ENPT

Cam1

Last updated:

Open255–915ms$0/M input

Quadrant scores

See the full quadrant

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

  • Maturity3.7/10

    Apache-2.0 LoRA adapter plus a decision head, with REPRODUCE.md, hashes, a published 20,000-row training set and a runtime zip. It needs a custom FP32 runtime on a 96 GB GPU, serves no /v1/systemone, one author, zero downloads.

  • Capability4.9/10

    Choice and Noul are trained and tested; Score is trained but not established. DI 0.3 public 30.72 (the board lists Full 30.78, rank 121), roughly half of Jev 1.13's 60.11. The author's own run: median 255 ms, p95 915 ms on an RTX PRO 6000. Calibration under FP32 unproven.

  • Adoption1/100

    Zero downloads and zero likes on the card; one author post on X; training data and DI submission published. No integrations or users.

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.

publicado (OK do Fabricio 17/10 BRT).

Read this first. Cam1 is a learning project with honest numbers, not a Jev competitor. Its score of 30.72 on the Decision Index 0.3 public suite is about half of Jev 1.13’s. The card itself says calibration under its serving runtime is “unproven” and that it is “not a maintainer Full score or leaderboard admission claim”.

Cam1 is a decision model by an independent author (cbuk10001): the pinned Qwen3.5-4B backbone, a rank-8 LoRA adapter and a joint-schema decision head whose architecture derives from Cloudflare’s Apache-2.0 Clef implementation. It returns Choice probabilities and Noul (yes/no) decisions directly. It was trained locally on an RTX 4070 SUPER (12 GB) on 20,000 published records (SNLI, BANKING77, GoEmotions, LocalLLaMA typed decisions, and controlled synthetic decisions).

Specs

Attribute Value
Author cbuk10001 (independent)
Created 8 Oct 2026 (card updated 13 Oct)
Weights Apache-2.0: LoRA adapter, decision head and tokenizer; the Qwen3.5-4B backbone is downloaded separately
Runtime Custom inference runtime (runtime.zip, cam1_engine:Cam1Engine); “the adapter alone is not the complete model”
Validated serving Full FP32, reference kernels, GPU-only, 16,384-token limit, fresh process per schema policy; one RTX PRO 6000 (96 GB)
Question types Choice and Noul trained and evaluated; Score trained but “not established”
Training data Cam1-training-data with hashes and licence attributions
Endpoint No /v1/systemone server
Downloads / likes 0 / 0 (17 Oct)

Boards

Decision Index 0.3 (Hugging Face Space, edition 0.3.2, generated 16 Oct 19:00 UTC): public score 30.72, same-skills 32.28, new-domains 28.12, Full 30.78, rank 121 (the 16 Oct radar recorded #115 at the time). Jev 1.13.0 is 60.11 on the same board. The card describes its 30.72 as the author’s own run of the official public suite; the board lists the entry. Not on JevBench or Artificial Analysis.

The author’s latency (own run, successful requests): mean 409 ms, median 255 ms, p80 399 ms, p95 915 ms, in FP32 on the RTX PRO 6000. The card notes these do not certify the maintainers’ held-out latency test.

Fit / anti-fit

Fit when you need: a fully documented, reproducible small decision-model recipe (data, hashes, runtime) to study or extend.

Anti-fit when you need: accuracy anywhere near Jev’s, a served API, Score questions, or modest hardware (the validated setup is FP32 on a 96 GB GPU).

Limits

  • Calibration was fitted under BF16 and its validity under the FP32 runtime is stated as unproven.
  • 2,000 PhishNChips inputs were rejected by the schema validator; the author’s follow-up compatibility engine fixes this but has no new full score.
  • English only.

Why it is in the catalog

Open Apache-2.0 adapter, published data, a documented decision-head design, and a board entry. Scored low on accuracy and integrations; kept for completeness of the open adapter line (compare Kev).

Score working (17 Oct 2026)

  • Maturity = availability 5 × 0.30 + docs 7 × 0.25 + integrations 1 × 0.25 + support 1 × 0.20 = 1.50 + 1.75 + 0.25 + 0.20 = 3.70 → 3.7
  • Capability = types 6 × 0.30 + calibration 4 × 0.25 + latency 6 × 0.25 + accuracy 3 × 0.20 = 1.80 + 1.00 + 1.50 + 0.60 = 4.90 → 4.9
  • Adoption = engagement 1 × 0.35 + ecosystem 2 × 0.35 + production 0 × 0.30 = 0.35 + 0.70 + 0.00 = 1.05 → 1

Availability 5: weights are public but need a custom runtime archive and an FP32 96 GB setup. Docs 7: REPRODUCE.md, manifests, hashes, data licences. Integrations 1: none. Support 1: one author.

Types 6: choice and noul only. Calibration 4: a fitted temperature with a method, but explicitly unproven in the served runtime. Latency 6: median 255 ms but p95 915 ms (author’s run). Accuracy 3: DI 0.3 public 30.72, about half of Jev’s.

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) Cam1 Full 30.78 (#121, tie group 31); public 30.72 (#121); same-skills 32.28; new-domains 28.12

Version / family note. Score kept at 4.9 (accuracy 3, as published on 17 Oct). Applying the audit rule to Decision Index Full 30.78 ÷ 60.11 × 8 = 4.1 would give accuracy 4 and capability 5.1; kept as a documented exception approved by Fabricio (17 Oct 2026).

Hugging Face on 17 Oct (cbuk10001/Cam1): 0 downloads, 0 likes, licence tag apache-2.0, last modified 2026-10-13.