposal, unruled.** The fleet-side counterpart (the founder's other systems) is kept out\nof this web-served doc and lives in the private dotfiles corpus.\n\n### 5a. Two facts that changed since § 3b\n\n- **Training and retention, read from the vendor's model page 2026-09-26**: *\"Jev is not trained on\n  customer requests or responses.\"* Zero retention exists only for enterprise customers under a\n  data processing agreement; the privacy policy allows retention *\"as reasonably necessary\"*, and\n  hosting is in the United States. Public graph claims are public anyway, so this bites only on\n  **compartment (blob) material and anything under `.private/`**, which stays off this route unless\n  a compartment's own membrane says otherwise.\n- **Language**: *\"English is the primary training language with best accuracy. Other languages …\n  are supported but less reliable — testing recommended before deployment.\"* A large share of what\n  this corpus will judge in the next months is Swedish (the election material, the compartments, the\n  wild weighing sentences). **Every Swedish-facing placement below carries a per-language\n  measurement as a precondition, not an afterthought.**\n\n§ 4's first step is therefore answered: the request budget is 64k tokens, 32k for state plus the\nlongest question.\n\n### 5b. The tests every placement is graded against\n\nA decision model returns a label and a probability and **no reason**. That one fact decides more\nthan speed or price does, because three of this project's disciplines assume a reason exists:\n*reasoning goes before the verdict it explains* (enforced by `tests/test_reason_before_verdict.py`),\n*propose, never assert* (a proposal a person weighs needs a stated ground), and *synthesizes only\ninspectably*. So the first test is about role, not accuracy.\n\n1. **Role.** Does the verdict *order, filter or flag* work that something else then does, or does it\n   *stand* as content a reader meets as the graph's judgment? The first is open to a reasonless\n   model. The second is not, whatever its accuracy, unless it falls under the ruled housekeeping carve\n   (a confident re-check of machine-written, unconfirmed structure, old link kept visible).\n2. **Shape.** A short input with a crisp two- or three-way label is its measured best regime (spam\n   98.7, sentiment 95.7). A many-way semantic pick is its worst (Banking77 76.0, behind both\n   comparison models). A fuzzy compound judgment should be split into atomic yes/no questions\n   first — the phishing result, 62.6% asked once against 95.0% split five ways.\n3. **Answerability.** Can the question be unanswerable? The measured failure is a confident answer\n   on the unknowable (0.74 while right 44.7%). ~~Our own classifier does the same two times in five\n   at 0.70 (§ 3c)~~ **Corrected 2026-10-04 (§ 3f)**: the items behind that figure were unfit. On\n   eight contested value claims our own model abstained every time and Jev took a side on four, so\n   this test separates the two. It is still worth running on every judge.\n4. **Error cost.** What does a false negative cost, and who ever sees it? A filter that drops a\n   true candidate makes the gap invisible — so every filter placement needs **an unscreened sample\n   that measures the screen**, or it becomes a guard that never fires.\n5. **Labels.** Can calibration be fitted *per question* on data we already hold? Accuracy is flat\n   from 0.50 to 0.95 confidence in the independent test, so only the extremes are usable until a\n   per-question fit exists.\n6. **Language** and 7. **membrane** (5a above).\n8. **Independence.** Is the value that a *different family* disagrees with the incumbent? That value\n   survives even when Jev is the weaker judge, because disagreement across families is the best cheap\n   fragility predictor measured (mean |ρ| 0.59, against 0.42 for repeat-consistency).\n\n**Proposed boundary, unruled — one sentence the founder may want to rule on:** *a decision model\nmay order, filter or flag the graph's work; its verdict never stands in the graph as a judgment a\nreader meets, except as the ruled housekeeping re-check.* The tell for a violation is a stored\nfield holding a Jev label with no reason beside it on a surface a reader sees.\n\n### 5c. The sweep, by phase\n\nVerdicts: **Trial** — cheap enough to shadow as soon as a key exists · **After labels** — worth it,\nblocked on a labelled set · **Park** — real but not binding now · **Refuse** — fails a test above.\n\n**Ingest — before any model reads a source**\n\n| Place | Today | Jev's question | Verdict |\n|---|---|---|---|\n| Injection guard (`extraction/injection_guard.py`) | six English regexes, propose-only flags, *\"publishing this list arms them\"* | yes/no per chunk: *is this text addressed to an automated reader rather than a human one?* | **Trial.** Flags only, so a false positive costs a note on the Source node. A model detector is not a published list, which answers the strategy-class limit the module states about itself. The detector reads attack text by design, but can only answer yes or no, so it cannot be steered into writing anything. |\n| Reported-speech split (`reported_speech.py`) | one call decides `is_reported_speech` and writes both halves | yes/no gate first; the writing call runs only on yes | **Trial.** The gate ruling of 2026-09-22 applied to a mixed call. Most claims are not reported speech, so most writing calls disappear. The false-negative rate is measured with the unscreened sample. |\n| Weighing detection (`weighing.py`) | lexical; 1 of 4 registers detectable, **0 of 9 Swedish** | yes/no: *does this sentence weigh one consideration against another?* | **After labels — and the labels exist.** The 72 blind-labelled wild Swedish sentences are exactly the per-question, per-language set test 5 asks for. This is the one Swedish placement with a ready measurement. The Jul-7 auto-open ruling was made under a deterministic detector, so a model-detected weighing would invite rather than open until precision is known (as the semantic-tier doc already leans). |\n| Sacredness brake | lexicon matched **0 of 1,678** claims | yes/no: *does this claim declare a value not open to trade-off?* | **Park.** The brake's surviving function after the 2026-08-25 re-scope is consent-and-timing choreography, and a detector for a declaration a strategic contributor can make at will is the detection-before-payoff-removal order the threat model warns against. |\n\n**Linking — deciding what relates to what**\n\n| Place | Today | Jev's question | Verdict |\n|---|---|---|---|\n| The relatedness gate in `auto_connect` | a required `related` field in the same call as the menu | the gate as its own yes/no call, in front of the keeper | **Trial — § 3's sharpest fit, unchanged.** The probe (§ 3c) shows the gate shape fixes the exact pairs that wrote 84 bad decomposition edges. **Measured 2026-10-04 (§ 3f): on those pairs Jev is not confident (0.17 to 0.52), so in front of the keeper it would discard little.** It discards clearly unrelated pairs well (83 of 100), which pays only if the candidate set is widened. |\n| **SIMILAR_TO auto-links** (`graph/linker.py` `auto_link_new_claims`, live on both extraction paths) | writes an edge on **cosine ≥ 0.80 alone**, before the keeper reads the pair, and nothing deletes it afterwards | — | **Withdrawn 2026-09-29 as a Jev placement.** First filed here as a contradiction check in front of the linker. The right fix is structural: the linker's write is the pre-ruling shape the sameness ordering (2026-09-13) exists to end, so the proposal is to **retire it and let the keeper alone write `SIMILAR_TO`** (`auto-connect-upgrade.md` § 2c, *The linker still writes*). A Jev check would have patched a step that should not exist. How many existing edges join opposing claims is still unmeasured. |\n| Candidate reranking before the keeper | cosine + concept route + structural signatures, top-k | Score: *how likely is it these two bear on each other?* | **Park.** The keeper call is where the money is; a reranker in front of it only pays if it cuts candidates, and cutting is a filter that needs test 4's sample. Worth it once the keeper runs on the full corpus, not before. |\n| The seven-key sameness pick | the keeper, one model call | — | **Refuse** (session 36 § 7): a seven-way semantic pick is the worst regime, and negation is on the vendor's own weakness page. |\n\n**Housekeeping — sweeps over what is already stored**\n\n| Place | Today | Jev's question | Verdict |\n|---|---|---|---|\n| The ruled confident re-check (*a machine-written, unconfirmed link that fails a confident re-check is superseded automatically*) | ruled 2026-09-13, re-check unbuilt | yes/no per edge: *does A actually support B?* | **After labels — and a fixture exists.** The **14 known-wrong SUPPORTS edges** are the first test: a re-check that passes them is inert. Only the extremes (≥ 0.9, ≤ 0.1) could act; the middle band goes to the triage feed. This is the one place where a Jev verdict would change the graph, and it is exactly the carve the boundary in 5b leaves open. |\n| Re-sweep ordering (`machine-work-as-a-standing-property.md` § 5, lever 5) | proposed, unbuilt: order re-sweeps by cross-family disagreement | the same question the incumbent pass asks, split to Jev's shape | **Trial.** The lever needs a second family, and at $0.0004 a decision the whole corpus costs cents. Its caveat is test 8's other side: disagreement with Jev partly measures Jev's own weak spots, so disagreement orders the work and never decides it. |\n| Decomposition kind (197 edges, measured twice: 76% not part-of (⚠ corrected 2026-10-07: 84 of those 197 were the automatic linker's premise links, stored backwards, which could not have been judged parts; among the other 113, 59% are not parts: `auto-connect-upgrade.md` § 2c)) | Sonnet with a `confident` flag | split: *is B a part of A?* then *is B evidence for A?* — two yes/no questions | **After labels.** The parts review produces human keep/reword/drop labels, which are the calibration set. The split is test 2: the current three-way-plus pick becomes two crisp questions. |\n| Concept backfill (tagging coverage **24.3%**) | one call per claim choosing senses | Choice among a term's recorded senses | **Park.** Senses are the concept layer's contested ground (\"claims are cores, senses are peripheries\"), the labels are model paraphrases, and a sense pick with no reason is the flattening the layer exists to prevent. The coverage gap is real; this is the wrong tool for it. |\n\n**Reading — what a visitor or an agent gets back**\n\n| Place | Today | Jev's question | Verdict |\n|---|---|---|---|\n| Retrieval for *Think with Deliberus* and the truth-graph answer (`truth_graph.retrieve_relevant_claims`) | cosine ranking | Score per retrieved claim: *does this claim bear on the question asked?* | **Trial — a directive currently served only by cosine.** *\"Conservative relevance beats semantic reach: tangential matches are worse than honest thinness\"* is a threshold decision, and a cosine score is the wrong instrument for it. A reranker only reorders and drops; the answer layer still cites and explains, so the reason-bearing surface is untouched. The drop rate needs the unscreened sample, since a dropped relevant claim makes the answer look thinner than the graph is. |\n| Head/tail citation balance | proposed, unbuilt | — | **Refuse for Jev.** It is a count over what was cited, and counting needs no model. |\n\n**Instruments — the system checking itself**\n\n| Place | Today | Jev's question | Verdict |\n|---|---|---|---|\n| **Amplification**, the named blind spot of disagreement preservation | `FLATTEN_THRESHOLD` 0.85 cosine measures collapse only; *\"a claim rendered with more conviction than its source passes the check cleanly\"* | Score, three levels: *compared with its source span, does the claim assert less, the same, or more certainty?* | **After labels — the most novel placement found.** It fills a gap the threat model names and no instrument covers, it is a telemetry reading rather than content on a claim, and a small hand-labelled set is cheap to make. Its own calibration test is Heilig's marker test: a judge that learned the concept has an optimum, one that learned hedging words is monotonic in their density. Run that before trusting it. |\n| The two-family jury | wanted since 2026-08-25; the Claude backend is the only second family built | Jev as a third family on the atomic questions | **Trial**, for independence rather than accuracy (test 8). |\n| The unanswerable probe (§ 3c) | built; ~~it caught our own classifier~~ its first five items were unfit, and eight contested value claims replaced them as the test (§ 3f) | — | **Keep running it on every judge, Jev or not.** On the eight, Claude abstains and Jev takes a side on four. |\n\n**Live session — the election format and the co-present rehearsal**\n\n| Place | Today | Jev's question | Verdict |\n|---|---|---|---|\n| Which spoken segments go to extraction | everything, ~27-min cycle | Choice per utterance: *new claim, restatement, objection, or procedure?* | **Park.** The cycle is bound by the writing passes, which Jev cannot take over, so a gate saves little until the reduced live profile exists. One property is worth writing down now: **a decision model cannot summarise**, so it structurally cannot commit the summarisation steering effect the DeepMind study found. Its failures are misclassification, which is visible, rather than restatement, which is not. |\n\n**Places it cannot go, restated so nobody reopens them**: every writing pass (splits, premises,\nquestion rendering, synthesis, short titles, the `does_not_fit` reason) · horizon readings, which are\n**read from the arrangement and use no model at all** (ruled 2026-09-15) · terminus proposals, which\nare contested ground and publish their reasoning · anything in the strength layer, which stays\ndeterministic · scoring contributors, which the bias literature killed empirically (no general\nbias-proneness factor) · any merge of value claims, which belongs to the sitting.\n\n### 5d. What this adds up to\n\n**Jev's best use here is not speed and not cost. It is a cheap, reasonless, different-family\nsecond opinion that can be put in front of work, behind work, or beside work — never in place of a\nreason.** The placements that pass every test share one shape: a yes/no that routes a more expensive\nor more consequential step, with the error measured by a sample the screen never touched.\n\n**Ranked by value against cost, as written on 2026-09-26, before any measurement** (what the first measurements change follows the list):\n\n1. The **relatedness gate** — yes/no, with the failure already measured. *(The SIMILAR_TO\n   contradiction check that shared this rank was withdrawn 2026-09-29; see the table above.)*\n2. **Beside the relevance judge** (§ 6, added 2026-09-29): the gate at link creation, the direction\n   census, and the veto on writes — all flag-or-withhold, none of them writing a strength.\n3. The **injection guard's second tier** and the **reported-speech gate** — ingest, both\n   flag-or-route, both nearly free.\n4. The **re-sweep instability index** — the one that changes the economics of every future\n   prompt fix.\n5. **Retrieval reranking** for the thin-honesty directive — worth a shadow trial, not a priority:\n   the corpus fit in one context window when last measured, so retrieval is not yet the bottleneck\n   (`structure-versus-scale.md`), and a wrongly dropped claim makes an answer look thinner than the\n   graph is.\n6. After labels: the **confident re-check** (fixture: the 14 edges), **decomposition kind** as two\n   questions (labels: the parts review), **Swedish weighing detection** (labels: the 72 sentences),\n   and the **amplification instrument** (labels: to be made).\n\n**What the first measurements change in this list (2026-10-04, § 3f and § 6e). Proposals, unruled:**\n\n- **The relatedness gate leaves first place** (ledger W6). In front of the keeper it would discard\n  little, because similarity already hands the keeper same-topic pairs, and on those Jev is not\n  confident. What it does well is discard clearly unrelated pairs, so its use is a **wide scan**:\n  pairs that similarity never proposes, most thrown away at a fraction of a cent per thousand, the\n  few it scores high handed to the keeper. Every pair among about two thousand substantive claims\n  is about two million requests, roughly 35 dollars at today's price, so a scan would start with\n  pairs across the sources of one debate.\n- **Beside the relevance judge** the veto behaved as designed on twelve links: it withheld two, and\n  both deserve a look (§ 6e). Its gate and direction answers were never confident there, so what\n  they give is a disagreement to read, never a decision.\n- **Ordering** has its first number: on 190 links its answer to *does this bear on that* ranks them\n  much as the reasoned pass's relevance does (+0.64). That is the *order* in *order, filter or\n  flag*.\n- **The boundary (W1) has evidence from our own claims**: on contested ground it gives a verdict\n  where a reasoned model declines.\n\n~~Nothing on this list needs code before the key exists.~~ The client exists since 2026-10-04\n(`deliberus/decision_api.py`), used by two scripts and by nothing in the pipeline. The Cypher count owed on SIMILAR_TO\nbelongs to the linker proposal now, and needs a local session, not a key. A read-only measurement of\n2026-09-29 already moved that proposal's cost: no high-similarity pair would lose its link, but the\nkeeper's pool admits only extracted claims, so minted reports must enter it before the linker stops\nwriting (`auto-connect-upgrade.md` § 2c).\n\n## 6. Jev and the relevance judge (2026-09-29)\n\n**Founder instruction**: *\"weave in how Jev could be employed for/as the relevance judge, think deeply\nabout that.\"* The judge (`deliberus/relevance_judge.py`, built 2026-09-28 under ledger Q33) is the\nmost Jev-shaped pass in the system: a no-answer gate asked first, then a direction, then a five-step\nordered scale. It is also the pass where a reasonless verdict would do the most damage, because its\nconfident reading **becomes the link's relevance and moves strengths**. **Everything below is a\nproposal, unruled.** It was written while what \"relevance\" means was still open\n(`strength-calculus-literature/bayesian-and-informal-logic.md` § 6, question 1).\n\n**Since 2026-09-30 that question has a for-now answer, and it changes two lines below** (ledger Q34,\n`relevance-and-the-two-factors.md` § 7). A reason that fails drops out, so relevance means **the gain**:\nhow much a reason adds if it holds. The judge's question asks exactly that. The objection in 6a.2 that it\nmeasures sufficiency rather than Δ therefore applies only to a subtracting rule. Its readings now wait on\nthe founder's approval of `--write` (it moves live strengths), not on a ruling about meaning. The\nconditionals in 6a.3 are still worth storing, for a different reason: if the starting-value path (ledger\nQ35) is taken, relevance becomes Δ again. A reason with starting value `s` then gains `(1 − s)·Δ` when shown\ntrue and loses `s·Δ` when shown false, so the necessity side is needed, and stored conditionals would keep\ntoday's readings valid.\n\n### 6a. What Jev brings that the current judge lacks\n\n1. **A distribution, not a switch.** The judge returns one step and a `confident` boolean, which is\n   the special case ρ ∈ {0, 1} of the reliability model the literature review derived:\n   E[relevance | reading] = ρ·rₘ + (1 − ρ)·r0 (§ 2.9). Jev's Score primitive returns **a probability\n   for each step**, so an expected relevance is Σ p(step)·value(step) directly, and the width of the\n   distribution says how much to trust it. That is the shape question 8 of § 6 asks for (discount a\n   machine reading by how reliable it is), and it arrives without a second call. **It is only worth\n   anything once calibrated**, and 5b's warning applies at full strength: out of distribution the\n   vendor's Score answers ran overconfident.\n2. **Cheap enough to ask the right number.** The ruled rule is Bayes-exact when relevance means\n   Δ = P(C | P) − P(C | ¬P) (§ 2.1), and today's prompt measures something closer to the direct change\n   and sufficiency, which is why its readings must not be written yet. Δ is two conditionals. The\n   literature says to ask them side by side, as verbal steps placed on numbers (van der Gaag et al.\n   1999), to ask for likelihoods rather than posteriors that need base rates (Li & Rui 2024), and to\n   check coherence with the complements (Zhu & Griffiths 2024). **A decision model makes that\n   affordable**: four Score questions per link — P(C | P), P(C | ¬P) and both complements — then Δ\n   computed in code. With a reasoned model that is four long calls per link; with Jev it is cents for\n   the whole graph.\n3. **Store the conditionals, not the relevance.** This half is independent of Jev and is the most\n   useful thing in this section. § 2.2 shows the portable number depends on how a target's reasons\n   combine: Δ for independent contributions, causal power for independent sufficient reasons, a\n   likelihood ratio for signs. **A stored pair (P(C | P), P(C | ¬P)) survives whichever the founder\n   rules**; a stored relevance number has to be re-swept after every ruling. The judge's readings\n   are already on hold for exactly that reason.\n4. **A third model family.** Most links were extracted by Gemini; the judge runs on Claude as the\n   second family. Jev, trained on a different objective, is the third. Disagreement across families\n   is the best cheap fragility predictor measured (mean |ρ| 0.59), so the Claude–Jev agreement rate\n   is the first available estimate of ρ per kind of link, until argued relevance claims exist to\n   measure it against.\n\n### 6b. Why it cannot simply BE the judge\n\n- **No reasons.** The judge writes `relevance_reason` for every reading, and that stored reason is\n  what a person argues against when they contest a link's relevance. A reasonless reading that moves\n  strength is the violation 5b's boundary names.\n- **Much of relevance is unanswerable.** On a value or weighing link, P(C | ¬P) often has no fact of\n  the matter, and the measured failure is a confident answer on the unknowable (0.74 while right\n  44.7%). Relevance readings on normative links would be exactly that failure, spread across the\n  strength layer.\n- **The instructive cases are its weak regime.** The two readings the second trial marked weak are\n  weak for reasons of scope: *\"most economists believe\"* is not the mechanism itself, and a\n  precondition is not a route. The vendor's own jaggedness page says scoping words and implied\n  conditions are read literally. **Pre-registered prediction**: on the second trial's 12 links, Jev\n  agrees with the Claude judge on most direction and gate answers, and reads at least one of the two\n  scope-driven *weak* links as strong or moderate. If it gets both right, this paragraph is wrong.\n- **English first.** The judge will soon read links extracted from Swedish sources.\n\n### 6c. The proposed arrangement — Jev gates, measures and vetoes; the reasoned judge writes\n\n| Stage | Who | What it does | Moves strength? |\n|---|---|---|---|\n| **At link creation** (a person's or a machine's) | Jev, one yes/no | *does the first claim bear on the second at all?* Stored as telemetry. A confident *no* puts the link in the triage feed | no |\n| **Census** over all argument links (509 on 2026-09-28) | Jev, direction yes/no | *does the first claim make the second more likely?* A confident answer against the filed polarity joins the judge's disputed filings | no — flags only |\n| **Conditionals** | Jev, four Score questions | P(C \\| P), P(C \\| ¬P) and complements, stored with their distributions and a coherence check | no |\n| **Reasoned reading** | the Claude judge, ordered by hinge price | reads the links where relevance would actually move a claim, with reasons, asked the same questions | only if the next row allows it |\n| **Write rule** | code | a reading is written only when Claude and Jev agree on gate and direction, and on magnitude within one step. Otherwise it is stored as disputed and goes to the triage feed | yes, carrying Claude's reasons |\n\n**Jev can only withhold, never add.** Every value that reaches the strength layer carries the\nreasoned judge's stated grounds, so 5b's boundary holds; Jev's role is a veto and a reliability\nestimate. That makes a machine write **rarer and safer**, which is the right direction for a pass\nthat moves strengths on a graph where almost nothing has been examined.\n\n**The at-creation gate touches the saboteur question** (ledger O13, open since 2026-09-19): a link a\nperson adds presumes relevance 0.5 under the build. A confident Jev *does not bear* could lower that\npresumption to 0 until someone argues for the link, which removes the payoff for bogus links rather\nthan detecting them. **That is a machine judgment changing the weight of a person's contribution, so\nit needs a ruling of its own**; as proposed above it is telemetry only.\n\n### 6d. What would settle it, in order\n\n1. ~~**The 12-link shadow run.**~~ **Run 2026-10-04: § 6e.** The registered prediction holds.\n2. **A founder sample of about forty links**, read by him, stratified by claim kind (empirical,\n   causal, normative). Zero disagreements in forty bounds the true rate near 7%; the strata say where\n   Jev can be trusted and where it cannot. It is the only ground truth the per-question calibration\n   can be fitted against.\n3. **The starting-value ruling (ledger Q35).** Relevance means the gain for now (Q34). If starting\n   values are adopted it becomes Δ, and the stored conditionals give every link its number without a new\n   sweep.\n\n### 6e. The shadow run, scored (2026-10-04)\n\n`scripts/decision_model_shadow_relevance.py`, five runs of the twelve links. The fixture is the\nsecond trial's own printed readings\n([`relevance-trial-2-2026-09-28.json`](jev-and-the-judgement-layer/relevance-trial-2-2026-09-28.json)),\nand the scoring rules were written into the script before the first call. The decision model was\nnot told how each link is filed. **Nothing was written to the graph.**\n\n| | Result |\n|---|---|\n| Gate | agrees with Claude on **12 of 12** (all bear). Its probabilities ran from 0.51 to 0.84: **none confident**, and three sat within 0.04 of one half. |\n| Direction | agrees with the filing on **11 of 12**, none confident. The one disagreement is the only attack link (below). |\n| How far | the same step as Claude on 7, within one step on 11. |\n| **The registered prediction (§ 6b)** | **holds.** Of the two links Claude read as weak for reasons of scope, Jev read one as strong (2.71, where 1 is weak and 3 strong) and the other as weak (1.34). |\n| Conditionals, asked as verbal scales | the difference between them has the filed sign on 10 of 12. Complements sum to within 0.09 of one on average and 0.32 at worst; the side that supposes the first claim false is the unsteady one. |\n| Conditionals, asked as a bare yes/no | the filed sign on 6 of 12. This shape is not usable. |\n| Steadiness | *how far* moved by at most 0.18 across five runs. |\n\n**Under the write rule of § 6c, ten of Claude's twelve readings would be written and two withheld**,\nand the two are the ones worth a look:\n\n- **The attack link.** *\"In expected-value reasoning, tiny probabilities multiplied by astronomical\n  value can dominate every other consideration\"*, filed against *\"A sufficiently large expected\n  future value can override the competing claims of people alive today.\"* Jev reads the first as\n  raising the second (0.79). So did the Gemini fallback in the first trial. Claude, told that the\n  link was filed as an attack, read a moderate attack. As worded, the first claim states the\n  mechanism the second needs. It is an objection only as a reductio, and the step that makes it one\n  (*and such domination is unacceptable*) is in no claim. The attack lives in a premise nobody wrote\n  down, which is the implicit crux again. Two of three model families read this link as support,\n  one of them told the filing and one not, so a disputed direction is worth reading as a pointer to\n  a missing premise before anyone rules on the filing.\n- **The scope link.** A claim about what minimum-wage laws do, filed for a claim about what most\n  economists *believe*. Claude: weak. Jev: strong. Withholding it loses a correct reading and writes\n  nothing wrong.\n\n**One link shows the two meanings of relevance coming apart.** *Serious loss of human control over\nAI systems is a coherent possibility*, filed for *there are multiple coherent causal routes from\nloss of control to catastrophe*. Both judges call the gain weak. The conditionals say 0.78 if the\nfirst is true and 0.02 if it is false, a difference of 0.75. The first claim is a precondition:\nestablishing it adds little, and losing it removes everything. Under the ruling in force, where\nrelevance is the gain (ledger Q34), that link is weak. Under starting values (Q35) it would be\nstrong. That is the case for storing both conditionals (6a.3, ledger W3) in one example, and it\nholds whichever model supplies the numbers.\n\n**What it does not show.** Twelve links, no human reading, and three of the twelve come from the\nsynthetic stress suite. Agreement with Claude is agreement between two machines. The second step of\n6d, the founder's sample of about forty links, is still the only ground truth any of this can be\nfitted against.\n\n**The same twelve links on four more models, the same evening**\n([whose-tokens-pay-for-mapping.md](whose-tokens-pay-for-mapping.md) § 3e). The reasoned judge's own\nprompt, unchanged, was asked of Gemini 3 Flash, Gemini 3.1 Flash Lite and GPT-6 Luna through a\nrouter key, and of Claude Sonnet 5 again as a baseline. Two things bear on this section. Claude\nagrees with its own earlier step on 10 of 12 and the three others on 4 to 6, and Gemini 3 Flash\nanswered *strong* to all twelve, so the five-step scale is read differently by each model. And the\none attack link, which the decision model read as raising at 0.79, was read as raising by all three\nother models and as lowering, as filed, by Claude both times. So four readers of two kinds agree\nagainst the judge that wrote the stored answer, on the one link where the filing is in doubt. That\nis the split the proposed arrangement of 6c is meant to catch, and it argues for a judge from a\nsecond family more than for any one model.\n\n---\n\n## Sources\n\n[Jev (AI model) — Wikipedia](https://en.wikipedia.org/wiki/Jev_(AI_model)) ·\n[TypeSafe's Jev scores 62.6% asked once and 95% split five ways — The Daily Brief](https://www.beri.net/article/typesafe-jev-typed-decision-model-calibration-decomposition-shadow-eval) ·\n[AINews: Jev, a \"System One Model\" — Latent Space](https://www.latent.space/p/ainews-jev-a-system-one-model-that) ·\n[What Is Jev AI? Inside TypeSafe's RLCD Model — Turing Post](https://www.turingpost.com/p/what-is-jev-rlcd) ·\n[Jev: TypeSafe's System One Model That Never Hallucinates — DataCamp](https://www.datacamp.com/blog/system-one-models-jev) ·\n[RLCD vs RLHF: what is TypeSafe's Jev actually claiming? — MindStudio](https://www.mindstudio.ai/blog/typesafe-jev-rlcd-vs-rlhf) ·\n[TypeSafe Jev: the first decision-only model class, benchmarked and priced — Developers Digest](https://www.developersdigest.tech/blog/typesafe-jev-system-one-models-release-guide-2026) ·\n[TypeSafe — Models (training, retention, language, limits)](https://docs.typesafe.ai/models) ·\n[TypeSafe/Jev Won't Train on Your Data. It Can Still Learn From It — Eduard Ruzga](https://wonderwhy-er.medium.com/typesafe-jev-wont-train-on-your-data-it-can-still-learn-from-it-563f0ad591b6)\n\n## Cross-references\n\nCode and data: `deliberus/decision_api.py` (the client; nothing in the pipeline imports it) ·\n`scripts/decision_model_bench.py` · `scripts/decision_model_shadow_relevance.py` ·\n[`jev-and-the-judgement-layer/`](jev-and-the-judgement-layer/) (the fixture and every raw answer of\n2026-10-04).\n\n[pass-quality-and-the-choice-of-model.md](pass-quality-and-the-choice-of-model.md) (2026-10-05: which\nkind of question each kind of model is enough for; closed questions travel across the cheap models,\nwhich is where a decision model sits too) ·\n`CLAUDE.md` § Principles (*a machine's no-answer belongs in a gate*, which creates the yes/no shape) ·\n[the-residual-error-taxonomy.md](the-residual-error-taxonomy.md) (same-hand bias, and the two-family\njury) · [fragile-checkers-and-the-verification-bottleneck.md](fragile-checkers-and-the-verification-bottleneck.md)\n(jury disagreement across families as the fragility predictor) ·\n[soft-canonical-clustering-and-reversible-merge-semantics.md](soft-canonical-clustering-and-reversible-merge-semantics.md)\n(the sameness sitting, which is the labelled data this needs) ·\n[relevance-and-the-two-factors.md](relevance-and-the-two-factors.md) and\n[strength-calculus-literature/bayesian-and-informal-logic.md](strength-calculus-literature/bayesian-and-informal-logic.md)\n(the judge § 6 would sit beside, and the Δ derivation it rests on)\n"}