01 / allocation policy
Six make, two check
Put 75% of every tick into producing candidates.
READY
Language · 语言 · 言語
English by default.
Same page, automatic translation. Text is sent to GTranslate / Google when translating; wording may vary from the original.
Weekly small experiment · 2026-W46
A busy pipeline can make more work than it can honestly finish.
Give five teams the same action budget and the same hidden-quality candidate stream. Only reviewed valid candidates count. Change review capacity, expiry and queue size; keep blocked, expired and idle work visible.
Loading the seeded workflow…
01 / allocation policy
Put 75% of every tick into producing candidates.
READY
02 / allocation policy
Split the fixed action budget evenly.
READY
03 / allocation policy
Keep six review actions even when the queue is empty.
READY
04 / allocation policy
Make below a visible target; check at or above it.
READY
05 / allocation policy
Choose make or check from a seeded 50/50 draw.
READY
A queue bar shows produced candidates that have not crossed review. Candidate validity is hidden from every policy until review; the reader sees outcomes, not an oracle available to the policy.
Try this first
Press Show outcome. Then let one review action inspect three candidates. The six-maker team changes from a clogged queue to the strongest fixed allocation. Restore one candidate per review and shrink the queue to 16. Activity stays high; accepted work does not.
Every tick contains eight make-or-check actions. Production places one candidate with hidden seeded validity into a first-in, first-out queue. Review perfectly accepts valid candidates and rejects invalid ones. Unreviewed candidates expire after the deadline. A full queue blocks production.
Every policy draws from the same candidate sequence and sees only its queue and past reviews. The queue-aware rule uses no quality oracle: it makes below a visible target and checks at or above it. Fixed policies produce first and review second within each tick.
| Setting | 6 make / 2 check | 4 / 4 | 2 / 6 | Queue-aware | Random |
|---|---|---|---|---|---|
| Default | 167.250 | 334.750 | 167.500 | 331.828 | 317.031 |
| Perfect quality | 240 | 480 | 240 | 476 | 454.828 |
| 35% valid | 83.734 | 167.281 | 84.625 | 165.859 | 158.391 |
| Two per review | 334.734 | 334.750 | 167.500 | 443.891 | 333.063 |
| Three per review | 503.234 | 334.750 | 167.500 | 499 | 333.516 |
In the default batch, six-makers average 346 expiries, 72 blocked production actions and 62 candidates still waiting. Two-makers accept nearly the same total but spend 480 review actions idle. Similar accepted totals conceal overload on one side and unused capacity on the other.
When one review action handles three candidates, six-makers accept 503.234 on average. The balance is conditional on review throughput—not a universal maker/checker ratio.
This is a deterministic teaching world, not an employee or model productivity benchmark, quality-assurance prescription, agent-framework test, biological model or financial system. Candidate validity is a bit. Review is perfect and equally difficult for every candidate. Rework, correlated errors, reviewer mistakes, task heterogeneity, learning, safety severity, compute, money and human judgment are absent.
All 640 matched seeded worlds and 3,200 policy outcomes are retained across ten predeclared settings. No seed was removed. Accepted, rejected, expired, blocked, idle and queued outcomes remain separate. No private conversation text is included.
Method & parameter guide ↓Source package ↓All batch results ↓Read the model
Previously: Bridge or Bus? · Explore the eight small worlds.