aminry/decisio-decision-index

decisio on the Decision Index 0.2.1, self-run

Per-request results of three runs of the Decision Index kit's whole 0.2.1 suite against decisio, an open-source decision server: typed questions about a piece of text go in, and a probability for every option comes out.

These are self-run scores. They were submitted to the Decision Index in apolinario/decision-index#62 on 2026-10-05 and are pending the maintainers' validation. They are not board numbers until the maintainers add them, and no rank is claimed.

The three runs

Run directory	decisio tag	Base checkpoint	Decision Index 0.2.1	Raw index	Breadth	Request latency, median / p95
runs/decisio-0.4.0-qwen3.6-35b-a3b	v0.4.0	Qwen/Qwen3.6-35B-A3B-FP8 at 95a723d	48.06	60.45	46.55	70.5 / 261.2 ms
runs/decisio-0.4.0-gemma-4-12b	v0.4.0	google/gemma-4-12B-it at 707f0a3, bf16	49.43	61.47	47.57	37.8 / 275.9 ms
runs/decisio-0.6.0-gemma-4-31b	v0.6.0	google/gemma-4-31B-it at 842da37, FP8 on load	57.58	67.33	56.50	56.0 / 439.3 ms

The five areas, as each run's scores.json gives them (skill, 0 to 100):

Run	Knowledge & Reasoning	Language Understanding	Retrieval & Classification	Tools & Automation	Arts & Human Taste
Qwen3.6-35B-A3B	32.45	51.39	53.56	68.43	31.57
Gemma 4 12B	31.37	51.66	58.00	71.62	32.59
Gemma 4 31B	45.57	61.02	62.75	75.04	37.43

Every base is a frozen official checkpoint, read in one prefill per question from the option letters' logits, with no generated tokens.

Edition 0.3

• runs/decisio-0.4.0-qwen3.6-35b-a3b-edition-0.3: Decision Index Edition 0.3, decisio tag v0.4.0.
• runs/decisio-0.4.0-gemma-4-12b-edition-0.3: Decision Index Edition 0.3, decisio tag v0.4.0.
• runs/decisio-0.6.0-gemma-4-31b-edition-0.3: Decision Index Edition 0.3, decisio tag v0.6.0.

How they were run

• Kit: apolinario/decision-index at 87d4650 ("Version 0.2.1"), edition 0.2.1 (panel decision-index-0.2.1).
• Suite: rebuilt from its sources with the kit and checked with suite verify --edition 0.2.1 against the lab's hashes (base rows b2b56d6f…, added rows 7429f3c9…, both matching). It sends 150,759 requests, of which 150,317 are scored; the edition excludes the other 442 at scoring.
• Engine: the kit's http engine with --compact, one request at a time, against decisio's /v1/systemone on the same machine. Each base was served by its tag with --base <name> and the port, nothing else; each server's /health is in decisio's run record.
• Hardware: one NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96 GB), vLLM 0.30.0.
• Capacity: no declared capacity limit was reached; every scored request was answered (150,317 of 150,317), and none was unsupported, truncated or in error.
• Latency: the kit's HTTP wall time per request on this card, one at a time, including decisio's prompt construction and excluding server start-up. It is not the maintainers' measurement, which they take themselves on their own sample.
• The 31B run was resumed once, after a container restart on 2026-10-05: the same server was started again from v0.6.0 with an identical /health, and the kit sent only the requests not yet answered. Its status.json records the resumed part; its results.jsonl.gz holds all 150,759 rows.
• No decisio parameter was fitted on any Decision Index request. What was chosen with the Decision Index in view is stated in sections 4, 6.4 and 7.4 of decisio's EVAL_CARD.md.

Files

Each run directory holds the kit's files, unchanged:

• scores.json: the index, the raw index, the five areas and every benchmark, with "complete": true.
• index.json and benchmark-summary.json: the kit's summaries of the same scores, and request counts and latency per benchmark.
• environment.json: the engine, the served model name, and the hashes of the runner and the frozen corpus.
• status.json: the kit's last status event.
• results.jsonl.gz: one row per request, 150,759 rows.

No item text

These files hold no benchmark item text. No row has the request's payload (the text and the questions sent) or a raw model output; each row identifies its item by run_id, group_id and the payload's sha256, and holds decisio's answer, token usage and timings. decisio's answer gives each question's options with their probabilities, so option labels and question keys appear as the suite sent them (a tool's name, for example). The items come from the kit's sources and are rebuilt with the kit. This was checked on the files as uploaded: payload and raw_output are in no row and payload_sha256 is in every row; the files hold no credential and no figure for any other system.

Re-scoring and re-running

With the suite rebuilt by the kit at 87d4650 (suite rebuild, suite import, suite verify), the kit re-scores a run from its results file:

python -m decision_index score --edition 0.2.1 --suite-dir <suite> \
  --results runs/decisio-0.4.0-gemma-4-12b/results.jsonl.gz --out <new directory>

This gives back 49.43 for the 12B run. The commands to serve each base and run the suite again are in #62. A run on another machine, or in another session, can move individual probabilities. The measured sizes are in the EVAL_CARD: section 6.5 for the 12B, and 7.3 and 7.5 on main for the 31B, measured after v0.8.1.

Links

• The project: github.com/aminry/decisio.
• decisio's EVAL_CARD.md, section 8, and its run record runs/2026-10-05_decision-index (each server's /health, the downloaded checkpoint commits, the manifest).
• The submission: apolinario/decision-index#62.
• The kit: apolinario/decision-index.

Licence

decisio's outputs in these files are under the Apache License 2.0, as decisio is. The benchmark items are not included, and stay under their sources' terms.