RUN 2026 10 08

Decision Arena run, 2026-10-08

The first real (not mock) arena run. It cost nothing: only the two deciders with no model ran. It follows the wording rules.

• Files: public/content/arena/run-1c8b7b636f83.json (1,615,849 bytes) and replay-1c8b7b636f83.json (131,598 bytes), listed in index.json.
• Run: arena-20261008T194257-ee72038, captured 2026-10-08 at 19:42 UTC from RISE commit ee72038dac6e958066302c1ca04bc5c97c67b9f0 (kept reachable by tag arena-run-2026-10-08-harness), clean tree, Node v22.23.3.
• Inputs: the 39 authored cases and the 50 known-probability controls (scripts/arena/controls.json, seed committed before capture), each asked 3 times: 534 results.
• Check: node scripts/arena/arena.mjs report --run public/content/arena/run-1c8b7b636f83.json prints the same bytes twice, with matchesRecorded: true and matchesReplay: true.

Who ran

Decider	Model	Status
rules	none	ran
rules-floor	none (first offered option)	ran
openai	none called	not run: no key
jev	none called	not run: no key
kev	none called	not run: hardware/setup

Kev was tried on the operator's Mac (Apple M5 Pro, 48 GB) through npm run local. The launcher found the Apple Silicon path and installed pinned Kev into ~/.cache/rise-kev, using a standalone Python 3.12 in ~/.cache, because this Mac has only Python 3.9. The download of the Qwen/Qwen3.5-4B-Base weights stalled at about 0.2 GB of roughly 9 GB, at a few kilobytes a second. With the 20-minute setup cap in sight, the launcher was stopped. Kev never answered, so this run says nothing about Kev on Apple Silicon.

What the run shows

Every number below is for run 1, as the report's agreement block states it. Both deciders are deterministic: all three runs gave the same answer for every case and control (89 of 89 for each), so the repeats add no information, and the report's pooled scores totals (123 / 147 for rules) must never be given an interval.

Agreement with author-written expectations (49 explicit checks across 39 cases, admitted decisions; raw decisions give the same numbers):

Decider	Agreement	95% case-cluster bootstrap
rules	41 / 49 (0.84)	0.74 to 0.93
rules-floor	12 / 49 (0.24)	0.07 to 0.44

The interval resamples the 39 cases, because the checks within one case are not independent either.

Contrast pairs (19 pairs in which two cases should differ): rules 18 / 19 (95% Wilson 0.75 to 0.99), rules-floor 0 / 19 (0 to 0.17).

Difference (rules minus rules-floor, explicit agreement, paired case-cluster bootstrap over 39 shared cases): 0.59, 95% interval 0.38 to 0.80. The interval excludes zero (withinNoise: false). It shows only that matching words in the request agrees with these authored expectations more often than always taking the first option. The cases were written and tuned by RISE's authors, and their wording may favour word overlap. These numbers say nothing about any model.

Out-of-menu answers: 0 of 267 results for each decider. There were no refusals and no missing answers, and every result was admitted.

Latency: median 0 ms and 95th percentile 0 ms for both deciders. The slowest of the 534 calls took 2 ms. The time is local computation only, with no network.

Cost: $0. Neither decider calls a model.

Calibration: none. The deciders that ran report no probabilities or confidence, so all 147 expected-question rows were skipped. The controls stay sealed. Even when the seed is revealed, they test only whether a decider puts the stated odds on the stated options, and these deciders give no odds.

What it does not show

Nothing here compares OpenAI, Jev or Kev: none of them ran. The results hold for this RISE release (ee72038), these 39 cases and this date. Frozen results age, and a later run is a new file, never an edit.

---

Published from docs/decision-arena/RUN-2026-10-08.md. Edits made here are overwritten on the next push to main.