run · GQ-011 / D-18 · 6 Sep 2026 · Mac Mini · native openai · gpt-6-astra

2026-09-06-j2

Does requested reasoning effort medium versus xhigh change usage, time or accepted work on one real brief (the E1 response-window summarizer)? Two clean arms, same seed, same prompt, no delegation.

measured with limitsn = 1 brief · 1 run per armeffort applied · null in bothboth arms pass 12/12

Preregistration Arms, and the effort actually applied Numbers, with n Acceptance verdict Runtime audit What it changes

11measured rows 2arms 5runtime-audit facts 5things it changes

Preregistration

what we expected, registered before inference · preregistration.json 2026-09-05T20:48:30Z

assignment: J2 / D-18 then E1
registered: 2026-09-05T20:48:30.689Z (6 Sep 03:48 UTC+07), status registered_before_inference
node · route: Shaans-Mac-mini.local · native openai / existing ChatGPT login · gpt-6-astra
arms · order: medium then xhigh; fixed order recorded as a confound
real brief: Implement reusable E1 response-window summarizer; winning accepted code used on actual first-three-owner E1 data
clean runs: same seed · same prompt · user config ignored · context management explicitly enabled · approval never · sandbox workspace-write, tool network disabled · no delegation
acceptance: 12 functional holdout checks + 2 provided tests + permitted-file scope + parent code review; first correct action = earliest captured version passing the parent holdout tests (file events + 1 s polling)
claim under test: factor 4, 'about four times'; metric from his chart not supplied → report separate ratios, no quota multiplier without debit evidence

Hashes registered before inference

seed/AGENTS.md: a6d75837…9236511
seed/BRIEF.md: b72244bc…19b8b2
seed/response-window.test.mjs: 261af503…637058
seed/context/observe.mjs: c2987380…2ddda
grade.mjs: 9610fa0a…c53a
run-pair.mjs: 16d79060…5293

Arms, and the effort actually applied

requested from run.json args · applied from the rollout's thread settings (metadata.json)

arm · medium

requested: medium
applied: null collaboration_mode.settings.reasoning_effort in the rollout's turn_context
window: 2026-09-05T20:48:34.706Z → 2026-09-05T20:51:01.031Z
thread: 01a07354-8af2-70f2-9966-1c0dfcb874d8
exit: code 0 · no violation · wall 146,325 ms

arm · xhigh

requested: xhigh
applied: null same field, same value; only requested settings are known
window: 2026-09-05T20:51:01.033Z → 2026-09-05T20:53:24.172Z
thread: 01a07356-c481-76e3-8c03-3d1365a17a14
exit: code 0 · no violation · wall 143,139 ms

Numbers, with n

acceptance.json usage · n = 1 brief, 1 run per arm

metric	medium	xhigh	ratio
input tokens	122,427	122,327	1.00×
cached input tokens	92,672	92,672	1.00×
uncached input tokens	29,755	29,655	1.00×
output tokens	4,101	4,027	1.02×
reasoning output tokens	240	248	0.97×
model responses	5	5	1.00×
user turns	1	1	1.00×
first correct action	68,135 ms	67,121 ms	1.02×
wall	146,325 ms	143,139 ms	1.02×
holdout checks passed	12 / 12	12 / 12	—
snapshots captured	1	1	—

Ratios are medium ÷ xhigh from acceptance.json. n = 1 brief and 1 run per arm: one real brief, one run per effort; not general model-quality evidence or a bill comparison. Quota or cost: not tested (no isolated subscription debit, invoice or per-run allowance measurement).

Acceptance verdict

j2.json · claim.verdict

inconclusive_effective_effort_not_verified

Both arms pass every holdout check and both provided tests; the medium-labelled output was selected for the real E1 analysis with no reasoning-quality winner inferred (selection_reason). Requested-effort runs are approximately 1× in usage, turns and time. Because reasoning_effort is null in both rollouts, this is not a verified medium/xhigh contrast: the 'about four times' claim is untested, not refuted.

Claim under test: 'Medium reasoning makes it stretch a lot further, the usage. about four times further' (Shaan, 6 Sep 03:00).

medium · acceptance

✓ twenty on each side; no padding or boundary leakage
✓ earliest duplicate time and key order cannot shift window
✓ usage conflict rejected
✓ turn and thread identity conflicts rejected
✓ counter resets do not change additive response accounting
✓ input remains untouched, including frozen rows
✓ tie order, ISO normalization and first-seen user turns
✓ missing IDs counted, other kinds ignored, identity null normalized
✓ empty totals and unknown optional totals
✓ invalid arguments rejected
✓ invalid reported usage and timestamps rejected
✓ top-level contract and default window

xhigh · acceptance

✓ twenty on each side; no padding or boundary leakage
✓ earliest duplicate time and key order cannot shift window
✓ usage conflict rejected
✓ turn and thread identity conflicts rejected
✓ counter resets do not change additive response accounting
✓ input remains untouched, including frozen rows
✓ tie order, ISO normalization and first-seen user turns
✓ missing IDs counted, other kinds ignored, identity null normalized
✓ empty totals and unknown optional totals
✓ invalid arguments rejected
✓ invalid reported usage and timestamps rejected
✓ top-level contract and default window

Runtime audit

facts from runtime-audit.json; confounds from j2.json are proposals until measured

fact · has a receiptproposal · not yet

Seed files unchanged in both work trees after the run: AGENTS.md, BRIEF.md, response-window.test.mjs, context/observe.mjs (4/4 unchanged: true per arm).runtime-audit.json seedChecks

Both invocations used --ignore-user-config with explicit -c arguments; launcher sha256 3b8363aa… identical before and after in both arms.medium/run.json · xhigh/run.json args, before/after

Rollouts recovered read-only by exact thread ID; medium rollout 192,878 bytes sha256 7fd47f21….runtime-audit.json source

Final implementation hashes differ per arm (medium 6f44c34e…, xhigh 7f209064…); parent review: pure computation, no imports or side effects.acceptance.json final_sha256 · source_review

Skill catalog hash identical and context guide non-empty in both arms; the same skill-load errors occurred in both.j2.json confounds[5] · medium/metadata.json world_state

Nine existing E3 owners were switched to medium around 03:04 by an actor outside Fable; effort drift would invalidate a clean comparison.confound; later resolved as Shaan's own change (source/2026-09-06-0345)

Sequential fixed order, cache state, shared-account/server load and small sample may affect timing and usage.confound, unmeasured

The Mini's global config hash changed during the medium arm, by an unobserved actor; both arms ignored user config.confound, unmeasured

The xhigh arm's initial SQLite metadata lookup failed; rollout recovered by thread ID.recovery path, not a result

Limits

Effective effort is not independently persisted in either exec rollout; both null values cannot be read as medium, xhigh or equal.
First-correct timing is the earliest captured passing version, not an instrumented model intention or server timestamp.

Version

j2.json status measured_with_limits · registered 2026-09-05T20:48:30Z

2026-09-06

What it changes

Agent entry

00_AGENT_ZERO/domains/compute/runs/2026-09-06-j2/preregistration.json

cd 00_AGENT_ZERO/domains/compute/runs/2026-09-06-j2
jq '.medium.usage, .xhigh.usage' acceptance.json
jq '.medium.records[] | select(.type=="turn_context") | .payload.collaboration_mode.settings.reasoning_effort' runtime-audit.json
node run-pair.mjs # re-run both arms on the Mini (native route)

Agent entry

Read first: 00_AGENT_ZERO/domains/compute/runs/2026-09-06-j2/preregistration.json
Owner: GQ-COMPUTE (closed) · last writeback 6 Sep 13:36
Machine: https://siso-shell.pages.dev/t/U19/example.json
Done when: Open this page and see two arms, 122,427 vs 122,327 input, ratio 1.00×, reasoning_effort null in both, and the verdict; then compare against acceptance.json.