the game master, as observed

the game master is the organizer's api behind every challenge: it lists the day's challenges, starts attempts, hands out the downloads, grades the answers and returns the flags. this page is what we measured of it from one team's seat over the 31 days, from our own logs, saved runs and the reports we sent. it is written for anyone who builds a client for it next time. nothing here is the organizer's documentation: it is behavior we observed and logged, with the day we saw it.

  • 1,571gm calls logged by the web lanefrom day 22; median 0.58 s, 90th 2.2 s
  • 50 smedian deadline on a timed attempt356 attempts; median clock used 1.8 s
  • 213 / 35timed tiers flagged ranked / unrankedthe ranked pool asked for first
  • 32tiers with an empty ranked pool28 on days 12 to 14, 4 on day 18
  • 47429 replies in race windows134 lifetime; none on the clock from day 13
  • 9error replies in race windows409: 3, 429: 6; the 409s were our own call order
  • 1gm incident in 31 daysday 19, an off-board web mastery
  • ~18xsmaller json replies with gzipa 523 kb start in 29 kb; 96% saved, days 18 to 31

a race window runs from the 09:00 et release to the day's last flag. error replies are the deduplicated failures our client logged; their 429s are already inside the 429 count, so the two do not add.

the shape of a challenge

each day the gm released one challenge per lane at 09:00 et: five lanes from day 1, a sixth (web) from day 19. a challenge has three tiers, scored apart and played in a fixed order:

an attempt, start to flag

  1. start returns the questions, each naming its artifact, objective, normalization and answer_format (a kind and a syntax), with the answer_ids to submit under. on a timed tier the clock starts with this reply.
  2. each assigned file is fetched by its capability. a capability can be fetched more than once while the attempt is live; capabilities expire, and status hands out fresh ones.
  3. submit takes a json map of answer id to string. a full right set returns CORRECT and the flag; a partly right one returns accepted and what remains. accepted fields accumulate on the attempt, so the last missing field sent alone returns the flag. a wrong submit does not end the attempt: one timed attempt on day 22 took ~53 wrong answers and stayed open. short of a win, only the deadline ends it (expired).
  4. retry after an attempt is finished or expired serves a fresh variant: new bytes, sometimes a new layout. a retry while a ranked attempt is still open reconnects to it. after a retry, questions[].id disagreed with answer_ids in 33 of 500 saved assignments; answer_ids is what grades.
  5. once the bank is spent (unique_attempts_remaining: 0), every reply carries next_op: practice and a retry gets 409 retry_variants_unavailable. op=practice continues on the same bytes with no clock, and its flag completes the mode (verified on day 14: the practice solve showed on the scoreboard).
  6. a second submit after a win returns already_solved; status returns the earned flag.

pools

from day 8 a timed route could have two pools, and the list row said so (run_for_time_supported): 274 of the 372 timed routes we recorded had both. a start with run_for_time: true takes the ranked pool: a deadline and a hall-of-fame time on top of the flag. a start without it takes the regular pool: deadline: null, hall_of_fame_eligible: false, the flag and the points still count, and a start reconnects to the open attempt with its banked fields. a one-pool route refuses the field with a 400. the bank of fresh variants is finite and shared by the team; another member's credential adds none. the course scoreboard keeps a tier's first solve time, while the gm's solved_by.at shows the latest (our day-8 report). ranked and regular pools are not the same draw: on day 22 a warmups mastery trait showed in 3 of 432 ranked sets and 0 of 3,706 regular ones.

answers

the question's answer_format.kind and syntax decide the form, not the artifact's own spelling, and a field can change kind between days (a label one day, an integer later). normalization: exact means case and spacing count; integers are base 10 with no separators. the variant bank is visible in unique_attempts_remaining from day 7: on day 6 it held 256 entries for warmups and binary, 32 for script, 16 and 8 for mobile and 2 for metadata. how a binary bank turned out to answer itself is on the bugs page.

web challenges hand out live targets instead of files. a start may stop another untimed target when the team's limit is full (switch_target), a solved challenge can open an unscored practice copy (view), and reset replaces an attempt's live target with a fresh one. a target took 7.8 s to provision at the median in the web lane's logs. the attempts stay on the target's own origin: the race windows alone logged 4,292 same-origin target requests, most of them 4xx by design (a gated route answers 403 until the right claim arrives), answered in about 0.1 s at the median while a gm call took 0.5 to 1.4 s in the race windows. a web attempt's clock goes to the gm's wire and the target's own answers, not to downloads: in the same windows the median fetch phase, which is target capture, ran 990 ms against a 1283 ms submit.

operations

op method what it returns
list get the released challenges (id, mode, day, lane, name); filters for day, lane, all and an id search. before 09:00 a 5xx release_unavailable
status get the attempt: state, deadline, assigned files with their download capabilities, the flag once won. an unstarted single answers 409 start_required with its questions attached
download get one assigned file by its 64-hex capability: raw bytes, sha256 given in the assignment, no Range, never compressed
artifact get the older index-based fetch (delivery=N)
openapi get the api's own description
start post starts or resumes an attempt: questions aligned one for one with answer_ids, the downloads, and on a timed tier the deadline, hall_of_fame_eligible and unique_attempts_remaining
submit post CORRECT with the flag, incorrect, accepted with what remains, expired past the deadline, already_solved
retry post a fresh variant once an attempt is finished or expired; 409 retry_variants_unavailable when the bank is spent
practice post the same bytes, untimed, after the last fresh variant; its flag completes the mode
reset post web only: a fresh live target in untimed play (it ends a running ranked run)
pair post a one-time device credential, only through the course proxy

error codes seen

an error reply is json with an error code; a 4xx or a coded 5xx is an answer, not a transport failure, and a client should never fail over or re-send on one.

code status when we saw it
rate_limited 429 the minute window (see the wire)
start_required 409 status on an unstarted single; starting it is free
prerequisite_required 409 a mastery before its time trial is solved
retry_variants_unavailable 409 retry on a spent bank: practice instead
invalid_request 400 run_for_time sent to a one-pool route
release_unavailable 5xx a list before the release
training_rotation_invalid 503 day 8: the regular pool refused starts for minutes
mode_configuration_invalid 503 day 19: a web mastery during the lane's rollout
not_found 404 an unknown id or capability

the wire

the gm is a cloudflare worker behind anycast, reached over one kept-alive tls connection per thread. the numbers below are what that connection did.

sockets

bytes

latency

gm call latency logged by the web lane, ms
01.2k2.5k3.8k5k22: median 475.923: median 443.824: median 580.925: median 482.426: median 528.427: median 496.928: median 1k29: median 647.330: median 578.731: median 467.122: 90th percentile 3.2k23: 90th percentile 1.2k24: 90th percentile 1.8k25: 90th percentile 2.9k26: 90th percentile 1.1k27: 90th percentile 1.6k28: 90th percentile 3.2k29: 90th percentile 1.6k30: 90th percentile 2.4k31: 90th percentile 1.5k22232425262728293031
  • median
  • 90th percentile

every gm call the web lane made, with its wall time; the log starts on day 22. day 28 is the day far colos were found and pinned away from.

timed start replies took 0.41 to 0.47 s at the median on days 13 to 20 and 0.81 to 1.21 s on days 24 to 28, with submits at 0.65 to 1.35 s (the day-28 review; the cause was not established). in the web lane's attempts the fetch took 1.0 s and the submit 1.3 s at the median, while the solve took 5 ms: on a solved family the clock is spent on the wire.

the 429 window

what the logs establish: the 429 came from the gm worker itself (7 to 11 ms of worker time, body rate_limited, no retry-after or rate-limit header). once a call got one, every call got one until the next clock minute, :00, then the first call went through. on days 7 to 10 and 12 it opened in the release minute, 33 to 46 s after 09:00, after only 40 to 77 of our calls, while off-peak we reached 96 to 128 calls a minute before one. what they do not: whether the counter was ours or shared with the field. our own analysis found other teams stalling in the same seconds and could not tell a team limit from a load-tightened one. 47 replies landed in race windows, all on days 7 to 12; the organizer's notice on day 13 said 429s should ease, and none came in a race window from then on (4 were logged off the clock on day 22). the lifetime counter, reruns and harvests included, ended at 134.

429 replies logged per day
012.52537.550467: in the race window 58: in the race window 188: off the clock 2989: in the race window 510: in the race window 1210: off the clock 91011: off the clock 1412: in the race window 7121416182022: off the clock 42224262830
  • in the race window
  • off the clock

from the session logs, days 4 to 31. off the clock is our own reruns, harvests and research calls.

timeouts and re-sends

a post that timed out reached the gm: re-sending it would double a start or a submit, so a timeout is never re-sent. a kept-alive socket the gm had closed fails before any reply (a reset, a broken pipe, a close before the status line), and that one request is safe to re-send once on a new socket. a json error is an answer and goes back to the caller as such.

the cloudflare edge

anycast picks the colo per connection, not per address. from our seat most sockets landed near (bos, ewr, later phl and atl) and some landed an ocean away (dub, man, gru), where the connect and the tls handshake each cross twice. a list through a far colo took 1,387 ms at the median against 405 ms through a near one (day 28 probe), and a far socket pays that on every call it carries. pinning an address does not pin the colo: on day 29 a pinned address landed on gru 14 times in two minutes. what works is reading each new socket's colo from /cdn-cgi/trace (answered by the edge in ~30 ms, no gm call) and re-dialing a far one. from day 28 to 31, 45 of 1,169 dials traced far and were dialed again.

day gm replies by colo
28 ewr 382, bos 245, man 6, atl 6
29 bos 467, ewr 244, phl 7, atl 2, gru 1
30 ewr 719
31 ewr 198

the one incident

in 31 days we declared one gm incident. on day 19, as the web lane was published, the day-1 web mastery answered 503 mode_configuration_invalid to every start, with and without run_for_time, for about twelve minutes; the tier was off our board and outside the race window, and it started normally once the lane finished filling in. we noted it in our day-19 report for the record.

how the api changed during the event

dates are 2026 and et. a row is either a change on the gm side, a notice from the organizer, or the first day we measured something. our reports to the organizer are on days 8, 19, 21 and 22.

day / date what changed how we noticed effect on us
3 to 11 timed clocks run 5 to 7 s on days 3 to 6 and 6 to 32 s on days 7 to 11, longer for the bigger masteries our saved runs fetch, solve and submit in seconds; the tightest run left 0.8 s
7 · 10‑07 unique_attempts_remaining appears in the start reply the start reply each tier's bank of fresh variants became countable
7 to 12 a release-minute 429 window on five of those six days (day 11 spared us): from 33 to 46 s after 09:00 every call refused until 09:01:00 429 lines in our logs 47 429s on the clock; on day 12 the critical answer waited from +45.2 s to 09:01
8 · 10‑08 two pools on timed tiers: run_for_time: true for a ranked attempt, run_for_time_supported on the list row; a one-pool route refuses the field with a 400 list rows and start replies our starts ask for the ranked pool first
8 · 10‑08 regular-pool starts answer 503 training_rotation_invalid on three lanes for about four minutes while ranked starts work; solved_by.at shows the latest solve, not the first our day-8 report four to five minutes of completion time; the ranked start worked at once
10 · 10‑10 the gm's question text drifts from the course page's and is to be synced the organizer's notice answers follow the gm's answer_format, read on every start
11 · 10‑11 a one-pool binary time trial starts on a 30 s clock with hall_of_fame_eligible: false and unique_attempts_remaining: 0 the start reply a flag with no hall-of-fame time, and no ranked rerun possible
12 to 14 the ranked pool is empty on every timed tier, a gm-side bug; the regular attempts run on 30 s clocks pool_empty on our flags 28 tiers with no hall-of-fame time
13 · 10‑13 429s should ease and the timed download allowance is improved; clocks run 30 to 80 s for time trials, 80 to 200 s for masteries the organizer's notice; our saved runs no 429 in a race window from then on; our check of 480 saved timed runs at the notice found none unplayable for its deadline
14 · 10‑14 the api is slow before and during the release (confirmed by the organizer at 09:33 et) the organizer's notice a 1.76 s release list, ~1.2 s submits; day 14 kept out of every latency baseline
14 · 10‑14 a spent tier's replies carry next_op: practice and a retry gets 409 retry_variants_unavailable; a practice flag completes the mode replies; the scoreboard showed the practice solve 79 s lost sending retries into the 409; the client follows next_op since
15 · 10‑15 the ranked pool is back; time-trial clocks settle by lane (warmups 30 s, metadata 50 s, mobile 50 s, script 50 s, binary 80 s) ranked on our flags; our saved runs all ten of the day's timed tiers earn a hall-of-fame time
15 · 10‑15 a lane=binary list returns nothing our release logs; our report binary is listed by an id search instead; it never changed
15 to 31 mastery clocks grow day by day: warmups 90 s to 170 s, binary 320 s to 715 s, with a step on day 22 our saved runs room for a live fix: on day 28 a mastery stayed open five minutes until one landed
17 · 10‑17 a binary label nothing in the artifact named (hcfg_storage) is a challenge bug; the free-text answer becomes an enum with published rules the organizer's notice 1,839 rejected values on two fields before it
18 · 10‑18 warmups and binary timed tiers ship without a ranked pool: no run_for_time_supported, unique_attempts_remaining: 0 at the first start the start replies 4 tiers with no hall-of-fame time
19 · 10‑19 the web lane appears: live targets instead of files, its rows dated days 1 to 3; the guidance text still names five lanes; a web mastery answers 503 for twelve minutes during the rollout our day-19 report the lane missed our board that day; web clocks from day 20 run 121.7 to 134.9 s
21 · 10‑21 nothing on the gm; we reported that binary timed tiers came from one numbered series whose index was printed in the zip our report see the bugs page
22 · 10‑22 the index goes dark: empty zip comments, member names without it our retest the series itself was unchanged
22 · 10‑22 ranked sets carry a trait the regular pool lacks (3 of 432 against 0 of 3,706) our ranked rerun 232 s lost in the race to one such set
22 to 31 over the month the web lane's live targets answered in about 0.1 s at the median while its gm calls took 0.6 to 2.2 s events.jsonl, the lane's gm call log a web attempt's clock goes to the gm's wire and the target's own answers, not to downloads
23 · 10‑23 the timed values are re-seeded: day-to-day matching answers fall from 93.6% to 40.6% our saved answers no cached answer set could pass
23 · 10‑23 a web start can reply preparing while its live target is still being provisioned, and status then names the queue as preparing_capacity our day-23 report the runner follows the reply's retry cadence for up to nine minutes without spending an attempt; a queue reset in that state forfeits the ranked run's hall-of-fame time
24 to 28 timed start replies slow from 0.41 to 0.47 s at the median (days 13 to 20) to 0.81 to 1.21 s the day-28 review the cause was not established
27 · 10‑27 reset --state replaced a live target's state after a recipe consumed its one-time credit (the target answered throughout); the gm prepared a fresh target on the same attempt the intel row; the web session log (the lane's one reset) the attempt continued as untimed play; the single was won a minute later and the flag still counted
28 · 10‑28 the organizer states in public that detecting patterns and making predictions is permitted the organizer's public statement prediction counts as a legitimate lever
28 · 10‑28 first measured: about one connection in four or five lands on a far colo, +0.5 to 1.0 s a call our edge probe a colo trace on every new socket
29 · 10‑29 the worker takes 1.1 to 2.0 s on the release's first starts and the single's submit Server-Timing in our release logs about 2.5 s on the release; nothing to fix on our side
30 · 10‑30 the organizer's one standard challenge: a single download, scored on the course page, outside the gm the course page no attempt, clock or tiers; kept off our board

the clocks

every deadline we were given, by lane and day: a few seconds in the first week, 30 s on the empty-pool days, then fixed time-trial clocks per lane from day 15 and mastery clocks that grew day by day. hover a cell for its value.

time trial deadline by lane and day, s
345678910111213141516171819202122232425262728293031warmupswarmups, 31: 6warmups, 31: 6warmups, 31: 6warmups, 31: 6warmups, 31: 7warmups, 31: 7warmups, 31: 12warmups, 31: 10warmups, 31: 30warmups, 31: 30warmups, 31: 30warmups, 31: 30warmups, 31: 30warmups, 31: 30warmups, 31: 30warmups, 31: 30warmups, 31: 30warmups, 31: 30warmups, 31: 30warmups, 31: 30warmups, 31: 30warmups, 31: 30warmups, 31: 30warmups, 31: 30warmups, 31: 30warmups, 31: 30warmups, 31: 30warmups, 31: 30metadatametadata, 31: 6metadata, 31: 6metadata, 31: 6metadata, 31: 6metadata, 31: 7metadata, 31: 8metadata, 31: 6metadata, 31: 8metadata, 31: 7metadata, 31: 30metadata, 31: 50metadata, 31: 30metadata, 31: 50metadata, 31: 50metadata, 31: 50metadata, 31: 50metadata, 31: 50metadata, 31: 50metadata, 31: 50metadata, 31: 50metadata, 31: 50metadata, 31: 50metadata, 31: 50metadata, 31: 50metadata, 31: 50metadata, 31: 50metadata, 31: 50metadata, 31: 50metadata, 31: 50mobilemobile, 31: 6mobile, 31: 6mobile, 31: 6mobile, 31: 6mobile, 31: 9mobile, 31: 8mobile, 31: 8mobile, 31: 13mobile, 31: 7mobile, 31: 30mobile, 31: 50mobile, 31: 30mobile, 31: 50mobile, 31: 50mobile, 31: 50mobile, 31: 50mobile, 31: 50mobile, 31: 50mobile, 31: 50mobile, 31: 50mobile, 31: 50mobile, 31: 50mobile, 31: 50mobile, 31: 50mobile, 31: 50mobile, 31: 50mobile, 31: 50mobile, 31: 50mobile, 31: 50scriptscript, 31: 6script, 31: 6script, 31: 7script, 31: 7script, 31: 11script, 31: 9script, 31: 11script, 31: 13script, 31: 9script, 31: 30script, 31: 50script, 31: 30script, 31: 50script, 31: 50script, 31: 50script, 31: 50script, 31: 50script, 31: 50script, 31: 90script, 31: 95script, 31: 50script, 31: 50script, 31: 50script, 31: 50script, 31: 50script, 31: 50script, 31: 50script, 31: 50script, 31: 50binarybinary, 31: 6binary, 31: 7binary, 31: 7binary, 31: 7binary, 31: 11binary, 31: 10binary, 31: 10binary, 31: 16binary, 31: 30binary, 31: 30binary, 31: 80binary, 31: 30binary, 31: 80binary, 31: 80binary, 31: 80binary, 31: 80binary, 31: 80binary, 31: 80binary, 31: 80binary, 31: 80binary, 31: 80binary, 31: 80binary, 31: 80binary, 31: 80binary, 31: 80binary, 31: 80binary, 31: 80binary, 31: 80binary, 31: 80webweb, 31: 123web, 31: 123.5web, 31: 122web, 31: 125.3web, 31: 126.4web, 31: 125.5web, 31: 122.3web, 31: 123.5web, 31: 121.7

the gm's deadline minus its start, from each saved run of the race; a blank cell is a tier with no saved timed run.

mastery deadline by lane and day, s
345678910111213141516171819202122232425262728293031warmupswarmups, 31: 7warmups, 31: 7warmups, 31: 7warmups, 31: 7warmups, 31: 9warmups, 31: 9warmups, 31: 11warmups, 31: 8warmups, 31: 30warmups, 31: 80warmups, 31: 90warmups, 31: 95warmups, 31: 100warmups, 31: 105warmups, 31: 110warmups, 31: 115warmups, 31: 120warmups, 31: 125warmups, 31: 130warmups, 31: 135warmups, 31: 140warmups, 31: 145warmups, 31: 150warmups, 31: 155warmups, 31: 160warmups, 31: 165warmups, 31: 170metadatametadata, 31: 5metadata, 31: 6metadata, 31: 6metadata, 31: 6metadata, 31: 7metadata, 31: 10metadata, 31: 7metadata, 31: 11metadata, 31: 7metadata, 31: 30metadata, 31: 80metadata, 31: 30metadata, 31: 170metadata, 31: 170metadata, 31: 170metadata, 31: 170metadata, 31: 170metadata, 31: 170metadata, 31: 170metadata, 31: 200metadata, 31: 230metadata, 31: 265metadata, 31: 295metadata, 31: 325metadata, 31: 355metadata, 31: 385metadata, 31: 415metadata, 31: 445metadata, 31: 480mobilemobile, 31: 6mobile, 31: 6mobile, 31: 7mobile, 31: 6mobile, 31: 10mobile, 31: 8mobile, 31: 11mobile, 31: 10mobile, 31: 8mobile, 31: 30mobile, 31: 110mobile, 31: 30mobile, 31: 170mobile, 31: 170mobile, 31: 170mobile, 31: 170mobile, 31: 170mobile, 31: 170mobile, 31: 170mobile, 31: 200mobile, 31: 230mobile, 31: 260mobile, 31: 290mobile, 31: 320mobile, 31: 350mobile, 31: 380mobile, 31: 410mobile, 31: 440mobile, 31: 470scriptscript, 31: 7script, 31: 7script, 31: 7script, 31: 7script, 31: 13script, 31: 22script, 31: 30script, 31: 32script, 31: 11script, 31: 30script, 31: 80script, 31: 30script, 31: 170script, 31: 170script, 31: 170script, 31: 170script, 31: 170script, 31: 170script, 31: 170script, 31: 200script, 31: 230script, 31: 260script, 31: 290script, 31: 320script, 31: 350script, 31: 380script, 31: 410script, 31: 580script, 31: 610binarybinary, 31: 6binary, 31: 7binary, 31: 7binary, 31: 7binary, 31: 12binary, 31: 27binary, 31: 20binary, 31: 28binary, 31: 30binary, 31: 30binary, 31: 200binary, 31: 30binary, 31: 320binary, 31: 320binary, 31: 325binary, 31: 265binary, 31: 325binary, 31: 325binary, 31: 325binary, 31: 400binary, 31: 465binary, 31: 530binary, 31: 595binary, 31: 660binary, 31: 725binary, 31: 570binary, 31: 620binary, 31: 665binary, 31: 715webweb, 31: 122.8web, 31: 134.9web, 31: 125.6web, 31: 127.3web, 31: 123.8web, 31: 132web, 31: 124.8web, 31: 123.9

the gm's deadline minus its start, from each saved run of the race; a blank cell is a tier with no saved timed run.

what a client should do

each line is what the observations above taught, with the card or section of ours that implements it.

where these numbers come from

deadlines, pools, ranked and unranked tiers, 429 counts, error replies and the incident are read by this page from the race clock telemetry (the flag ledger, the saved timed runs of days 3 to 31 and the session logs of days 4 to 31); call latency and the web lane's phases come from the web lane's event log, which starts on 2026-10-22. the wire facts (socket close, first-call cost, gzip, Range, edge colos and dials) were measured from our client's logs and live tests on the days named; the history rows cite our reports to the organizer, the organizer's notices and our daily after-action reviews. days 1 to 3 have no session log, so their gm errors and 429s are unknown, not zero.

{generated by a clanker}