QuorumWiki
Jugar
Diseño

Measurements

Balance figures the bots take against the build, taken at Simulation version d51095e0.

Esta página aún no está traducida al español.

How these are taken#

Every figure on this page comes from playing the Balanced bot against the Simulation in packages/simulation/src/pacing.ts, with each user sending 0.02 requests a second and the seeds the Simulation itself uses. The tables were taken at Simulation version d51095e0 on 2026-11-16, by pnpm measurements. A change that moves pacing changes the tables, and the page is generated again with it.

Read a row as the time a milestone happened on the seeds the section names. A milestone no seed reached inside the run shows as never, which is a fact about the bot, not about the game.

Challenge pace#

The Balanced bot at the Challenge pace (CHALLENGE_TUNING), seeds 1 to 5, until 90 sim minutes. Times are sim minutes.

MilestoneMeasured
First research0.7
First upgrade6.6
Database split6.6
First Feature live1.4
Pre-seed11.4
First hire18.9
Second App Server46.9–67.7
First incident19.2–21.4
Seed35.1–42.3
Engineers at Seed2–4
Features live at Seed5
Margin at Seed entry49%–66%

The M2 gate asks for Seed on every seed (Scope and roadmap) and pacing.test.ts holds it. Nothing else in this table has a target.

Live pace#

The same bot at the live tuning (LIVE_TUNING), seeds 1 and 3, until 130 sim minutes. The live pace doubles growth and shortens the incident gap, so it reaches Seed sooner and recovers more often in the Eras after it.

EraMeasured
Pre-seed6.1
Seed26.6–32.5
Series A56.4–62.8
Series B96.5–107.7

Seed#

The Balanced bot at the Challenge pace, seeds 1 and 3, to Series A. Seed length is the time between the two raises.

MilestoneMeasured
Seed35.1–38.4
Search Engine placed37.2–40.7
WAF placed39.2–42.7
Second App Server46.9–53.9
Series A64.4–75.4
Seed length29.3–37.0

Seed length is 29.3–37.0 sim minutes, and growth sets most of it: from 5,000 to 25,000 users against a 35,000 market at g = 0.0016 takes 28 minutes on its own. The Team's Features raise growth and the market, which is what buys the bot room to build them.

The autopilot builds Features in catalog order, so Profile photos, which gates uploads, arrives late in Seed or in Series A. Until it ships the Company sends no uploads, and the Queue and Worker branch that takes them off the App Servers waits for it. The Seed Sandboxes play that branch instead.

Series A and Series B#

The Balanced bot at the Challenge pace, seeds 1 and 3, until 130 sim minutes.

MilestoneMeasured
Series A64.4–75.4
Second zone and standbys68.5–79.5
Series B118.8, never on some seeds
Users at 130 min117,157–224,676
Availability at 130 min96.35%–97.89%

Series A leaves at 150,000 users and Series B at 1,000,000 (Progression). The deepest run holds 117,157–224,676 users at 130 minutes, so the table says whether a seed raised Series B inside the window, not how long the Era takes on its own schedule.

The bot spends its Dev Hours on a fixed priority list: two Studies queued, no Spikes, and Features in catalog order. It places the fix a Design needs once the bottleneck names it. A player reads the Board better than that, so these times are an upper bound on the real ones.

In Series A the bot researches Multiple zones first, then spreads every App Server and Worker across zone B and gives every data Node a standby. Its autopilot ships Video, Chat, and Checkout, so media, rt, and buy reach a bot Company. It ships none of the Series B Features, so job, report, and ai do not arrive there, and the Sandboxes play them instead.

The bot spends its error budget several times over. A 99.9% week allows ten seconds of full outage, and its incidents do not spread the way the design intends. A slow query or a zone outage before Multiple zones takes a third of its users, and a Company that loses that many rarely recovers inside the next Era. Which seed survives Series A turns on a percent or two of load, so read the tables as the shape of the curve rather than a promise.

Series C to Public#

No recorded run reaches Series C. The bot ends at 130 sim minutes holding 117,157–224,676 users, against the 10,000,000 that Series C asks for. sandbox.test.ts plays the Series C to Public mechanics on built Designs, and no bot plays to the IPO, so no IPO time is recorded. What the checks cover lists the checks.

Exit and Legacy#

A second Company on the same seed, with Angel network 1, Pattern recognition 1, and Founder’s reputation 1 (10 Legacy), against the first Company, both played by the Balanced bot. Sim minutes.

SeedFirst: Pre-seedFirst: SeedSecond: Pre-seedSecond: Seed
111.438.410.538.2
211.439.510.532.6
311.435.110.534.9

The M7 exit check, a second Company that reaches its Exit in 3 to 6 hours, is not measured: the Balanced bot does not reach Public.

Recorded by hand#

These sections record what the bots cannot take: replay timings, the state of the interface, and the research calibration. They were taken at Simulation version 17 and describe that build. The pacing tables above come from the current one.

Challenges#

Verification time for each reference replay (wall clock). Each run replays a Challenge's reference solution from t = 0 at its own Step.

ChallengeStepsNodeworkerdChromiumFirefoxWebKit
Hello, World180140 ms48 ms122 ms75 ms120 ms
Split the Box240218 ms101 ms150 ms218 ms279 ms
Slashdotted3001.5 s0.96 s1.4 s1.8 s1.6 s
The Event Loop240191 ms121 ms210 ms189 ms156 ms
Too Many Connections2403.7 s2.3 s3.2 s3.1 s2.9 s

The longest replay, Too Many Connections (81 Nodes), takes about 2.3 s of CPU in workerd. That fits the Workers Paid default of 30 s CPU per request but not the Free plan's 10 ms, so the API Worker needs a paid plan; no limits.cpu_ms change is needed for these five. In Node, every other reference replay takes under 2 s, and Five Nines (40,320 Steps of 60 sim seconds) takes about 6 s. workerd and browser times for the newer Challenges are not measured.

Polish#

Recorded state of the interface at this version.

TopicRecorded
SettingsMute, Reduce motion, and Text at 150% in quorum.prefs.v1; M toggles mute. The text scale is --type-scale on the interface; board cards scale with the board's zoom, and the Research tree's Node boxes and spacing grow by 1.5 while a long item name wraps onto a second line. Copy save as text and Import a save move a Company as one QSV1: line, refused when its version is another build's
Board zoomZoomed out past 0.38× a card drops its rows, and from 0.36× it becomes a chip: its block, its name and a ring of its load, with its size pips, zone and footer gone. At a third of full size, every Node on the Far Away Sandbox still reads as a named block
Region linksA belt between two regions carries the round trip its calls pay under its midpoint: 75 ms NA↔EU and 160 ms EU↔APAC from rttOf, and 150 ms from a Users Node, which pays a user's two round trips. A link into a global Component pays no place, so it shows no plate
SoundIncident, Era, and achievement sounds, skipped when muted
StatusNode cards mark ok, warn, danger, and down with a dot, a triangle, a square, and a cross as well as colour
Request kindsPackets are one of four shapes in the kind's colour; kinds that share a colour share a shape
Reduced motionPackets hold still and become arrows under the rate each belt carries; the Bottleneck stops pulsing; panels and toasts appear at once
KeyboardArrows select the nearest Node, Enter moves focus into the Inspector and places a part in hand, L links the selected Node, Shift+T tidies, Shift+F fits, C collects, Tab reaches a Node list that opens any Node and the Dashboard's scroll area
ContrastLight --ok at L 0.50 and --danger at L 0.52, so status text is at least 4.5:1 on wells
Achievements40 of the 42 are detected each tick in a Company, never in a Sandbox or Challenge; Serial founder and Five Nines come from an Exit
GlossaryEvery entry of the glossary catalog and every CVE, as a page on the wiki, and a test checks its references; a term chip on a Tradeoff card, a Research item or a Security advisory opens its page in a new tab
DashboardFive annotation groups, each with a switch; yesterday's line under the Traffic and Latency series; a Request kind and a Region that narrow the per-kind and per-region panels; a table view for every time series; Cash per second among the Overview's stats, from the Samples' cash; an App Server row of copies and busy workers; an SLO row of availability this week, the error budget gauge, the SLA credit and the burn rate. Research marks read the Inbox, since no field on the Game dates a Study
Incident loopThe pager and its Ignored rules; the postmortem Mail per run with MTTD, MTTA and MTTR beside each other, a dash where the Game has no time; the Incidents row's three tables; and a timeline that splits each lane at its first signal and its acknowledgement, so the not-seen and awaiting-ack spans show beside the SEV-coloured run, the running Incident last. The HUD adds an SLA row under the runway while the credit pays

axe-core 4.11 (WCAG 2.0/2.1 A and AA plus best practices), before and after the contrast fix:

ScreenBeforeAfter
Title0 violations0
Lab, Team panel open0 violations; aria-hidden-focus incomplete (build dock)0; none
Lab, panel closed1 violation0
Daily drill (Challenge)1 serious color-contrast (.text-ok, 3.64:1)0

Insight by Era#

Measured at Simulation version 17, before the catalog grew, on seed 1 at 130 minutes, with the Balanced bot's own Design and a larger Analytics tap. The first column is research available in each Era then, and the last column is how long the tap took to buy all of it. The second column is the current catalog's total for the Era, summed from packages/simulation/src/catalog/research.ts. It is not a measurement, so the time column does not describe it.

These figures assume no Product lever. A lever raises a pool's share and shortens its Era's Studies, so the table is the no-lever baseline (Progression).

EraResearch in the Era (measured at version 17)Research in the Era (current catalog)Time to buy all of it (measured at version 17)
Garage455 ◇455 ◇~8 min at size 1, the Era is ~11 min
Pre-seed2,580 ◇2,880 ◇20–33 min at size 1, the Era is ~17 min
Seed11,150 ◇20,750 ◇~22 min at size 4 and 10%, for 6% of revenue; the Era is ~21 min
Series A59,000 ◇161,800 ◇~4 min at size 8 and 50% (the 512 samples/s cap), the Era is ~38 min
Series B78,000 ◇409,000 ◇~5 min at the cap

Calibration

Measured at Simulation version 17, before the catalog grew. Two bots, five seeds each, 130 minutes. The Balanced bot follows its priority list with 2 Studies queued and no Spikes. The Skilled probe (not committed) queues 8 Studies, spikes every priority Study, and fills the queue with the cheapest open items. Each cell is when the Era's last priority item completes, as a share of the Era, and how much of the Era's catalog was done in it.

EraBalanced: last priority itemBalanced: catalogSkilled: last priority itemSkilled: catalog
Garage47%7/1548%15/15
Pre-seed42–59%7–9/2165–81%11–15/21
Seed45–81%9/2053–78% (one seed 146%)11–20/20
Series A94–109%6–7/2165–149%10–15/21

The catalog counts in the denominators are the version 17 counts. The current catalog has 15 items in the Garage, 22 in Pre-seed, 36 in Seed and 58 in Series A, so the shares of the catalog above are not comparable with it.

Open points from the calibration:

  • tiered-caching costs media, and Video ships late in Series A, so it often completes in Series B. That is the recipe working as designed, but it pushes the Series A tail past 100%.

  • Series B's priority items (hosted-models, data-warehouse) need ai and report, which arrive with their Features. The bots stop at 130 minutes before they ship.

  • Postmortems: an action item goes first in the queue, so incidents slow other research.

  • The Seed figure is the no-lever baseline. upload limits the default recipe, so its first lever step should shorten the Era's Studies. The measurement waits for the mechanic.

Data tier#

Taken on 9 November 2026, before and after the Read Replica rebalance. A Design of one App Server and one Database, at the Pre-seed mix with every Node running, is given one fix. The table gives the most users it keeps at an Apdex of 0.95 or better and an error share of 2% or less, and what the fix adds to the Burn. The App Server is at its largest size, so the Database is the limit. "Small" is a Database one size above the smallest; "medium" is one size above that.

DatabaseFixUsers beforeUsers afterBurn added beforeBurn added after
Smallnone2,6452,64500
SmallOne size up5,5715,571$0.20/s$0.20/s
SmallCache5,1815,181$0.24/s$0.24/s
SmallOne Read Replica4,5964,791$0.20/s$0.14/s
SmallTwo Read Replicas5,9616,156$0.40/s$0.28/s
Mediumnone5,5715,57100
MediumOne size up11,03311,033$0.40/s$0.40/s
MediumCache8,6928,692$0.24/s$0.24/s
MediumOne Read Replica9,2779,472$0.40/s$0.28/s
MediumTwo Read Replicas8,10711,423$0.80/s$0.56/s

The stale reads a replica adds fell from 3.2 to 0.8 a second at the Medium row, and from 3.2 to 1.5 with two replicas. Before, the second replica made the Design worse, because every replica added stale errors faster than it added capacity. The replicas now cost 0.7× the box, so one replica adds about 13,900 users for each dollar a second against 13,700 for a size step and 13,000 for the Cache.