ENPT

reflex

Last updated:

Open80–250ms$0/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

  • Maturity4.7/10

    Public LoRA, local serve, and a browser WebGPU demo; research packaging without a hosted SLA keeps maturity mid-band.

  • Capability6.9/10

    Official boards (22 Oct): Decision Index 0.4.1 Full 58.77 (#53) for the 27B pmscr build; JevBench Intelligence 30.9, calibration 87.0, p50 1.358 s.

  • Adoption15/100

    Playground plus HF LoRA visible in the replica wave; still early and builder-niche versus hosted Jev.

Vendor claims

publicado (OK do Fabricio 17/10 BRT). Board audit of 17 Oct 2026: see the “Board audit” section below; where it conflicts with older board statements on this page, it wins.

reflex is an open LoRA (+ calibration) reconstruction of the System One contract by kshetrajna12 — default backbone Qwen3.5-4B — not TypeSafe and not a frontier base model.

JevBench v1.4.2

Metric Score Rank
Overall 54.0 #5
Intelligence 46.8 -
Calibration 74.2 -
Speed 81.3 -
Cost 56.8 -

Top-performing LoRA adapter on JevBench — the calibration story (ECE ~0.039 after temperature fit) holds up under independent evaluation. Beats several full fine-tuned models.

Pass state (text/JSON; images supported in the engine) plus typed questions (choice, score, noul → catalog boolean) and get distributions without free-text decode. API shape mirrors TypeSafe POST /v1/systemone for local base_url swaps. LoRA: kshetrajna12/reflex-qwen3.5-4b-lora; repo: github.com/kshetrajna12/reflex; playground: kshetrajna12.github.io/reflex.

Limits (read before you ship)

  • Author latency (~100ms cached) is self-reported; ECE figures now validated by JevBench v1.4.2 (calibration score 74.2).
  • Browser demo uses a smaller 0.8B path; do not assume 4B calibration transfers.
  • Local serve / WebGPU — fine for demos, not a hosted waitlist SLA.
  • Do not ship people-affecting thresholds without your own labelled eval.

Why it is in the catalog

Public typed decision I/O, open LoRA weights, and a real browser demo. Open/local peer next to Kev (artifactKind: lora-adapter) and below hosted Jev on adoption.

Board audit (17 Oct 2026)

publicado (OK do Fabricio 17/10 BRT). Checked against the official boards on 17 Oct 2026: Decision Index 0.3.2 (data of 16 Oct, 19:00 UTC), JevBench and ImageJevBench live. The boards are separate sources and are never averaged. Older board statements and rankings further down this page are superseded where they conflict with this section.

Board Row Result
Decision Index 0.3.2 (16 Oct 19:00 UTC) reflex Qwen3.8-27B-FP8 (pmscr) Full 54.72 (#48, tie group 11); public 57.34 (#45); same-skills 57.39; new-domains 42.74
JevBench open weights (live 17 Oct) reflex 4B Capability 59.0; Intelligence 30.9; calibration 87.0; p50 1.358 s; $0.017/1k
ImageJevBench v0.3.0 Reflex 4B Capability 53.3; Intelligence 27.4; calibration 79.3; p50 0.48 s; $0.052/1k
ImageJevBench v0.1.5 archive (frozen) Reflex 4B Capability 73.5; Intelligence 63.3; calibration 83.7; p50 0.46 s; $0.049/1k

Board-status statements on this page to recheck against the table (may be superseded):

  • “Author latency (~100ms cached) is self-reported; ECE figures now validated by JevBench v1.4.2 (calibration score 74.2).”

Score update (audit, 17 Oct 2026). Capability 6.8 → 6.7, first full working (the earlier score had no stored breakdown) = types 10 × 0.30 + calibration 6 × 0.25 + latency 3 × 0.25 + accuracy 7 × 0.20 = 6.65 → 6.7. Types from decisionTypes; calibration from the JevBench axis 87.0 (≥90 → 7, 85–89.9 → 6, 75–84.9 → 5, 60–74.9 → 4, below 60 → 3; no author ECE counted); latency from the board’s self-hosted p50 1.358 s without p99 (≤0.25 s → 7, ≤0.5 s → 6, ≤1 s → 5, above → 3); accuracy Decision Index Full 54.72 ÷ 60.11 × 8 = 7.3, half-up. Maturity and adoption unchanged.

Hugging Face on 17 Oct (kshetrajna12/reflex-qwen3.5-4b-lora): 245 downloads, 1 likes, licence tag mit, last modified 2026-09-20.

Board audit (20 Oct 2026, DI 0.4)

publicado (OK do Fabricio 20/10 BRT).

Decision Index 0.4 (20 Oct; Full now from the private parts only) lists reflex Qwen3.8-27B-FP8 (pmscr) at Full 58.94 (#46); it was 54.72 on the archived 0.3.2 edition. Accuracy rule, reference Jev 1.13 on DI 0.4 Full (61.26): 8 × 58.94 ÷ 61.26 = 7.70 → 8 (was 7). Capability = 10 × 0.30 + 6 × 0.25 + 3 × 0.25 + 8 × 0.20 = 6.85 → 6.9 (was 6.7). Other sub-scores, maturity and adoption unchanged.

Board audit (22 Oct 2026, DI 0.4.1)

publicado (OK do Fabricio 22/10 BRT).

Decision Index 0.4.1 (22 Oct; reference Jev 1.13 Full 59.22, was 61.26 on 0.4) lists this page’s row at Full 58.77 (#53). Accuracy rule: 8 × 58.77 ÷ 59.22 = 7.94 → 8 (was 8). Capability unchanged. Maturity and adoption unchanged.