---
title: "JEV → Tally + Tapestry Looms"
description: "Date: 2026-09-23 · Status: Tapestry LIVE · Tally committed, deploy pending (Douglas)"
created: 2026-09-24
updated: 2026-09-24
authors: ThinkingCap R&D
topics: [Tally]
status: published
canonical: https://console.thinkingcap.com/rd/Tally/2026-09-23-jev-tally-tapestry
summary: "JEV is a typed-judgment engine that answers questions over text states, deployed in Tapestry as a live quality gate and committed in Tally's Max analyzer. The system runs JEV answers in shadow mode against Anthropic's analyzer to measure agreement before fully switching over, with deployment pending once task-type match reaches 85% accuracy. JEV's cost comes from a pre-paid token pool, eliminating per-token charges for faster, cheaper routing decisions."
audio: https://thinkingcap.blob.core.windows.net/rd-home/summaries/0155df3febd8b66e34c402978786e068ed0b2f604c0263820bafdb04df9872d0.mp3
audio_full: https://thinkingcap.blob.core.windows.net/rd-home/summaries/full-d9d40481bb32d83d6e499fe9a286d25c83a635e7bebd5e0dc327a69d92591f32.mp3
---

# JEV → Tally + Tapestry Looms

**Date:** 2026-09-23 · **Status:** Tapestry LIVE · Tally committed, deploy pending (Douglas)

---

## What JEV is

JEV AI ([jev-ai.pro](https://jev-ai.pro/jev-api)) is a **typed-judgment engine — not a chat model**. It cannot hold a conversation turn, so it never enters Courier's model menu. What it does: answer typed questions over a state text, fast and cheap.

```
POST https://jev-ai.pro/api/v1/systemone
Authorization: Bearer sk_jev-ai_...

{
  "model": "jev-latest",
  "state": "<the text being judged>",
  "questions": {
    "urgent":  { "type": "noul",   "instructions": "Does this need immediate attention?" },
    "route":   { "type": "choice", "instructions": "Which team handles this?",
                 "criteria": { "support": "...", "engineering": "...", "billing": "..." } },
    "severity":{ "type": "score",  "instructions": "Rate the severity",
                 "criteria": ["trivial", "minor", "major", "critical"] }
  }
}
```

- **noul** → yes/no probability (`{ "noul": 0.87 }`)
- **choice** → pick + per-option probabilities + confidence
- **score** → fractional level index + probabilities + confidence

**Limits:** 64k model context; state + longest question ≤ 32k tokens; ≤64 questions/call; 256 KB body.
**Billing:** input tokens draw from a paid pool, **output tokens are free**; 1 credit/call when the pool is empty. No $/Mtok schedule.
**Models:** `jev-latest`, `jev-preview`, `jev-1.13.0` (pinned), `laya-english` / `laya-multilingual` (small strict-limit classifiers).
**Account state (2026-09-23):** 75M paid input tokens, 5 credits. Key verified live.
**Helper endpoints:** `GET /api/v1/models`, `GET /api/v1/credits` (free — used for key health checks).
**Docs' own positioning:** LLM Router, LLM-as-judge, RAG eval, classification.

The API key lives in Tapestry's `system_api_keys` as `tapestry_jev_1` (with base_url + models in metadata). A capcom creds copy is Douglas's to add.

---

## What's live in Tapestry — `seamstresses @ e2955ff` (deployed 2026-09-23)

| Piece | File | What it does |
|---|---|---|
| JEV client | `src/services/jev.ts` | Singleton `jevSystemOne(state, questions)`. Key from `tapestry_jev_%` row (`JEV_API_KEY` env override). **Fail-open everywhere** — every caller keeps a non-JEV path. |
| Startup | `src/server.ts` | `loadJevConfigFromDb()` beside the Tally config load. |
| **Quality gate pre-pass** | `src/services/stream.ts` | The per-turn gate verdict is a yes/no question — JEV's native shape. Confident PASS (**p ≥ 0.8**) skips the haiku call entirely (input-token-only cost, ~0.4s vs ~2s). Contested drafts escalate to haiku because the retry loop needs the prose REASON/PATTERN. Watch logs for `[gate] jev fast-pass p=…`. |
| Key health | `src/routes/adminExtended.ts` | `tapestry_jev_%` pool in the SDK health check via free `GET /credits`. `jev-sdk` service card (totalServices 13). |
| Services page | `frontend/.../ServicesPage.tsx` | `jev-sdk` entry in the provider key list. |

**Validation:** live `systemone` smoke test passed; `tsc` clean; jest 575/577 (2 failures pre-existing on the clean tree — directLoom repo-shorthand, unrelated).

---

## What's committed in Tally — `main @ 4651b75` + `7a6d050` (deploy = Douglas)

**`4651b75` — JEV provider + Tally Max analyzer path**
- `model-registry.ts`: all 5 JEV models registered (`supportsToolCalls: false`, cost 0-as-unknown since billing is account-token based). Registry documentation + analyzer cost accounting — JEV is never seeded into chat routing.
- `max/analyzer.ts`: `callAnalyzerJev` behind `TALLY_MAX_ANALYZER_PROVIDER=jev` (default `anthropic`). The analyzer's refined shape is a 1:1 map onto JEV question types: task_type→choice(8), structure_type→choice(4), complexity→score(5 levels, normalized 0..1), determinism→noul, capability tags→one noul each. JEV failure falls back to the Anthropic analyzer.
- **Documented tradeoff:** typed answers can't carry `replacement_text`, so the JEV path returns `rawOps: []` — shape refinement only, lower `est_savings`.

**`7a6d050` — analyzer shadow (agreement measurement)**
- Analyze-and-discard retention makes post-hoc comparison impossible, so after each uncached `/route/max` verdict, the *other* brain runs fire-and-forget into `max_analyzer_shadow` (migration 016). Never delays or fails the response.
- Gates: `TALLY_MAX_JEV_SHADOW=1` (JEV shadows Anthropic — pool tokens only) · `TALLY_MAX_ANTHROPIC_SHADOW=1` (reverse — real $, own gate).
- `tools/jev-agreement.sql`: head-to-head match rates per field + per-brain cost/latency + a posture section that works with no shadow data.

---

## The decision already made (2026-09-23)

**Should JEV run the Tally Max analyzer? Yes — shadow first, flip on evidence.**

Key fact that settled it: **Tapestry never reads `context_ops` or `est_savings`** from Max — only `recommended_model`, confidence, exploration flag, `why`. The one thing JEV can't produce is the one thing we don't use.

Also in JEV's favor: it eliminates the Anthropic path's malformed-JSON retry fragility (typed answers by construction), and its cost is pool tokens already paid for (~3k–15k analyses from 75M tokens depending on state size, fewer with the 24h verdict cache).

**Deploy sequence:**
```bash
# 1. Deploy tally main (applies migration 016), set:
TALLY_MAX_JEV_SHADOW=1
JEV_API_KEY=sk_jev-ai_...   # in system_api_keys / creds

# 2. After a few hundred analyses:
psql "$DATABASE_URL" -f tools/jev-agreement.sql

# 3. Flip when satisfied:
TALLY_MAX_ANALYZER_PROVIDER=jev
```

**Flip rule of thumb:** task_type match ≥ 85% with complexity mean-abs-delta ≤ 0.15 → JEV is interchangeable on routing.

---

## Where things live

| Thing | Location |
|---|---|
| JEV client + docs | `tapestry-node/src/services/jev.ts` (header comment) |
| JEV API key | tapestry DB `system_api_keys.tapestry_jev_1` |
| Agreement query | `tally/tools/jev-agreement.sql` |
| Shadow migration | `tally/apps/api/src/db/migrations/016_max_analyzer_shadow.sql` |
| Memory | `jev-tally-tapestry.md` in Claude memory |
| Session log | `~/.claude/sessions/2026-09-23-jev-tally-tapestry.txt` |