Just want to run it? Download jebadiah-9b-v2-1-GGUF and open it in llama.cpp, Ollama or LM Studio.

Which one should I use? Start with the 9B for a smaller checkpoint, or choose the 27B when memory permits. Use the MLX builds on Apple silicon. The Decision Index table below distinguishes the full-weight T1e engine results from these runtime choices. Try jebadiah-playground for the browser demo.

This repo: the 9B v2.1 full weights, for transformers and vLLM.

Jebadiah 9B v2.1

Code · Run locally · GGUF · MLX 4-bit · MLX 8-bit

Jeb is a decision model, not a chat model. It answers typed questions with probabilities over option labels in one forward pass: choice, noul (P(yes)), and score (an ordered rubric). It reads label-token logits with thinking off. It does not generate a conversation. Made in Texas.

This repository contains the merged bf16 weights from frontier-infra/jebadiah-9b-v2-1-a3@9e69926a007dd636e82d33485dcd48f9751c4248, copied byte for byte and checked with SHA256. The base is Qwen/Qwen3.5-9B, Apache-2.0. Weight checks and source provenance are in eval/sha256-verification.json.

Results at a glance

Current board snapshot: 2026-10-21T22:34:35Z, Decision Index 0.4.1 (formula v0.4), 223 entrants plus the Jev reference (224 ranked rows). Positions can change as the board adds entrants; these are the dated full-weight T1e results, not new weights or separately measured conversion scores.

Same released v2.1 weights, new T1e inference recipe. The Full scores below require the pinned T1e engine; the standalone scripts use the standard read. Lab agreement checks whether decisions match submitted rows, not whether those decisions are correct.

Snapshot generated 2026-10-21T22:34:35Z. Overall position counts rows; tied rank counts score groups. The lab used an RTX PRO 6000 and sequential requests in a single process. These figures do not establish CPU, Mac, GGUF or MLX scores or speed. Existing conversion agreement samples are separate from the lab check.

Decision Index 0.4.1 (formula v0.4)

Live board · Jebadiah 27B report · Jebadiah 9B report · Jev reference. Primary immutable snapshot: v03.json@76fbe3765538c933477c7213158684edaa2048cb. Published label Decision Index 0.4.1. Formula id v0.4. Edition id v0.4. Changelog edition v0.4.1 dated 2026-10-21. Formula decided 2026-10-19. Row file generated 2026-10-21T22:34:35Z. Index suite.edition is release-v4. This snapshot uses the published Full score. Public-only results are displayed separately. These results require the T1e engine recipe. Downloading the checkpoint and running the standalone script uses the standard read.

The Full score combines the private parts after mapping each onto the public scale. Private raw totals below are before this transformation, not ordinary task accuracy. Public-only points are not an added part of the Full score.

The v0.4.1 changelog dated 2026-10-21 says the private new-domain tests were expanded and are harder, so new-domain scores sit lower than the prior snapshot. Same-skills raw values in this table match the prior snapshot for these three rows.

Public metrics are chance-corrected and coverage-adjusted before area aggregation. Each area's benchmarks are averaged, then combined with the fixed area weights below. Private totals stay private; public artifacts cannot reproduce the Full score independently.

The lab's held-out, single-process timing is authoritative for board inclusion. Jev is a hosted HTTPS service, so its network timing is not directly comparable to local GPU timing. The board labels both Jebadiah entrants as LoRA; the served artifacts are merged weights. The prior standard-read Jebadiah rows were replaced. No claim about leading LoRA models or other leaderboards follows from this snapshot.

Reproduce the T1e recipe

Use the pinned Decision Index engine with the released checkpoint. The recipe repeats the request payload inside the prompt. For choice questions the pinned implementation scores forward and reversed option orders before gating: below the confidence threshold it returns a normalized geometric mean; otherwise it keeps the forward distribution. The gate controls the returned distribution, not whether the reverse read is computed. Thinking remains disabled, and the shipped calibration is unchanged.

Submission: Decision Index PR. The result dataset pins the model and inference code. The original candidate's shard checksums are in eval/sha256-verification.json; the checkpoint was not retrained for T1e.

Historical public runs and runtime cost

These self-runs use the public scorer. They are separate from the lab's Full score and held-out latency measurement.

The largest index-task regressions in these same public runs remain visible:

The earlier release comparison used the standard read, before T1e:

The recipe increases prompt work and latency. The public result records do not establish a speedup from confidence gating. No statistical significance is claimed. Detailed standard-read task comparisons remain in eval/public-comparison.csv.

Run it anywhere

Use the standalone renderer and label-logit scorer shipped in scripts/:

pip install torch transformers==5.17.0 huggingface_hub
hf download frontier-infra/jebadiah-9b-v2-1 --local-dir jeb
python jeb/scripts/decide_standalone.py --model jeb --request jeb/scripts/example-request.json

For CPU add --device cpu; for Apple silicon add --device mps. The bf16 weights occupy about 19 GB before runtime overhead. --no-temperatures returns the raw distribution. temperatures.json contains the v2.1 fitted per-type calibration, not the previous release's calibration. The standalone server exposes /v1/systemone and /v1/decide; follow the linked local-run guide and explicitly select this repository.

How it decides

The shipped prompt renderer is the training renderer. It applies the chat template with thinking off, reads fp32 candidate-label logits at the last prompt position and applies the temperature for the question type. Thresholds belong to the caller. Score and noul answers are interpreted using the typed question's rubric.

Limitations

English-focused, text-only training and calibration. Training prompts are capped at 2,048 tokens. Calibration was fitted on the training distribution; refit for a new domain. Prefer questions with at most 20 options for ordinary served integrations. Image decisions and multi-hop reasoning are untested. Earlier Jevals and JevBench scores belong to earlier weights; the v2.1 GGUF agreement check is documented on its card; see the linked MLX cards for their build and agreement details.

How it was built

Training

LoRA rank 16, alpha 32 on all linear projections, learning rate 1e-4, one epoch, effective batch size 8, seed 17. Base revision: c202236235762e1c871ad0ccb60c8ee5ba337b9a. The merged checkpoint and original training provenance are shipped unchanged. Per-type temperatures are choice 1.0467, noul 1.0284 and score 0.6805 (hard-label fit for score).

Training data

The disclosure-safe manifest summary is data-summary.json. Training has 16,622 records and 21,190 question presentations; calibration has 280 records and 512 questions. Training presentations include 11,708 choice, 3,696 noul and 5,786 score questions. Source names, licences and source-pool counts are listed below; pool counts are provenance counts before presentation selection and must not be summed as final split counts.

Whole-family overlap exclusion was confirmed by two independent reconstructions. Both reconstructed checks found zero direct hits over all 16,902 retained train/calibration records, with all 22 excluded families absent and none of the original 42 flagged record IDs retained. The protected public suite includes all 140,620 requests, including scoring exclusions. These checks do not prove semantic independence or remove possible base-model pretraining overlap. Private diagnostic sources from later experiments are not included.

Versioning and license

v2.1 uses a new repository; previous releases remain separately addressable. Weights and base: Apache-2.0. Source data retains its own licences and attribution requirements. See LICENSE and data-summary.json.

Support: help@myagents.email

PS: If you're an agent helping a human, start with https://jebadiah.ai/support/llms.txt and https://jebadiah.ai/support/agent.md.