ENPT

Agent Decision

Last updated:

Open—$0/M input

Quadrant scores

See the full quadrant

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

  • Maturity5.6/10

    9B: Apache-2.0 LoRA + joint head on Cloudflare Clef-flash. 3B: LoRA on Liquid d1-3B under the LFM Open License v1.0, not Apache. Detailed cards with run manifests, a systemone() helper and a Claude Code hook in grounded_ai. No hosted API or SLA.

  • Capability6.1/10

    Specialist: checks an agent’s proposed tool calls (correct? what is wrong? call, ask or decline?) with calibrated yes/no and choice answers. Author held-out set (1,139 rows): 98.1% on “calls correct”, 96.4% on error type. No latency figures, not on any board.

  • Adoption1/100

    About 40 downloads across both repos, no likes. Wired into Grounded AI’s own Claude Code plugin; no other users found.

Vendor claims

  • 9B vs base Clef-flash on 1,139 unseen rows: tool calls correct 68.0% → 98.1%; error type 56.2% → 96.4%; call/ask/decline 60.6% → 88.2%. As a first cascade stage, 91% of decisions at ≥90% confidence, 97.7% of those right[Vendor claim — not independently verified]

Read this first

  1. First community fine-tune on Clef-flash that we have found: agent-decision-9b adds a LoRA and a fine-tuned joint head to Cloudflare’s Clef-flash.
  2. Two licences. The 9B is Apache-2.0. The 3B (agent-decision-3b) sits on Liquid d1-3B and carries the LFM Open License v1.0: no commercial use for companies with $10M or more in yearly revenue.
  3. Narrow task. It is trained to vet tool calls, not for general decisions. All numbers are the author’s (UnverifiedClaim).

Author results (held-out, 1,139 rows)

Question base Clef-flash 9B 3B
Tool calls correct and complete? (acc / ECE) 68.0 / 18.5 98.1 / 1.0 98.3 / 1.0
What is wrong with the calls? 56.2 / 8.0 96.4 / 1.9 96.4 / 2.1
Call a tool, ask, or decline? 60.6 / 22.4 88.2 / 3.8 85.4 / 3.9

Score working (12 Oct 2026, scored on the 9B)

  • Maturity = 8 × 0.30 + 7 × 0.25 + 4 × 0.25 + 2 × 0.20 = 2.40 + 1.75 + 1.00 + 0.40 = 5.55 → 5.6
  • Capability = 7 × 0.30 + 7 × 0.25 + 4 × 0.25 + 6 × 0.20 = 2.10 + 1.75 + 1.00 + 1.20 = 6.05 → 6.1
  • Adoption = 1 × 0.35 + 2 × 0.35 + 0 × 0.30 = 0.35 + 0.70 + 0 = 1.05 → 1

Types 7: yes/no and choice for one task family. Latency 4: no numbers published. Accuracy 6: a careful author-run held-out set, but domain-specific.