Skip to content
voxlang.orgv0.6.0SSOT
Internals — research note: investigation or findings, not a description of shipped behaviour.

Jev / System One for Vox harness & orchestrator — research 2026-10-28

Integrate System One / RLCD ideas by feeding calibrated answers into build_selection_request and existing repair/stop hooks — not by invoking PreferFree or rewriting FREE_FLOOR. allow_free_in_performance_mode only lets :free models compete in the scorer. ChatARC composes skill, MCP, and command catalogs. Do not override clutch, emit model ids, or distill into MENS chat. See §17.

Revision (2026-10-28, pass 4). Pass 3 = redundancy + alternative performance (§11–§15). Pass 4 drops facade framing: map RLCD / System One research onto a fully plumbed harness turn (when/how select runs, PreferFree via intent fields, mid-loop escalate/stop)—worked example in §16. Do not treat hypothesis tables as a committed roadmap. Companions: autonomous-orchestration-policy-research-2026.md, multi-provider-local-cloud-routing-research-2026-07-30.md, model-catalog-ssot-2026.md, agentos-ssot-2026.md, search-retrieval-ssot-2026.md, voxmens-hub-and-spoke-ssot-research-2026-06-18.md, voxmens-finetuning-boundaries-research-2026-06-21.md, vox-harness-graph-audit-2026-07-30.md. Normative generative boundary: AGENTS.md §Model-Agnostic LLM Boundary. … [lean-ctx: omitted 1 lines]

0. Direct answers (operator FAQ)

Section titled “0. Direct answers (operator FAQ)”

| Question | Answer | … [lean-ctx: omitted 2 lines] | Can we download Jev weights / self-host? | No public weights or self-host binary. Hosted API only (POST /v1/systemone). Enterprise page lists VPC/on-prem as a sales question, not a documented product. Community “Jev-like” weights are not Jev. | | May we distill / imitate Jev into MENS? | Likely prohibited. MCA language reported by secondary digests bans distillation / imitating Service outputs / competing products. AutoResearch-style use of Jev probs as features for your own labels is what TypeSafe’s cookbook shows — still get legal review before any training on API outputs. | | Should we put Jev on the hot path for faster responses? | Only when cloud egress is already allowed and the judgment is closed-set. Vendor E2E is ~70–500 ms plus RTT/gateway. For local-only / privacy-strict / sub-50 ms floors, use rules → local classifier / MENS constrained pick → optional cloud Jev → frontier LLM. Jev is often cheaper than a frontier call, not always faster than a local SLM. | … [lean-ctx: omitted 2 lines]

1. Critique of the pass-1 draft (what was wrong or weak)

Section titled “1. Critique of the pass-1 draft (what was wrong or weak)”

| Issue | Why it was bad | Correction | … [lean-ctx: omitted 1 lines] | Implied “new crate vox-system-one” | Crate-edges / layer policy; premature | Prefer a thin client in existing egress (vox-http-client / orchestrator module) if adopted; no new workspace crate without USER-AUTHORIZED edges | … [lean-ctx: omitted 1 lines] | “Replace cross-encoder with Jev” | vox-search already has term-density / hybrid RRF; CLERC cookbook ≠ SOTA claim | Optional stage after shortlist, A/B against current reranker; never rip out lexical/hybrid first | … [lean-ctx: omitted 1 lines] | Banking “wire transfer” examples | Not Vox’s MCP surface; confuses readers | Use real irreversible classes: destructive VCS, secrets write, external effect, policy grant (oversight/floor.rs) | … [lean-ctx: omitted 3 lines] | Conflated registry “confidence” with Jev confidence | autonomic::ModelConfidence ≠ Choice confidence | Keep namespaces separate in telemetry | What pass 1 got right: non-generative contract; parallel questions; Kahneman System-1/2 split; do not stuff through llm_chat; escalate on low confidence; jaggedness (math/dates/counts in code). … [lean-ctx: omitted 1 lines]

2. What Jev is (primary)

Section titled “2. What Jev is (primary)”

| Axis | System One / Jev | Frontier / MENS chat | … [lean-ctx: omitted 4 lines] | Hosting | TypeSafe API and OpenRouter (typesafe/jev-*, Decisions API / System One base URL) | Multi-provider | … [lean-ctx: omitted 3 lines]

  • Aliases move (jev-latest → currently jev-1.14.0); pin versioned IDs when thresholds are tuned. … [lean-ctx: omitted 1 lines]
  • Coding agents: Jev is not a drop-in for Cursor/Claude Code models (coding-agents). … [lean-ctx: omitted 1 lines]
  • Evaluating and Benchmarking the System One Model Jev (jev-1.13.0, 37 datasets, ~346k requests): strong on several classification/NLI sets vs Qwen3.8-27B / Gemma-4-E4B on identical option-prob protocols; choice probs well calibrated for selective prediction; binary Nouls need threshold tuning (UNFAIR-ToS micro-F1 0.50→0.75); option rotation does not change accuracy.
  • Jev-ToM: typed ToM selection; Jev competitive; confidence calibration task-dependent; candidate design matters. Treat vendor 40–400× speed/cost claims as workflow-self-eval (blog discloses bias), not universal. … [lean-ctx: omitted 1 lines]

3. Fine-tune, download, and “should we build our own?”

Section titled “3. Fine-tune, download, and “should we build our own?””

3.1 Fine-tune Jev — no

Section titled “3.1 Fine-tune Jev — no”

Official: not fine-tuned / LoRA’d per customer; same weights; customize via request + code (Models — Customizing Jev).

3.2 Download / self-host Jev — no (public)

Section titled “3.2 Download / self-host Jev — no (public)”

No HF weight drop, no self-host binary in docs. OpenRouter and Vercel AI Gateway are hosted proxies to TypeSafe, not local runtimes (OpenRouter Jev).

3.3 Train something like System One for Vox — maybe, carefully

Section titled “3.3 Train something like System One for Vox — maybe, carefully”

| Path | Fit for Vox | Risk | … [lean-ctx: omitted 3 lines] | C. Jev probs as features → classical head (AutoResearch pattern) | Good when cloud Jev allowed offline for dataset build | MCA: features-for-your-labels OK in cookbook spirit; imitation/distillation banned — legal review | | D. Community Jev-like (Jevlike, decider/Qwen3.5, JevForge, ModernBERT “open jev”, etc.) | Possible air-gap System-1 API shape | Quality/calibration unproven vs Jev; supply-chain; not TypeSafe; do not claim parity | … [lean-ctx: omitted 1 lines] | F. Llama-Guard / ShieldGemma / NeMo-style rails | Local guardrails | Different task than general Choice routing | Recommendation: Do not fine-tune Jev (impossible). Do not distill Jev into MENS. For faster local responses: invest in B/E (labeled routers + constrained gen) and keep hosted Jev (via OpenRouter) as an optional cloud tier when D8 allows. Community Jev-likes are a spike candidate for air-gap only after an eval harness, not a default dependency.

3.4 Latency reality

Section titled “3.4 Latency reality”
Local rule / embedding kNN     ~0.1–5 ms
Local SLM / MENS constrained   ~10–200+ ms (device-bound)
Hosted Jev (vendor claim)      ~70–500 ms compute + WAN/gateway RTT
Frontier LLM                   seconds (often)

… [lean-ctx: omitted 2 lines]

4. Fit / anti-fit (revised)

Section titled “4. Fit / anti-fit (revised)”

4.1 Strong fits (independent or hybrid)

Section titled “4.1 Strong fits (independent or hybrid)”
  • Closed-set intent / stakes / complexity when cloud OK → feed SelectionIntent / skip Elite.
  • I/O guardrail Noul battery (jailbreak, exfil, self-harm, …) with code thresholds (guardrails cookbook). … [lean-ctx: omitted 1 lines]
  • Skill shortlist re-check after lexical vox skill search (skill suggestion cookbook). … [lean-ctx: omitted 1 lines]

4.2 Anti-fits

Section titled “4.2 Anti-fits”
  • Open-ended generation / “discover new themes.”
  • Physical / hard-realtime control. … [lean-ctx: omitted 2 lines]
  • Loops that need a textual rationale from the discriminator for the next LLM step — pass structured fields, not fake prose. … [lean-ctx: omitted 2 lines]

5. Integration boundaries (revised)

Section titled “5. Integration boundaries (revised)”

| Concern | Home | … [lean-ctx: omitted 2 lines] | System One judgments | Separate HTTP client → TypeSafe or OpenRouter System One / Decisions API (reuse OpenRouter secret if already resolved) | | Secrets | vox_secrets::resolve_secret — TypeSafe key or existing OpenRouter key; never raw env::var in consumers | … [lean-ctx: omitted 1 lines] | Detector | llm_provider_call must not treat System One as a banned chat SDK — document egress allowlist in the adopting PR | … [lean-ctx: omitted 2 lines] Fail policy: guardrail path → fail closed on System One outage (or fall back to local Guard-class model). Cascade prefilter outage → fail open to existing select() (do not block the harness on Jev). … [lean-ctx: omitted 1 lines]

6. Hypothesis backlog (not a roadmap)

Section titled “6. Hypothesis backlog (not a roadmap)”

| ID | Experiment | Exit | … [lean-ctx: omitted 1 lines] | H0 | OpenRouter System One spike + golden fixtures | Measured p50/p95 from CI region; cost/call | | H1 | Optional guardrail battery behind flag | FP/FN budget on labeled prompts | | H2 | Optional intent prefilter when cloud eligible | Token $ drop without quality regress on chat golden set | … [lean-ctx: omitted 3 lines]

7. Claim ledger

Section titled “7. Claim ledger”

| Claim | Status | … [lean-ctx: omitted 3 lines] | No public self-host weights | Primary docs silence + secondary; treat as hosted-only until TypeSafe publishes otherwise | | OpenRouter typesafe/jev-* | Primary OpenRouter docs | | MCA bans distillation / imitation | Secondary digests quoting MCA — verify PDF before training | … [lean-ctx: omitted 1 lines] | “1k entries <6s / <$0.10” | Unverified tertiary | … [lean-ctx: omitted 1 lines] | Vendor 193×/444× workflow speedups | Vendor self-eval | … [lean-ctx: omitted 1 lines]

8. References

Section titled “8. References”

Primary

Section titled “Primary”

Independent / secondary

Section titled “Independent / secondary”

9. Before / after — vox research enrichment (inline)

Section titled “9. Before / after — vox research enrichment (inline)”

9.1 What we ran

Section titled “9.1 What we ran”
vox research up   # SearXNG was offline; Tavily plan was over quota (432)


# Pass A — broad deep research (hypothesis generation)
vox research run --json --scope web --lane deep --waves 2 --max-sources 12 \
  "TypeSafe Jev System One: can weights be downloaded or self-hosted; is fine-tuning
   allowed; OpenRouter hosting; alternatives for local calibrated decision models;
   latency and privacy tradeoffs vs cloud API for agent harness routing and guardrails"


# Pass B — focused + claim verification
VOX_DATA_DIR=/tmp/vox-research-scratch \
  vox research run --json --scope web --lane fast --waves 1 --max-sources 8 \
  --verify-claims \
  "TypeSafe official docs: is Jev fine-tunable and are weights downloadable"

… [lean-ctx: omitted 13 lines]

9.2 Pass-1 doc gaps (before)

Section titled “9.2 Pass-1 doc gaps (before)”

Snapshot: /tmp/jev-research-BEFORE.md (237 lines, sha256 2da17353…). Missing or weak relative to pass 2: … [lean-ctx: omitted 3 lines]

  • Explicit “do not put Jev on every select()” … [lean-ctx: omitted 2 lines]
  • Honest critique of proposing a new crate / replacing rerankers

9.3 What deep research added (after) — distilled

Section titled “9.3 What deep research added (after) — distilled”

| Lead | Source pass | Disposition | … [lean-ctx: omitted 1 lines] | OpenRouter lists typesafe/jev-1.13 / 1.14 at $0.042/M in | A | Confirmed via OpenRouter docs | … [lean-ctx: omitted 1 lines] | Fine-tuning not offered / same weights all accounts | B (improved) | Confirmed against TypeSafe Models; Pass A wrongly said “not addressed” | | Local alternatives: cascades, NeMo/Llama-Guard, community System One-compatible servers | A | Adopted as path D/F with “not Jev” warning | … [lean-ctx: omitted 1 lines]

9.4 Excerpts — research executive summaries (raw)

Section titled “9.4 Excerpts — research executive summaries (raw)”

Pass A (deep, unverified citations):

Weights and self-hosting: evidence says Jev’s weights are not available… Fine-tuning: the evidence does not address whether fine-tuning is allowed… OpenRouter: Jev is listed… Local alternatives: community local server… NeMo Guardrails or Llama Guard-style classifiers… Privacy: hosted-only means prompts go to a third party, and you need a fallback if it is unavailable. … [lean-ctx: omitted 1 lines] Fine-tuning: The evidence says no. Jev is not offered as a customer-fine-tunable model. TypeSafe is quoted as saying Jev is “not fine-tuned or LoRA-adapted with customer data”… Weights: The evidence says no. TypeSafe has not published Jev’s weights… Caveat: TypeSafe-owned pages in the retrieval set were snippet-thin; claims still need primary-doc confirmation. … [lean-ctx: omitted 1 lines]

9.5 Net effect on this document

Section titled “9.5 Net effect on this document”

Pass 2 replaced soft roadmap language, answered the fine-tune/download/speed questions, wired OpenRouter as preferred cloud egress when D8 allows, added MCA distillation caution, added arXiv evidence, fixed research tooling blockers (SearXNG, VOX_DATA_DIR, dead free_floor slug), and showed that --verify-claims is required for non-zero citation precision.

10. Decision summary

Section titled “10. Decision summary”

| Question | Answer | … [lean-ctx: omitted 5 lines] [… truncated at ~4055 of 7464 tokens — use ctx_read with lines= parameter to see specific sections]