rules paid for in seconds
the runbook every solver read on race morning ran to 882 lines by day 31, and nearly every rule in it has the same four parts: the rule, the day it was learned, what that day cost, and the mechanism that enforces it now. it exists because dead time lost races, not slow solving: a value nobody submitted, a cause nobody posted, a claim nobody advanced. this page lays the rules out in day order, one card each, with the loss that bought it and what refuses the mistake today. every cost is the runbook's own figure, a lane brief's, a day's review or the race telemetry; a card with no time names the outcome instead.
- 57rules with a dated lossdays 3 to 31
- 5886 sthe most expensive lessonday 17, 1h38m on two labels
- 11,705 sthe dated losses that state a time3h15m over 31 rules, each loss once
- 20mechanisms that refuse or capguards, refusals, caps, locks
- 39commands refused by the solved guardshell commands, over the month
- 26commands refused by the scratch guardthrowaway scripts, scratch writes, installs
- 8fields that reached the guess cap25 rejected values on one field
- 16 of 31days finished under 600 srelease to the last flag, web included
the sum adds each card's stated time once (minutes and m:ss converted to seconds); a range, a lower bound, a replay and a second card on the same loss are left out. the last four tiles are the race telemetry's.
the timeline
each card is a rule. the day badge is the race that paid for it, coloured by the size of the loss (green: a board place or a minute; cyan: tens of seconds; grey: a turn or hygiene). the yellow badge names what enforces it; a dashed one means nothing mechanical does. the top edge says what the rule is about: green submitting, cyan the solver edit, purple claims and coordination, orange the shell and tools, pink the parents, yellow the game master and the clock.
days 1 to 7
days 1 and 2 opened late (2h28m on day 1 and 6:07 on day 2 from release to the first start) and priced nothing; from day 3 the week taught that two agents on one attempt, one field or one lock cost more than either alone.
one runner per timed tier
two agents ran retry loops on the script mastery. each start replaced the team's attempt, so they cancelled each other: one of them logged 210
expired.6 min and 7 attempts on one mastery
- day 03
- enforced by the timed lock (a second
startexits 4,busy)
in
AGENTS.md "solving" 4,gm.pyno clock before the solver reproduces the single
a helper started the binary time trial on a solver with
D03_FORMAT=Noneand no symbols reader. it raised aKeyErroronce the clock ran and sat until the parent noticed it idle and re-ran it.60 s+ idle on a started clock
- day 03
- enforced by the runner's preflight (no clock while the solver disagrees with the graded single)
in
AGENTS.md "solving" 4,src/ctfkit/hlib/runner.pyone guesser per label field
three agents probed the same binary format label: 221 unique guesses before the right one and 108 duplicate calls. the probes peaked at 521 calls a minute and drew 38 http 429s.
9.5 min on one field
- day 03
- enforced by
gm.py guess(one guesser a field, exit 4 for the rest; every rejected value skipped; probes share one pace)
in
AGENTS.md "sharing",gm.pymail is pushed, never pulled
mail was pull-only and nobody ran
inboxmid-task. five messages to the parent went unread, among them a helper's field split and an urgent question for the format that won the single.5 messages unread, one marked urgent
- day 03
- enforced by
MAIL n unread, printed bynext,checkand the runner
in
AGENTS.md "the loop",coord.pystop the moment the flag is in
the grader client logged 54 refusals on challenges the team had already won. the refusal came at the probe, after the analysis that led to it.
54 calls refused on won challenges
- day 03
- enforced by the solved guard (the next shell command is refused)
in
AGENTS.md "the loop", stop rule,tools/comp/solved_guard.pya won single chains its own tiers
every tier went back through
nextto another agent: 21-82 s from single to time trial, 9-69 s on to mastery. metadata waited 78 s because a hand probe won its single and no solver existed.78 s for one lane's solver
- day 04
- enforced by autopilot (no model in the loop; it runs every tier on CORRECT)
in
AGENTS.md "the loop",tools/comp/autopilot.pythe commands are printed, never looked up
with no saved assignment, agents learned the grader client by hand: 19
--helplookups, and an unstarted single's status answered 409start_required.16-38 s from assignment to the artifact
- day 04
- enforced by
next(every race command's usage) andnewchal(the assignment, saved)
in
AGENTS.md "solving" 3,docs/comp/commands.mdone writer a file, and the lock lapses
20 edit-lock blocks came from one library-wide lock that also gated every timed runner. on day 8 a forgotten decode lock was still refusing fixes at 194 s.
20 lock blocks in one race
- day 04
- enforced by
coord.py edit(a lock per file or module; it lapses at 120 s, was 300)
in
AGENTS.md "solving" 6,docs/comp/timeouts.mdone runner owns the lanes
lane owners on the second runner were 2-5x slower on days 4 and 5. the first runner owns every non-web lane; the second comes in on a stall, and owns a task only once it sat unclaimed 3 minutes.
2-5x slower owners
- day 05
- enforced by
next(OWN_AFTER, 3 minutes unclaimed)
in
docs/comp/parents.md,coord.pyscope is the asked answer ids
one unknown property crashed the metadata single. the fixer then wrote 13 pdf properties and ran a full replay before it submitted the one field asked; a rival had the flag at +3 s.
40 s of the 63 s to the first answer
- day 06
- enforced by: the runbook only
in
AGENTS.md "solving" 3,docs/comp/lanes/metadata-forensics.mdread the answer syntax before the first probe
the script descriptor joined its steps with commas; the question's
answer_format.syntaxasked for>. the single drew 28 incorrect submissions.86 s from the first answer to the flag
- day 06
- enforced by
answerfmt(the syntax read clause by clause, negation first; a wording read two ways sends both forms)
in
AGENTS.md "solving" 3,src/ctfkit/hlib/answerfmt.pyevery zip-based member is a container
a jar was skipped as a plain zip, and one long analysis ran about a minute before the find was shared.
a minute on the binary single
- day 06
- enforced by
triageanddelta(every zip-based member opened; a format the lane never had printsNEW FAMILY)
in
docs/comp/lanes/new-family.md
days 8 to 14
the second week priced the waits: a hang nobody capped, a stall nobody heard, a value held and not sent.
every wait has a bound
an unseen mastery shape sent the warmups decoder into an unbounded search. nothing capped it; the claim went stale three times and each new holder re-ran the same hang in the background.
344 s of silent hang on the critical tier
- day 08
- enforced by the time caps (the deadline, else 15 s; a dry run 15 s a set; a watchdog kill 5 s past a cap)
in
docs/comp/timeouts.md,src/ctfkit/proc.pya stall wakes the parent
the crash was in intel at 09:05:50 and in the parent's mail at 09:08:54: a timed stall alerted only when someone ran
checkon that tier.184 s with no escalation
- day 08
- enforced by the parent's watcher (every claimed timed tier swept every 5 s; exit 6 wakes the parent)
in
docs/comp/parents.md "timed stalls",coord.pyan incident is the parent's; workers wait
regular timed starts drew gm 503s. finding the new
run_for_timefield took 4 minutes, and the script lane slept 3 more past the mail that had the fix.4 min + 3 min on the script lane, hidden behind the warmups hang
- day 08
- enforced by
GM INCIDENT(no timed start into it; everysolve.pyre-checks offline and resumes onincident resolve)
in
docs/comp/parents.md "gm incidents",gm.pypost the cause the moment you know it
the warmups mastery crashed on a line the library could not read. a 60 s fix waited behind two holders who never posted what they knew.
190 s of waiting on a 60 s fix
- day 09
- enforced by: the runbook only
in
AGENTS.md "when it fails" 2a parent routes; it never holds a tier
the busy parent held the crashed warmups mastery 90 s, found its cause at 09:00:33 and never posted it, and the stall alert went to that same parent.
90 s held by the parent
- day 09
- enforced by
next(a parent gets assists only) andsolve.py(it refuses a parent)
in
START.md 2,docs/comp/parents.mda failed tier keeps the free agents
nobody else could work the crashed tier, and five idle agents retired during the stall.
5 idle agents retired mid-stall
- day 09
- enforced by offline assists on the saved run, and
retire(exit 4 while a failed timed tier is open)
in
AGENTS.md "the loop", offline assist,coord.pya parent's wait is short once live
a stall wake sat unread behind a parent's long foreground wait.
2 min with a wake unread
- day 09
- enforced by
wait-live(a parent's call cut to 30 s once the day is live)
in
START.md 3,coord.py wait-live --max-wait 30a probe-won single chains past the owner's lock
a probe won the metadata single while the solver already held the graded value as a candidate, and the tiers waited out the owner's solver lock. the board finished 2nd, 16.9 s behind the fastest team; without the wait it led by 13.1 s.
30.1 s on the critical lane, and 1st place
- day 10
- enforced by autopilot's follow (it chains once
solve.pyreproduces the single, whatever lock the owner holds)
in
AGENTS.md "a value in hand" 7,tools/comp/autopilot.pythe backup kickoff detaches itself
a backup kickoff started in a foreground shell was gone 2 s later.
2 s the backup's whole life
- day 12
- enforced by
kickoff.py --backup(it detaches; the owner's t-5m check mails both parents when none runs)
in
START.md 2,tools/comp/kickoff.pya value in hand is submitted now, as stored
three agents read the right installer product code from the bytes by +44 s. each reshaped it, posted it, mailed it or hunted for another value; the stored form went out last, from the one agent that never decoded the file. the board finished 4th, 28.2 s behind the fastest team.
+44 s to +87 s from the value to the flag
- day 13
- enforced by
gm.py guess --values(every written form in one command, the stored form first) andintel --value(it submits before the note)
in
AGENTS.md "a value in hand" 1-3,gm.pynever relay a value instead of sending it
the parent had the stored form at +38 s, posted it with a wrong shape note and polled 32 s. the braced variant sat in a mail 9 s and the value in intel 21 s; the second runner sent one form per turn, about 5 s each.
32 s polling with the value in hand
- day 13
- enforced by
intel(on an open single it prints the reminder;--valuesends every form first)
in
AGENTS.md "a value in hand" 3, 6,coord.pythe first working decoder is patched in
four agents wrote four msi decoders. none was patched in, and the chain waited for the owner's own.
4 decoders written, none used
- day 13
- enforced by: the runbook only
in
AGENTS.md "cold start" 3,docs/comp/lanes/new-family.mdfollow the gm's next op
every reply on the spent warmups tier named
next_op: practice. the client hardcoded a retry, and the parent asked the human for an op the reply already named.3m28s until the op that won
- day 14
- enforced by
gm.py(it followsnext_op: practiceitself; every reply code is mapped and tested)
in
AGENTS.md "solving" 4,gm.pyone fixer a lane, one timed client
five agents wrote the same caesar +7 fix and three patches went unused; one claim was stolen mid-edit from under its holder's edit lock.
3 patches unused, about 50 s of hand-off
- day 14
- enforced by
AutopilotOwns(no timed start but autopilot's), a lane-boundnext, an edit lock that keeps its claim
in
AGENTS.md "the loop",coord.py
days 15 to 21
the third week moved the fix into the solver: a flag means an edit, a hand computation means a handler, and a field nothing names goes to the organizer.
the solver follows the flag, at once
the binary single flagged at +62.7 s. the fixer's edit lock came at +123.1 s, after an intel row and a 39 s turn with no tool call. the board lost 1st by 36.9 s.
60.4 s from the flag to the edit lock, and 1st place
- day 15
- enforced by
CHAIN WAITS ON solve.py(printed bygm.py,checkandnextuntil the tier runs; a save re-runs in under a second)
in
AGENTS.md "a value in hand" 7,tools/comp/autopilot.pyone edit at the anchor, no source reading
six fixer runs (the race and three replays): the change was 4-12 lines and still took 12-75 s from the flag. 17 s composing an intel row, 13-40 s reading library source, a 39 s planning turn, 15 s retyping a whole solve.py.
12-75 s for a 4-12 line fix
- day 15
- enforced by: the runbook only
in
AGENTS.md "fixing in seconds" 1, 3an ambiguous rule is a ranked list
in a gate replay the metadata fixer spent 22 s from the flag telling sorted order from relationship-id order; the binary fixer, handler first, had the single and all three tiers within 9 s of its save.
22 s before the edit, in a replay
- day 15
- enforced by: the runbook only
in
AGENTS.md "fixing in seconds" 7source first; escalate a field nothing names
202 of 204 fields were confirmed by +175 s. the next 95 minutes went to two storage labels and 1839 rejected guesses, until the organizer called it a challenge bug. the board finished 15th of 15.
5886 s on two labels
- day 17
- enforced by
CEILING(5 min on a single, nothing confirmed in 2.5, mails the parent; the parent drafts the organizer note)
in
AGENTS.md "when it fails" 7,docs/comp/parents.md "ceiling"a pivot changes the method, not the word list
the 25-guess cap stopped each label, and 14 pivots lifted it, each for more of the same kind of word list. a guess only varies a fact read from the source.
14 pivots, one method
- day 17
- enforced by
gm.py guess(25 rejected values a field, exit 4; a pivot note unlocks 25 more)
in
AGENTS.md "stall and pivot",gm.pytwo helpers a single, at most
fifteen agents worked one field, each re-deriving the same missing token.
15 agents on one field
- day 17
- enforced by
TEAM_ASSISTS(2:nexthands a third agent other work)
in
AGENTS.md "the loop", assist,coord.pythe hand computation is the handler
three fixers ran a throwaway
python3 -c, guessed its value, then retyped the same logic into solve.py while the tiers waited. days 19 and 20 again: handlers saved 15 s and 16 s from the flag.13.5-39 s from the guess to the solver rule
- day 17
- enforced by
gm.py guessandprobe(a fixer's guess on its own single exits 4 until handlers.py exists; day 20)
in
AGENTS.md "fixing in seconds" 0,gm.pya slow search hands the single over at 3 s
a slow dry run held the warmups single from its fixer until the dry run's 15 s cap; the fixer had it at +17.2 s.
13 s of a fixer with nothing to fix
- day 17
- enforced by autopilot's early hand-off (still searching at 3 s, it releases the single and keeps searching)
in
AGENTS.md "the loop",tools/comp/autopilot.pyone parent runs the live cycle
both parents ran the morning's live cycle and raced one attempt.
389 POSTs before 09:00
- day 18
- enforced by
preflight.py(one parent runs it; a second inside 20 minutes prints who ran it and its verdict)
in
START.md 1,tools/comp/preflight.pyafter a flag, the handler is the graded rule
the metadata fixer typed 40 lines with rotation and aspect-ratio branches nothing graded while autopilot held the tiers. the graded rule was one line.
22 s of a 45 s board finish
- day 18
- enforced by: the runbook only
in
AGENTS.md "fixing in seconds" 7parents route, they do not diagnose
the parent spent 40 s on a scratch unpacker for the new overlay while the binary fixer held the edit lock.
40 s of a parent on the fixer's work
- day 19
- enforced by: the runbook only
in
AGENTS.md "when it fails" 8the web matcher first
a web racer spent 62 s by hand on a time trial the deterministic matcher solves in under a second.
62 s on a known class
- day 20
- enforced by the web autopilot (it runs the matcher on every base before a racer gets it)
in
docs/comp/lanes/web.md
days 22 to 31
the last stretch turned rules into refusals: the scratch guard, the hand cap, the crash row and the wait cap each answer a loss from these days; day 31 priced the per-instance read, 57.5 s on the web lane.
the lost row names the set's bytes
one set of 108 lost the warmups mastery. the fixer ran seven scratch scripts on the saved run and took 3m44s to a save; no cause row was posted.
230 s on the tier, rank 3 to 42 on its board
- day 22
- enforced by the runner (a lost field's row starts with the set's bytes) and
replay(it prints what autopilot's gate checks)
in
AGENTS.md "when it fails" 1,src/ctfkit/hlib/runner.pya one-set loss re-rolls
107 of 108 fields were in by +16.7 s. the tier waited for a solver change, and the fresh attempt that followed the save flagged 1.5 s into it.
1 of 108 fields held the tier
- day 22
- enforced by autopilot (a failed attempt whose losses sit in one set of several is re-rolled at its deadline, at most twice a tier)
in
AGENTS.md "the loop" 2,tools/comp/autopilot.pya crash names every crashing field
a set crash hid the next cause, so the script time trial learned one cause per spent attempt.
388 s 3 causes, 5 attempts
- day 23
- enforced by the runner (
crash fields on <run>in intel within ~20 s;replayprints the same)
in
AGENTS.md "when it fails" 1,src/ctfkit/hlib/runner.pynever the scratchpad
the mobile fixer installed a tool mid-race, ran five scratch scripts and tested a draft in the scratchpad before copying it in; the script fixer ran 13 scratch scripts and read library source 6 times.
277 s of a 318 s window
- day 23
- enforced by the scratch guard (a fixer's
python3 -c, heredoc, scratchpad write or blocking install is refused while the board is open)
in
AGENTS.md "fixing in seconds" 3,tools/comp/solved_guard.pythe readiness check runs the night before
no night-before check had run, so the 08:49 rehearsal ran in the parent's foreground.
211 s of the parent's last minutes before release
- day 24
- enforced by the readiness stamp (preflight names it; a stale one is rehearsed detached by the first kickoff)
in
START.md 1,tools/comp/autopilot.py --checkthe gap row carries the handler draft
a fixer composed the same seven lines the
is yoursrow could have printed.10.7 s composing
- day 25
- enforced by autopilot's
is yoursrow (it ends with a handlers.py draft, a skeleton at worst)
in
AGENTS.md "fixing in seconds" 1,tools/comp/autopilot.pyone complete web edit
day 28's web race ran fail open friday and if a breach falls in a log: fail open friday's first edit switched to the policy route but dropped its
prefixandquery_parameter, and if a breach falls in a log took three edits to find the post-then-audit flow and missed the two-minute goal.205 s on three edits
- day 25
- enforced by: the runbook only
in
docs/comp/lanes/web.mda wait never outlives the shell
three of five fixers let
wait-livefall to the tool's 120 s default and read their task a turn late. day 29 found the harder limit: the shell backgrounds a call at 180 s whatever its timeout.25-35 s a task read a turn late
- day 27
- enforced by
coord.wait_cap(any--max-waitpast 170 s is cut; a test refuses a runbook line past it)
in
AGENTS.md "you",coord.py wait-live --max-wait 170the gm's status is the only arbiter of a completed tier
the socket-to-me tiers were completed by our own team's accounts outside the agent ledger, never another team's, so the local board read them open and the gm said complete: the single closed at +25:55.3, and the local stop mails at 09:26:45 and 09:27:03 put about a minute of work after it was solved.
about 1 min spent on a single our own account had already completed
- day 27
- enforced by the external-wins checklist: query the gm's status first
in
WEB.md "external wins (checklist)"a parent never hunts on a fixer's lane
the parent hand-probed about 25 binary fields while the binary fixer held the edit lock. the single won; its tiers then waited on
CHAIN WAITS ON solve.pywith no code reproducing the 25.10m54s of tiers waiting
- day 28
- enforced by
gm.hand_cap(one accepted field, then a non-fixer's hand probe is refused) andnext(no parent assist before the fixer's 90 s stall)
in
AGENTS.md "a value in hand" 8,gm.pyno lane without a fixer
no fixer existed until 09:07:46. the restarted parent oriented 3m44s before its first spawn, about a minute of it copying the prompt by hand.
7m46s with no fixer on any lane
- day 28
- enforced by headless fixers (kickoff launches one per open lane with none live) and
spawn --promptsas a restarted parent's first command
in
START.md 1,tools/comp/fixer_launch.pyfirst read, small saves
the mobile fixer had the
ANS-…answer readable in the quoted puzzle json and sent nothing; the binary fixer composed 352 s in silence and left a patch for its own posted cause unapplied 3m18s.5m53s with the answer readable
- day 28
- enforced by: the runbook only
in
AGENTS.md "fixing in seconds" 8a single-set run sends its clean fields
the binary time trial answered 391 of 392 fields and sent nothing: one crashing field sank the set, and the attempt expired at 87 s.
391 of 392 answered, none sent
- day 28
- enforced by the runner (crashing fields are halved out; the clean ones bank on the attempt while the fix lands)
in
HUNTRESS.md "premise and modes",src/ctfkit/hlib/runner.pya handler reads the member's real format
mastery alternated elf and pe builds of one tag: an elf-only handler crashed a ranked time-trial attempt and 7 of 12 mastery sets on a pe.
7 of 12 mastery sets, and one ranked attempt
- day 28
- enforced by the field registry (a handler that raises hands its field to the lane reader; day 29)
in
docs/comp/lanes/binary-reverse-engineering.md,src/ctfkit/hlib/fields.pya handler guards its dispatch
a handler read
puzzle["data"]on the time trial's v2 puzzle, which carries none; the set crashed though the lane solver had the value.36 s and a ranked mastery set
- day 29
- enforced by the field registry (a raising handler's field goes to the lane reader)
in
AGENTS.md "fixing in seconds" 1,src/ctfkit/hlib/fields.pyone entry per rule, duplicates kept
a handler deduplicated its ranked candidates. on fresh bytes the graded position moved, and the time trial led with the by-digest order.
1 rejection inside the attempt
- day 30
- enforced by: the runbook only
in
AGENTS.md "fixing in seconds" 7,docs/comp/lanes/metadata-forensics.mda web racer's wait stays inside its shell
the web runner's shell backgrounds a call still running at about 60 s, and a yielded turn never woke on its follow-up.
3 racers frozen on the release morning until killed
- day 30
- enforced by: the runbook only
in
docs/comp/lanes/web.mdthe shell is the human's shell
the agent's shell is a zsh login shell with ugrep and eza aliases: a recursive grep skips gitignored logs,
lsprints colour codes, and an unmatched glob aborts the whole line.4 retried turns
- day 30
- enforced by: the runbook only
in
AGENTS.md "watching"a per-instance value is read per instance
the solver carried the single instance's claim-header name to the fresh timed target, where the name had rotated; two no-proof capture runs came back inside one ranked attempt (
17d313d3, 15 ranked attempts remaining throughout), and the third derived the name from the current target's own 403 vary.57.5 s from the attempt's start to the end of the second no-proof capture, and the attempt stayed ranked throughout
- day 31
- enforced by the solver reads every per-instance value from the current target's own responses, never from an earlier assignment
in
challenges/day30/web-exploitation-day-30-framework-roulette/solve.py,docs/comp/lanes/web.md
what the rules bought
web ran its own machinery and is left out. the scale stops at 20:00; a dashed bar is a day beyond it (days 1, 17, 28; day 1 opened 2h28m late). the reviews of days 3, 8, 9, 14, 16, 17, 20, 23, 28, 29 wrote a mechanism (hover a bar for which); each ran from the next release.
- solved-guard refusals
- scratch-guard refusals
- singles whose tiers waited on a solver edit
a chain wait is a won single whose tiers printed CHAIN WAITS ON solve.py. the session logs start on day 4, so days 1 to 3 read 0 for every series.
on the on-time days 3 to 10 the non-web lanes finished in a median 4:31; days 11 to 21 in 1:06; days 22 to 31 in 3:32. they finished inside ten minutes on 5 of 8 days early, 10 of 11 in the middle and 9 of 10 at the end. failed timed runs did not fall with them: 18 of 98 in days 3 to 10, 3 of 112 in days 11 to 21, 23 of 123 in days 22 to 31, 15 of those on days 23 and 28, the days four new formats and three lanes' new fields arrived with no reader. the refusal counts rise late because the guards came late: a refusal is a rule the machine caught, not a mistake that never happened. the challenges changed under the rules, so this is a correlation, not proof: what the rules target is the dead time around a failure, not the failure.
the ones the machine enforces
since day names the race whose review wrote the mechanism; it ran from the next release. "by N" marks one whose origin no review dates, first noted on day N.
| mechanism | what it refuses or caps | since day | rule it enforces |
|---|---|---|---|
one runner per timed tier (gm.py start) |
a second start on a held timed tier (exit 4, busy) |
3 | one runner per timed tier |
| the timed preflight (the runner) | the clock, while the solver disagrees with the graded single | 3 | no clock before the solver reproduces the single |
one guesser a field (gm.py guess --tokens) |
a second token guesser on a field (exit 4) | 3 | one guesser per label field |
the edit lock (coord.py edit) |
a second writer on a solver, a file or a library module (exit 4); the lock lapses at 120 s | by 4 | one writer a file |
the guess cap (gm.py guess) |
a 26th rejected value on one field (exit 4); a pivot note unlocks 25 more | by 8 | a pivot changes the method |
the 90 s stall and 300 s ceiling (coord.py check) |
a single quiet past 90 s prints STALL; past 300 s CEILING mails the parent |
by 8 | source first; escalate a field nothing names |
the time caps (runner, proc) |
a solve past its deadline (else 15 s), a dry run past 15 s a set, a process 5 s past its cap (the sidecar watchdog kills it) | 8 | every wait has a bound |
the gm incident (coord.py incident) |
every timed start into an open incident; workers wait and resume on resolve | 8 | an incident is the parent's |
the solved guard (tools/comp/solved_guard.py) |
the next shell command of an agent whose task another agent won, while the board is open; coordination commands pass | by 9 | stop the moment the flag is in |
parent routing (coord.py next, solve.py) |
a task for a parent to own; solve.py refuses a parent |
9 | a parent routes; it never holds a tier |
the retire refusal (coord.py retire) |
a helper's retire while a failed timed tier is open, a fixer's while its lane is open (exit 4) | 9 | a failed tier keeps the free agents |
the team 429 hold (gm.py) |
a ranked start into the gm's 429 minute waits for the next minute; every process now shares the hold, and a start or probe resends once | 9 | never retry by hand |
one timed client (AutopilotOwns) |
anyone's timed start or retry on a lane autopilot owns; solve.py <id>/<tier> exits 4 |
14 | one fixer a lane, one timed client |
the parent's wait (coord.py wait-live) |
a parent's call past 30 s once the day is live | 16 | a parent's wait is short once live |
the 3 s single hand-off (tools/comp/autopilot.py) |
autopilot's hold on a single still searching at 3 s: the single goes to its fixer, the search goes on | 17 | a slow search hands the single over at 3 s |
two helpers a single (TEAM_ASSISTS) |
a third helper on one single, every runner counted | 17 | two helpers a single, at most |
the handler-required probe refusal (gm.py guess, probe) |
a fixer's hand guess on its own lane's single until handlers.py exists (exit 4); --hand is logged and mailed |
20 | the hand computation is the handler |
the scratch guard (tools/comp/solved_guard.py) |
a fixer's or helper's python3 -c, heredoc, scratchpad or /tmp write and blocking install while the board is open; parents exempt |
23 | never the scratchpad |
the hand cap (gm.hand_cap) |
a non-fixer's hand probe on a fixer's lane past one accepted field | 28 | a parent never hunts on a fixer's lane |
the 170 s wait cap (coord.wait_cap) |
any --max-wait past 170 s, under the shell's 180 s background limit (exit 5 at the cap) |
29 | a wait never outlives the shell |
what still runs on trust
these rules have no guard: only the runbook holds them. the day in brackets is the race that wrote each; a later failure is named where the record has one.
- post the cause before the fix (day 9). last failed on day 22: the warmups fixer took 3m44s to a save with no cause row, and the parent's cause row came once the human asked. a crashed run now posts its own
crash fieldsrow; a rejected one still waits on an agent. - one edit at the anchor, no library source reading (day 15). last failed on day 28: two fixers made about 26 reads of library source.
- a save every field, never a silent compose (day 28, a 352 s compose). the watcher's 600 s window for a fixer holding its edit lock only keeps it from being respawned; nothing makes it save.
- parents route, they do not diagnose (day 19). last failed on day 28: about 25 fields probed by hand. the hand cap now stops a parent's probes after one accepted field, but the scratch guard exempts parents and nothing stops a parent reading bytes.
- one complete web edit, from the fingerprint alone (day 25, 205 s on three edits). last failed on day 30: ~10 s of detour and ~25 s of silent compose on a 45 s miss.
- one entry per rule, duplicates kept (day 30, one rejection). the runner carries the graded position to the timed tiers; a handler that dedupes its list still moves it.
{generated by a clanker}