How a Machine Used Its Genius to Solve a Game About Itself

On speedrun.com the published human record is held by christopho: 1:33:42 IGT · 1:33:46 RTA. On September 20 the controller used native timers only: 1:21:23 IGT · 1:21:19 RTA — about 12:19 under christopho's IGT and 12:27 under his RTA, which is the column the board actually ranks. That is plate 1 in the book. It is the run this page was first written to explain.

{
  "Quantum Computing": 395.6,
  "Release": 1290.4,
  "Space Exploration": 2963,
  "victory": 4883.0
}

Controller: sim.js with WR tune (cleanrun.ps1, budget set to the ~87-minute IGT bar). Site: official decisionproblem.com. Timers native — no virtual clock. Milestone flag 15. Matter 0. The game’s victory condition and our stop condition aligned.

The first record was not a machine that kept winning. The same route later finished in 1:23:31, then failed to finish at all, then collapsed in Space with the clip count frozen at six octillion and the fleet dying. A person watching the tab kept naming the same two things: yomi starvation, and a probe design that abandoned self-replication for speed and exploration at the wrong moment.

Both were right. Operations sat pegged at the cap, spilling into creativity, while the seven strategy projects waited for 171 factories. Buying them by hand — 162,500 ops, paid from overflow — moved Theory of Mind from 51:40 to 44:20 and Space Exploration from 52:05 to 50:27. That is now automatic at 20 factories.

The collapse was a one-way gate. The TAS exit schedule hands fifteen of twenty trust points to speed and navigation once the fleet is “done.” The TAS can afford that because it buys maxTrust past twenty with honor. We have honor 0 all stage. The gate also accepted an affordable clip bank in place of an actual fleet, with a probe floor of zero — so a stronger economy tripped it at ~1e29 probes and 22% explored. Replication went to nought. The fleet halved every thirty seconds.

Underneath both of those, the game's clock is a count of completed loop passes. Overlay HTML rebuilt twice a second, and a leftover local copy of the game in the same browser, held the loop at 67 ticks/s. A third of every second of replication, harvesting, and conversion simply never ran. That is why the “same route” could win once and stall the next time.

{
  "Release": 1291,
  "matter gone": 2611.7,
  "Space Exploration": 2973.5,
  "victory": 4856.9
}

The finish design held the argument: replication 12, hazard remediation 5, one point each to the three production dials, speed and exploration at 0 because the universe was already 100% found. Fleet 2.11e32 against 3.45e31 drifters. The previous attempt had been nine-to-one the other way.

The next morning the same tune was run again, to see whether 1:20:57 was the route or one finish. Edge had started delivering the game's 10 ms loop only 66 times a second, on a visible tab, with the anti-throttle flags still set. A bare setInterval(10) did the same. A message-channel timer, gated on performance.now so it cannot exceed 100 a second, brought the mean to 98.7 across 163 samples.

The game's sentence is a new best. The wall clock is not. Sixty-five seconds of missed passes turned an eight-second IGT gain into about a minute of real time lost. Plate 3 holds the IGT sentence; plate 4 took the RTA record two days later.

Christopho's human world record enters Space with about 146,000 yomi banked — enough to buy probe trust immediately and skip the slow tournament grind in the universe phase. Our Sep 23 route launched with fifteen thousand. The Space phase was fast for what we had, but we were playing with a smaller hand.

The fix was structural, not a dial tweak: hold the Space Exploration gate until the yomi bank nears 146,000. While dismantled and waiting, run tournaments through drone pauses. On Earth, pick strategies adaptively (ExtraTricky-style helper). Buy AutoTourney as soon as creativity allows; Theory of Mind doubles tournament yield before the bank fills. In Space, spend the bank on probe trust first — fourteen points in the first minute, twenty by 73:04 — then let replication carry the universe home.

{
  "Release": 1358.3,
  "matter gone": 2604.7,
  "Space Exploration": 3222.3,
  "victory": 4701.1
}

Earth cost four minutes to farm the bank. Space gave six back. Net: two minutes thirty-six under plate 2 on the wall clock — the number speedrun.com would rank. The game printed 1:17:26 IGT; the controller overlay read 78:21 RTA at the finish. Plate 5 took that title the next evening.

Plate 4 fixed Space by banking yomi. Plate 5 fixed Earth by never letting yomi idle after Release. The controller had been waiting for twenty factories before buying strategies or running tournaments; on screen that looked like eleven minutes stuck at 4,300 yomi while harvesters ramped.

The correction: buy AutoTourney on Earth as soon as Donkey Space exists (the Business trust≥90 gate never fires after Release). Buy the strategy field immediately with wrStratEarly. Farm with two strategies instead of eight. Yomi climbed from 4,900 at Release to 35,000 by minute 27 — then through the drone pauses to a dead-planet bank above 130,000.

{
  "Release": 1316.3,
  "matter gone": 2479.5,
  "Space Exploration": 3053.3,
  "victory": 4560.1
}

Earth and Space both moved earlier. The universe phase was slightly slower, but the wall clock still dropped two minutes — 1:16:21 RTA, 1:16:00 IGT. Seventeen minutes under christopho's published 1:33:42.

Plate 5's tune was right; a regression in sim.js was not. With wrProjects enabled, the project queue preferred wire extrusion and autoclipper upgrades immediately after Quantum — ops spends that starved creativity and pushed Hypno Harmonics from 9:28 to 16:34 on two consecutive live runs. The record plate 5 run never bought that chain at all.

The fix restored Business-phase ops discipline: only Creativity, the demand multipliers, Strategic Modeling, and Photonic Chip may draw operations before the yomi engine runs. The queue skips blocked ops projects instead of selecting them every tick. Catchy Jingle, New Slogan, and Hypno moved ahead of wire and autoclipper entries.

{
  "Release": 1260.5,
  "matter gone": 2339.3,
  "Space Exploration": 2867.1,
  "victory": 4389.9
}

Business, Earth, and Space all moved earlier. The wall clock dropped three minutes — 1:13:59 RTA, 1:13:10 IGT. Twenty minutes under christopho's published 1:33:42.

Plate 6 fixed the controller. Plate 7 asked whether the tune still had slack. Sim sweeps over thirty investment seeds were flat — with WrRng on, the market sequence is already pinned. Earth and Space knob sweeps moved seconds, not minutes. Two changes did carry:

• stake: 150000000 (was 170M) — Release 41 seconds earlier in sim; live matched at 20:21.
• wrExitRepFloor: 3 (was 5) — lower replication reserve at universe exit; Space finishes sooner.

{
  "Release": 1221.0,
  "matter gone": 2261.8,
  "Space Exploration": 2816.0,
  "victory": 4347.2
}

Forty-three seconds under plate 6 on both clocks. Still twenty-one minutes under christopho's published 1:33:42. Sub-hour remains a sim-policy problem, not a tune typo.

Plate 7 exhausted most tune knobs. A ramp sweep tested probe seed at Space launch: 200k, 1e6, 1e7, 1e8, 1e9. Sweet spot was 1e8 — sim victory 66:14, about fifty-one seconds under plate 7. Early exit thresholds (wrExitAt below 146 nonillion) and hazard cuts all hurt. explore-99 DNF'd at ninety minutes.

• wrProbeSeed: 100000000 — buy 100M probes at Space entry instead of 200k; replication curve starts higher.
• Everything else unchanged from plate 7 (stake 150M, repfloor 3).

{
  "Release": 1222.9,
  "matter gone": 2277.3,
  "Space Exploration": 2813.7,
  "victory": 4335.1
}

Twelve seconds under plate 7 on IGT. Sim/live gap persists: fifty-one seconds in sim became twelve on the wall clock. Sub-hour remains a Space replicate policy problem — not another tune typo. Still twenty-one minutes under christopho's published 1:33:42.

Plate 8 held the RTA column. A week later the same route, with yomi gates trimmed to 139k on both Earth and Space launch, printed 1:10:52 on the victory banner — eighty-three seconds faster than plate 8 on IGT. The controller overlay read 1:12:26 RTA, so this plate does not replace plate 8 on a speedrun.com-style sort. For PCIA the game clock is the primary score we optimize; the screenshot is the receipt.

Tune snapshot: snapshots/tune-plate8.json with wrLaunchYomi / wrEarthYomi at 139000, probe seed 1e8, stake 150M, repfloor 3. Logged on logs/decisions.jsonl run turns ~13815–13946.

Plate 9 landed once. October was about making it reproducible and fixing a cosmetic-but-real endgame mismatch: many runs ended at 3.00×10⁵⁵ clips while exploration sat below 100%. Sim sweeps suggested map-close gating; live Chrome runs did not beat plate 9 on IGT.

• Harness — paperclips/launch-chrome.ps1 and plate8run.ps1 prefer Chrome on port 9333 with throttle disabled.
• Mapfix — wrExitMinProbes drives wrExitAt; wrClipExitMapPct waits for ~98% map before clip-bank exit; wrDoneProd tuned for late replication.
• Verification — best clean live ~1:12:42 IGT (run 684618), Space 46:59, ~98 ticks/s. Sim-only wins on the same port are invalid.
• Next lever — Earth work/think slider staging (factories vs drone army) if the campaign resumes; not another probe-seed knob.

No new screenshot plate — this chapter is process. Campaign paused 14 Oct 2026.