TemplatesU5 · God QuestionGQ-011
GOD-QUESTIONS.mdJ2 run
GQ-011 · compute · host hold · priority 1 with GQ-008

How do Astra, Luna and delegation yield more accepted work?

Strong ownership versus bounded Luna work versus doing a small operation inline: which tier for which task shape, measured on accepted outcomes, not on tokens alone.

researchstate · measured with limitsmodules · 1 landed · 3 partial · 1 unallocatedLibrary page · GQ-011 is Mathematics (D-20); compute is GQ-023
The question and the decision it changesResearch stateEvidence mapMind mapCompute modulesExpected answer package
5evidence lines5register modules4pieces of the answer package3seats on it
01

The question and the decision it changes

GOD-QUESTIONS.md card · his words that raised it
Decision target

Strong ownership versus bounded Luna work versus doing a small operation inline: which tier, for which task shape, by accepted work, allowance, latency and recovery. Dependencies: GQ-008 (which model and route earn each task), GQ-012 (context), GQ-016 (owners without a bottleneck).

Counterexample rule

If, on ≥10 real dispatches of one task class on one unchanged route, the cheaper tier's accepted-work rate and repair cost are as good as the stronger tier's, the ownership premium is refuted for that class. Conversely, if setup + verification + repair erase the delegation gain, 'Luna for the dumb 80%' is refuted. Until then every multiplier is a proposal.

apparently reasoning's a big thing. A lot of things don't need more than medium reasoning, according to the chart. Medium reasoning makes it stretch a lot further, the usage. about four times further along with this config the compact oneShaan → Fable FA0 · 6 Sep 03:00
we needed a smart way to deploy lunar sub have astras use astra for reasoning and lunar sub agents for work because i'll be real we've just rinsed three weekly plans in the last 24 hours this is like unsustainableShaan → Fable FA0 · 6 Sep 03:45
we're going to go ultra compute efficient on this so every one of these guys are going down to medium thinking and they can use lunar sub agents to do 80 of the dumb stuffShaan → Fable FA0 · 6 Sep 15:30
If you can use Haiku 4.5 for sub-agents, that might be better … we're running low on our GPT and I kinda wanna reserve that for, like, more important shit … Just use Haiku for nowShaan → AGENT-ZERO · 6 Sep ~18:45
02

Research state

finding · unknown · next, from the card
Finding

One helper lost its assignment through a rewrite and completed after a file-packet workaround; earlier worker DONE gates were narrower than intended outcomes. P6's integrated tier / lean-dispatch rules use one Playbook source and the two originally checked harness links; September 6 native discovery exposed a conflicting Hub projection in .agents/skills, so a name-only PASS cannot certify P6.

Unknown

Net gain after setup, verification and repair. Account debit per route. Whether 'effort applied' can be read back from any rollout at all (J2: null in both).

Next

Score naturally occurring work by task class and accepted outcome; do not launch another benchmark swarm. Re-run J2 only when a seat can persist applied effort in the receipt.

GQ-008 routeGQ-012 contextGQ-016 owners
03

Evidence map

every fact cites a receipt path and a date and carries n; the rest is proposal
fact · has a receiptproposal · not yet
J2 · medium vs xhigh, one real brief, one run per arm. Input tokens 122,427 vs 122,327 (ratio 1.00×); output 4,101 vs 4,027 (1.02×); reasoning 240 vs 248; uncached input 29,755 vs 29,655; wall 146.3 s vs 143.1 s; first correct action 68.1 s vs 67.1 s; 12/12 holdout checks pass in both; 5 model responses each. reasoning_effort: null in both rollouts, so this is not a verified medium/xhigh contrast.domains/compute/runs/2026-09-06-j2/acceptance.json · runtime-audit.json · j2.json · 6 Sep 03:48–03:53 UTC+07 · Mac Mini · native openai
E1 · 20-response windows before/after the runtime context guide on three owners. INTENT-DISTILLATION: uncached input ×6.34, total ×1.09. GQ-COMPUTE: output ×3.22, reasoning ×4.27, uncached ×12.6, total ×1.26. 20 user turns per side were not available (15/10 distinct), and the E3 route change coincides: observational, confounded.domains/compute/runs/2026-09-06-j2/e1.json · 6 Sep
Oracle token-waste lesson. About 75 Astra seat-hours at xhigh landed nothing until the landing rule became a ratchet; then 44 commits in 15 seat-hours (the lesson's numbers); origin/main by commit date shows 13 then 23 commits.siso-harness-lab/docs/lessons/2026-09-06-oracle-fleet-token-waste.md · plan/projects.json#oracle · 6 Sep 15:31
Nine E3 owners switched to medium at 03:04: his own change in the Codex app, not a leak. Standing rule: medium everywhere except Oracle streaming.source/2026-09-06-0345-shaan-medium-was-me-luna-subagents-mini-agents.md · 6 Sep 03:45
D-30: sub-agents that find, read or list run on Claude Haiku 4.5 while the GPT/Codex window is reserved.plan/decisions.json D-30 · source/2026-09-06-1832-* · 6 Sep 18:32
Medium reasoning stretches usage 'about four times'.his chart, 03:00; J2 measured ~1× on requested effort with applied effort unrecorded
Luna is 'free / unlimited'; delegation carries ~75% overhead; 3×/4×/10× multipliers.GOD-QUESTIONS.md: 'not measured facts'
P6 tier table is certified across both harnesses and both machines.name-only PASS; conflicting Hub projection found in .agents/skills (p6-native-discovery.md)
Limits
  • One brief, one run per effort: not general model-quality evidence or a bill comparison.
  • No isolated subscription debit, invoice or per-run allowance was measured.
  • Effective effort is not persisted in either rollout; only requested settings are known.
Version
GOD-QUESTIONS.md · j2.json measured_with_limits
2026-09-06
04

Mind map

question → subquestions → modules → runs → Works
GQ-011Modules 4G11 · compute allocation, mea…G11 · compute allocation, measured — holdP6 · dispatch tier tableP6 · dispatch tier table — codex · partialE1 · overhead on 10 dispatchesE1 · overhead on 10 dispatches — codex · partialE2 · re-measure after P4/P5/P6E2 · re-measure after P4/P5/P6 — unallocatedRuns 22026-09-06-j2 · medium vs xhi…2026-09-06-j2 · medium vs xhigh — measured with limits2026-09-05-gq-compute · P6, d…2026-09-05-gq-compute · P6, dispatch overhead — observationalWorks 2GQ-023 · compute (Library)GQ-023 · compute (Library) — D-20 homeGQ-008 · routeGQ-008 · route — siblingSubquestions 4Which tier for which task sha…Which tier for which task shape? — G11Does medium ≈ xhigh on accept…Does medium ≈ xhigh on accepted work? — J2What does a dispatch cost end…What does a dispatch cost end to end? — E1 / E2Can both harnesses read one t…Can both harnesses read one tier table? — P6Decisions 3D-18 · J2 then E1D-18 · J2 then E1 — assignmentD-19 · Luna defaultD-19 · Luna default — in forceD-30 · Haiku sub-agentsD-30 · Haiku sub-agents — in forceGQ-011
Subquestions4
Modules4
Runs2
Decisions3
Works2

Each leaf links to its record. Branches are laid out by subtree size, not by hand.

05

Compute modules

allocated / done / unallocated by host, each with its verification line (plan/modules.json)
  1. 1
    G11 · GQ-011 compute allocation: which tier for which task, measured partial · host holdOpen the dispatch tier skill and see per task shape a measured token and acceptance row from at least 10 real dispatches where the applied effort is recorded in the receipt (J2 today: ratio 1.0×, effort null in both runs).
  2. 2
    P6 · dispatch: tier table as a skill both harnesses read; native Luna delegation partial · host codexOpen ~/.codex/skills/subagents and ~/.claude/skills/subagents and see the same tier table; one Luna dispatch from each harness returns a result and a measured overhead line lands in domains/compute/runs/ (today: wording only, no measurement).
  3. 3
    E1 · measure brief/handoff tokens vs output, time-to-first-correct-action, pings per outcome, on 10 real dispatches partial · host codexOpen domains/compute/runs/<run>/overhead.json and see, for 10 real dispatches on one unchanged route, brief+handoff tokens vs output tokens and time-to-first-correct-action (today: 20-response windows confounded by the E3 route change).
  4. 4
    E2 · cut and re-measure after P4/P5/P6 land; report the actual multiplier unallocated · host holdOpen the before/after page and see the same 10 dispatch shapes re-run after P4/P5/P6 with the multiplier stated and how many runs support it.
  5. 5
    G0 · GQ registry rewritten from the full corpus; falsifiers on all partial · host webuiOpen GOD-QUESTIONS.md and see 22 cards each carrying a decision target and a counterexample line (today 2 cards use the word falsifier).
Preregisterhashes of seed, brief, tests and graders before inference (preregistration.json)
Run both armssame seed, same prompt, user config ignored, no delegation
Gradeholdout checks + provided tests + scope + parent review (acceptance.json)
Auditseed unchanged, effort applied read from the rollout (runtime-audit.json)
Readratios with n; requested vs applied shown separately (j2.json)
06

Expected answer package

what would close it
A tier table with numbersper task shape: accepted-work rate, tokens in/out, time to first correct action, repair count, from ≥10 real dispatches on one unchanged route.
Applied effort in every receiptnot null: the rollout or the harness persists what actually ran.
A cost lineaccount debit or allowance per run, not a chart multiplier.
A decisionwhich tier by default for which class, written into the subagents skill both harnesses read, with the counterexample that reverses it.
07

Who is on it

closed 6 Sep 13:36 · host codex · MiniGQ-COMPUTE22 cards; J2 two clean runs; P6 native-discovery diagnosis; E1 windows for three owners; branch checkpoint/gq-compute-20260906-d24 (52 files, 2 commits ahead).
handoffs/GQ-COMPUTE.md
closed · host claude · laptopFable FA0Captured his 03:00 / 03:45 / 15:30 words; assigned J2 (D-18).
E2 · G11 re-run · host holdunallocatedNo seat; waits for a route where applied effort persists (GQ-008).
08

Timeline

runs · decisions · his words, 5–6 Sep
09-0509-06runs2026-09-06 03:48 · J2 preregistered 03:48:30; medium arm 03:48–03:512026-09-06 03:51 · J2 xhigh arm 03:51–03:532026-09-06 12:00 · E1 windows written (e1.json)decisions2026-09-06 02:30 · D-18 · J2 then E12026-09-06 02:40 · D-19 · Luna default2026-09-06 18:32 · D-30 · Haiku sub-agentshis words2026-09-05 17:12 · fleet respawn and roles2026-09-06 03:00 · about four times further (medium)2026-09-06 03:45 · rinsed three weekly plans; Luna sub-agents for work2026-09-06 15:30 · ultra compute efficient; medium + Lunas for the dumb 80%2026-09-06 18:32 · Haiku for sub-agentsseats2026-09-05 18:17 · GQ-COMPUTE first writeback2026-09-06 13:36 · GQ-COMPUTE closed
runsdecisionshis wordsseats
10

Agent entry

00_AGENT_ZERO/GOD-QUESTIONS.md#GQ-011
  1. sed -n '/### GQ-011/,/### GQ-012/p' 00_AGENT_ZERO/GOD-QUESTIONS.md
  2. cat 00_AGENT_ZERO/domains/compute/runs/2026-09-06-j2/j2.json | jq '.claim, .acceptance'
  3. node 00_AGENT_ZERO/domains/compute/runs/2026-09-06-j2/run-pair.mjs # re-run both arms (Mini, native route)
Agent entry
Read first
00_AGENT_ZERO/GOD-QUESTIONS.md#GQ-011
Owner
GQ-COMPUTE (closed) · next seat unallocated · last writeback 6 Sep 13:36
Machine
https://siso-shell.pages.dev/t/U5/example.json
Done when
Open the dispatch tier skill and see, per task shape, a measured token and acceptance row from ≥10 real dispatches with applied effort in the receipt (G11).