the human's calls: the decisions that steered the month

one human ran the team. the agents raced, fixed and wrote; the human decided what to optimise, what was allowed, who ran what and when to stop. this page lists those calls in date order, each with the situation that forced it and what changed after it. the sources are the decision sections of the day reports, the runbook's dated rules and the parent's notes of the human's decisions; a call none of them states is not here.

  • 77calls logged2026-10-03 to 2026-10-31, one card each
  • 13standing orders in force at the endthe yardstick and the always-on rules
  • 3 to 31days of the earliest and the latest callthe gm direct; first to every lane solved
  • 40% to 97%largest measured effect: autopilot the sole timed client (day 14)timed tiers autopilot sent, days 1-13 against 15-31 (90% on days 10-13)
  • 3dated citations of the human in AGENTS.md32 dated citations across the runbook
  • day 14the day with the most calls7 calls on 2026-10-14

the standing orders

13 of the calls stayed in force from the day they were given to the last flag. the yardstick comes first: every after-action measured itself by it. the rest are the always-on rules.

the timeline

one card per call, in date order and grouped by week. the card's colour is its area (race green, tooling cyan, agents purple, web pink, publishing orange, infrastructure yellow); the date badge is green for a call that changed what submits flags or how agents work every day, cyan for one mechanism, grey for housekeeping.

week 1: days 1 to 7, 5 calls

the first week set the terms: a direct line to the gm, a race counted in seconds, and automation that needs no agent turn.

  • the gm direct with the paired token, the homebase vm off

    a raw gm request sent through the homebase vm spent the human's course credits: the vm bills compute and tokens for as long as it runs.

    after every gm call goes direct; gm.py fails over to the vm only when direct is down, laggy or intermittent, and the vm stays off unless the human turns it on.

    • 2026-10-03 · day 3
    • area: infrastructure

    source: the parent's notes

  • a first-solve race: poll in a second, assist after 10 s

    the team finished day 3 second by 3 points, and 10 s retry polling was too slow. a single one agent held waited 120 s before anyone could help.

    after kickoff and wait-live were built that day; agents poll local state at about 1 s, and the assist delay went from 120 s to 33 s to 10 s, with the order never to raise it.

    • 2026-10-03 · day 3
    • area: race

    source: the parent's notes

  • 'insanely fast' automation; expect a change every day

    on day 4 a rival team had all three tiers of a lane by +59 s, while each agent turn cost 20 to 80 s.

    after kickoff started autopilot, which runs the prior solver on the new bytes, submits and chains the tiers with no agent turn: it sent 0 of 61 flags on days 1 to 4 and 80 of 135 on days 5 to 13.

    • 2026-10-04 · day 4
    • area: tooling

    source: the parent's notes

  • the human pushes

    the repo's ssh key asks for a passphrase on every push, which no agent can type.

    after agents commit when asked and say when a push is due; the end-of-day runbook lists pushing among the things never done alone.

    • 2026-10-04 · day 4
    • area: infrastructure

    source: the parent's notes; AFTER_ACTION.md, the standing order

  • equal points and a faster finish is first, never 'tied #1'

    a status reply called the team 'tied #1 overall' because another team had the same points. the board ranks on total score, then on the fastest completion.

    after retros, writeups and status lines report the rank the scoreboard implies.

    • 2026-10-06 · day 6
    • area: publishing

    source: the parent's notes

week 2: days 8 to 14, 18 calls

the second week named the yardstick and the standing order, and handed every timed tier to autopilot.

  • the yardstick: the first correct flag on every tier, not the fastest attempt

    day 8 swift justice mastery: the fastest attempt on the board took 2.65 s and ours 7.29 s, but our flag landed at +340 s against +146 s for the best other team. 194 s went before the attempt that won started.

    after every after-action ranks a finding by the seconds it moved the board finish, race logic and tooling first, solvers second.

    • 2026-10-08 · day 8
    • area: race

    source: AFTER_ACTION.md "what we optimize"; the parent's notes

  • ranked starts first: 'go with what we know works'

    the untimed-first design of day 9 had never raced; days 1 to 10 raced on ranked starts. untimed play gets the flag but no hall-of-fame time.

    after a timed start asks for the ranked pool and plays untimed only when the gm refuses it: 213 ranked tiers over the month, 35 unranked.

    • 2026-10-10 · day 10
    • area: race

    source: the parent's notes

  • cancel the backup kickoff to head off the 429s

    at 08:48 et the human told the backup parent to cancel its backup kickoff and let one kickoff run, to try and prevent the release-minute 429s.

    after the review found the backup had never started and sends no gm call before it takes over; it was restored as the runbook describes, on the human's call, and now prints that it is armed with no gm calls.

    • 2026-10-11 · day 11
    • area: race

    source: day 11 report, f3 and dd11-4

  • no second solve before a chained tier

    the runner re-solved a tier the chain had just won in memory before starting the next one; only a solver that changes its answer on the same bytes could fail it.

    after the human approved removing it: a correct run remembers its code stamp and the next start skips the re-solve. day 11 binary mastery's check went from 0.905 s to 0.001 s.

    • 2026-10-11 · day 11
    • area: tooling

    source: day 11 report, dd11-1

  • a field one lane cannot read goes to every lane's reader of that name

    a new metadata field named a managed-pe table column the binary lane already parsed. nothing on the metadata path pointed to it, three agents rebuilt the decode by hand, and the lane lost 38.3 s.

    after the human widened the fix from one property to every lane: crosslane asks every lane's reader of exactly the field's name, and the dry run prints which lane read it.

    • 2026-10-12 · day 12
    • area: tooling

    source: day 12 report, dd12-1

  • hold out the 429 window; never pace a lane that is ahead

    the metadata single's value was in hand at +40.3 s, but the first gm contact at +45.2 s fell inside the team's release-minute 429 window and the flag waited to 09:01:03.

    after a team-wide hold after the retry ladder: no process sends into a known window, and fast lanes keep their pace.

    • 2026-10-12 · day 12
    • area: race

    source: day 12 report, f2

  • a 429 on a timed download never expires the tier

    another team lost a tier to a 429 on its timed download.

    after a download refetches the same attempt when the clock covers the window, else opens a fresh attempt.

    • 2026-10-12 · day 12
    • area: race

    source: day 12 report, dd12-8

  • ignore the empty ranked pool

    from day 12 to day 14 the gm's ranked pool was empty, a gm-side bug, so every timed tier of those days played untimed.

    after nothing was built around it; ranked starts came back on day 15, and a later empty pool is a note to the organizer (day 18).

    • 2026-10-12 · day 12
    • area: infrastructure

    source: the parent's notes

  • 'i want everything fixed, i want to get 1st place'

    the end of day had been chained by hand: read the after action, pull variants, check hints, look for waste, then the blog.

    after 'read AFTER_ACTION.md' runs every stage without being asked for the next; the reviewer fixes whatever makes a flag faster or more accurate and asks only on a tradeoff, untested gm behaviour, attempts past the budget, deleting data or pushing.

    • 2026-10-13 · day 13
    • area: agents

    source: AFTER_ACTION.md, the standing order; the parent's notes

  • a value held is a value submitted: agent logic first, every lane

    on day 13 the product code was readable at +31 s and in intel at +41 s; the flag came at +87 s after five rejected shapes, while three agents reshaped, mailed or posted the value. the human: an agent problem, not a code problem.

    after AGENTS.md 'a value in hand': the next command after reading a value submits it as stored with every variant, and the evidence pack flags a value held 15 s unsubmitted (day 13's sat 54.9 s).

    • 2026-10-13 · day 13
    • area: agents

    source: day 13 report, phase 2; AGENTS.md "a value in hand"

  • spend unique attempts on variants, keep a third of each pool

    the pools were refilled (233 to 254 left on days 1 to 11) and the variants carried what the next days brought. the rule of day 11 was to spend at most a third.

    after the harvest keeps 0.34 of each tier's original pool; on day 17 the human allowed pulls past it for a day already solved and rerun.

    • 2026-10-13 · day 13
    • area: infrastructure

    source: the parent's notes

  • autopilot is the sole timed client; one fixer per lane

    every losing day (9, 10, 13, 14) was one lane falling to the agent path, 50 to 270 s lost in hand-offs, claims, steals and holds; lanes autopilot ran finished in 9 to 14 s. day 14: +276.7 s against the fastest team's +13.6 s. the human: 'too many cooks', and 'i want to win this competition'.

    after agents write fixes and autopilot runs every timed tier: it sent 97% of the timed tiers on days 15 to 31, against 40% on days 1 to 13 (90% on days 10 to 13).

    • 2026-10-14 · day 14
    • area: race

    source: day 14 report, decisions; the parent's notes

  • each autopilot slot lists only its own lane

    the release list names every lane; on day 14 it answered in 1.76 s on a gm the organizer confirmed was slow that morning.

    after every slot sends its own lane list from t+0.01 s: on day 16 a lane list came back at 2.28 s against +4.08 s for the full one.

    • 2026-10-14 · day 14
    • area: race

    source: day 14 report, decisions; tips page

  • the practice flag counts

    day 14 lost 79 s sending retries into 409s while every gm reply advertised practice, and the team was unsure a practice flag would score.

    after the human confirmed it on the scoreboard (day 14 warmups mastery, 277 s, rank 24); gm.py follows the gm's next op itself, so a spent tier is never blocked.

    • 2026-10-14 · day 14
    • area: race

    source: day 14 report, decisions; the parent's notes

  • keep submitting low-confidence answers

    the runner submitted 109 times on low confidence in one attempt between 09:00:06 and 09:00:35.

    after kept: a low-confidence answer is still submitted; by the last day it went in its own group, so it never sank the confident fields.

    • 2026-10-14 · day 14
    • area: race

    source: day 14 report, decisions

  • tiers run in order; no speculative next-tier start

    sessions kept proposing a time trial start beside the single's submit; on day 13 the human stated the gm needs the single complete first.

    after settled: single, then time trial, then mastery; speed comes from the hand-off after a flag (19 to 64 ms in a day 13 sim), not from overlapping tiers.

    • 2026-10-14 · day 14
    • area: race

    source: day 14 report, decisions; the parent's notes

  • be predictive and proactive

    the two tiers that lost day 14 (a new line codec, a new lnk family whose reader already existed) could be read off each lane's daily pattern; 13 of 15 tiers were autopilot's by +13 s. the human: a failure in our tooling, not only in agent logic.

    after reviews weigh tooling beside agent logic and prepare the next day: a synthetic next day per lane, day patterns, readers mapped to a new family's fields.

    • 2026-10-14 · day 14
    • area: tooling

    source: the parent's notes

  • large static data to the network share

    the data disk sat at 93 to 98%, and a full disk broke the test suite mid-session.

    after cold files move to the share behind symlinks; on day 17 the human narrowed it to big artifacts, session logs and scoreboards with all intel kept local, and on day 19 sent every fixture artifact there, untracked in git.

    • 2026-10-14 · day 14
    • area: infrastructure

    source: the parent's notes

week 3: days 15 to 21, 26 calls

the third week set what agents may do: source before guessing, the handler before the guess, and the web lane walled off.

  • a replay with real fixers gates every agent-logic change

    day 15 placed fourth by 36.9 s; binary's fixer edited its solver 60 s after its own probe won the single. scripted sims never showed that cause; a real fixer on the day's bytes reproduced the race within 8 s (+143.1 s against +151.2 s).

    after stage 5.3 of the end of day: an agent-logic change is done when a replay shows the gain. the day 15 replays finished at +118.7 s with the agent rules alone and +43.5 s with the role routes added.

    • 2026-10-15 · day 15
    • area: agents

    source: day 15 report, d4; AFTER_ACTION.md stage 5.3

  • report the empty binary lane list to the organizer

    the gm's lane list for binary came back empty on every day it was asked.

    after the human reported it; the slot keeps listing binary by its id prefix.

    • 2026-10-15 · day 15
    • area: infrastructure

    source: day 15 report, d1

  • ooxml relationship targets: ascending order leads

    day 15 graded a sorted relationship target list; the next ooxml day could want document order.

    after the sorted form leads on the next ooxml day.

    • 2026-10-15 · day 15
    • area: tooling

    source: day 15 report, d2

  • no attempts spent measuring the gm clock

    measuring the gm clock offset with untimed starts would spend about five regular attempts.

    after skipped: no regular attempt was spent on it.

    • 2026-10-15 · day 15
    • area: race

    source: day 15 report, d3

  • an ungraded c field answers with a ranked guess

    the c function_count field had never been graded, and the solver crashed on it.

    after kept as an unsure guess (symbols first, then the unwind table) with a fallback.

    • 2026-10-15 · day 15
    • area: tooling

    source: day 15 report, d6

  • start the predictable warmups single at t+0, beside the list

    day 16's release list answered at +4.0 s, and the warmups single's id is known before the release; a start before the list was untested gm behaviour.

    after its slot sends the start at t+0.01 s and falls back to the listed id when they differ: the warmups single landed in a median 1.9 s on days 17 to 31, against 2.4 s on days 5 to 16.

    • 2026-10-16 · day 16
    • area: race

    source: day 16 report, d1

  • no replay for a printed-page change; make sure the cut never recurs

    dd16-4 changed what a tool prints to fixers, and a page had been cut by the agent harness's output cap.

    after the shell output cap went to 64000 characters, pages stay under 8000, and a test keeps the cap at four pages.

    • 2026-10-16 · day 16
    • area: agents

    source: day 16 report, d2

  • the docs agents read match the code every night

    the human named the documents day 16's changes had not reached: the capabilities doc, HUNTRESS.md, AGENTS.md, START.md, parents.md and four lane briefs.

    after stage 7 step 1b: for every commit since the morning, each doc an agent reads describes it as it now works.

    • 2026-10-16 · day 16
    • area: agents

    source: day 16 report, stage 7b; AFTER_ACTION.md stage 7

  • source first; a guess only varies a fact read from the bytes

    day 17 binary: two storage labels took about 95 minutes and 1889 rejected values; the organizer confirmed the answers should have been readable from the source, a challenge bug.

    after the 25-a-field cap stays, no automated label sweeps, and a field nothing can derive goes to the human as an organizer note: 102 rejected values on days 18 to 31 together, at most 30 in a day.

    • 2026-10-17 · day 17
    • area: race

    source: day 17 report, d1; AGENTS.md "when it fails" 7

  • tools that make a flag faster are pre-approved

    a disassembler, a decompiler, a parser or a library that would make a flag faster or more accurate.

    after installed without asking through the project's own manifests, ready offline on race morning, and used by the item that wanted it in the same commit.

    • 2026-10-17 · day 17
    • area: infrastructure

    source: AFTER_ACTION.md "always"

  • the human runs the live steps

    the testing grounds retest a past lane against the real gm with a fresh agent, and auto mode blocks a gm submit or a worker launch from the session.

    after the human runs open and launch by hand: one regular attempt per timed tier, ranked untouched, and every run leaves the same record.

    • 2026-10-17 · day 17
    • area: infrastructure

    source: the parent's notes; AFTER_ACTION.md "the testing grounds"

  • comp/ for the race, after-action/ for the rest

    race agents and reviewers share one checkout.

    after an active race uses only docs/comp, tools/comp and logs/comp, and AGENTS.md bars race agents from every after-action folder.

    • 2026-10-17 · day 17
    • area: infrastructure

    source: the parent's notes; AGENTS.md agent boundary

  • the parent runs the morning's live gm cycle; the backup only backs up

    on day 18 both parents ran preflight's live cycle at 08:47 on the same regular attempt; the loser bisected 383 already-solved submits (389 posts) ten minutes before the release, a team 429 risk.

    after one live cycle per morning, team-wide; the backup parent sends nothing unless the main runner is low or out.

    • 2026-10-18 · day 18
    • area: infrastructure

    source: the parent's notes; parents.md

  • a stopped agent is marked stopped and respawned a rung down

    a fixer cut short by a classifier stop or an out-of-context exit kept its lane for its 300 s working window.

    after coord.py stopped frees the name at once; the respawn carries a resume note, a rung down the model ladder after a classifier stop.

    • 2026-10-18 · day 18
    • area: agents

    source: parents.md "scaling the team"

  • tell the organizer about the four unranked tiers

    warmups and binary time trial and mastery played untimed: the release rows had no ranked flag, and the first start said no attempts were left.

    after the human sent the note with the evidence; the flags were first on the board either way, and a new unranked tier is now checked against the first start's count.

    • 2026-10-18 · day 18
    • area: race

    source: day 18 report, d1

  • every tool's output by its words, not a cipher word list

    day 18's only gap was a known tool's output asked by other words (video dimensions as ffprobe's width x height): 31.8 s of a 45.5 s finish.

    after readers that parse a tool's output expose every key by its words, and a reader's miss lists the keys it has.

    • 2026-10-18 · day 18
    • area: tooling

    source: day 18 report, d2; the parent's notes

  • web branch work stays off main

    audit work had been done on the web lane's branch. the human: not on the web branch, just on main, and leave the web branch alone.

    after the non-web audit items were rewritten fresh on main; on day 26 a rebuilt web lane merged into main and the old branch was archived unmerged.

    • 2026-10-18 · day 18
    • area: web

    source: the parent's notes

  • one regular attempt to test gzip on the live gm

    a 3.8 mb script mastery reply crossed the wire whole on the clock.

    after the gm's edge gzips json on request: a 523 kb start crossed in 29 kb, and gm.py now asks on every call but a download.

    • 2026-10-18 · day 18
    • area: tooling

    source: the parent's notes

  • every agent's context says this is a sanctioned ctf

    a model safeguard stopped the parent on exploitation-named records; the workspace is full of malware-shaped words that are challenge content.

    after the authorized-context block opens AGENTS.md, START.md and every fixer prompt; on day 28 the attack-word lists left the shared docs and a fresh parent read START.md with no stop.

    • 2026-10-19 · day 19
    • area: agents

    source: the parent's notes; AGENTS.md "authorized context"

  • a checklist for every after-action

    the end of day runs eight stages, and its items belong to the non-web reviewer, the web runner or the human.

    after the first act copies the template; each item is ticked with its evidence in the same commit, and the after-action is done when every item is ticked.

    • 2026-10-19 · day 19
    • area: agents

    source: AFTER_ACTION.md "always"

  • remove the parent-spawn race replay

    a whole race day replayed with real parents spawning fixers was built that afternoon; both live attempts failed (a safeguard stop while orienting, a sandbox that blocked the network).

    after reverted: 'they arent working and need to be better thought out'. a design comes first; the single-lane testing grounds stay.

    • 2026-10-19 · day 19
    • area: tooling

    source: the parent's notes

  • web is the web runner's, and the web runner works only web

    the web runner's uncommitted edits to AGENTS.md, coord.py and gm.py sat in the shared checkout beside the non-web after-action, and mixing the two clobbered each other's work.

    after the parent and fixers never claim, probe or read web; on day 20 the human added web workers on a model of an earlier generation, and a web rewrite swept into a non-web commit by a plain git add made commits go by hunk.

    • 2026-10-19 · day 19
    • area: web

    source: the parent's notes

  • the handler before the guess, enforced

    fixers wrote the handler after the flag (day 19: 15 s, day 20 metadata: 16 s) while the tiers waited on solve.py. the human: 'this is imperative'.

    after a fixer's guess on its own single is refused until handlers.py exists, --hand only for a value no code read, and a hand win still chains from a learned header rule.

    • 2026-10-20 · day 20
    • area: agents

    source: day 20 report, d4; AGENTS.md "fixing in seconds" 0

  • no start ahead of the release list

    the speed review offered to start the predicted warmups single a few hundred ms before the release list.

    after declined: the start waits for the list. the human: 'we will win by not having to hand it to fixers no matter the lane'.

    • 2026-10-20 · day 20
    • area: race

    source: day 20 report, d5; the parent's notes

  • no next-tier start before a correct

    the speed review offered to fire the next tier's start right after the submit post, before the correct came back (0.7 to 0.9 s a tier).

    after declined: speed work went to solver coverage of the next day's change and the chain after a flag.

    • 2026-10-20 · day 20
    • area: race

    source: day 20 report, d6; the parent's notes

  • skip the live-agent replay

    with day 21's fixes no fixer gets a task, so no agent-logic change needed the gate; a replay costs two runs and about 15 minutes.

    after 'we dont need to run the live-agent replay': the gains are tooling, measured by the traced sim, and the printed-line changes got unit tests.

    • 2026-10-21 · day 21
    • area: agents

    source: day 21 report, d1

week 4: days 22 to 28, 18 calls

the fourth week closed the gaps a fixer leaves (re-rolls, no scratch work, headless fixers, partial submits) while the web lane was rebuilt.

  • the parent moves to the current main model; every fixer pinned to it

    the parent had run on another model since day 17, and a spawned agent inherits the parent's model unless the spawn names one.

    after the start script pins the parent, and every fixer and helper spawn names the main model; the web workers keep theirs.

    • 2026-10-22 · day 22
    • area: agents

    source: START.md "models"; parents.md; the parent's notes

  • re-roll a timed attempt that lost a few fields

    day 22 warmups mastery lost 1 of 108 fields to a slug past its stated shape; every other field was accepted, and the tier waited for a clean re-solve.

    after autopilot opens one fresh attempt at the deadline (+2 s) when the losses sit in one set of several, at most twice a tier; by the report's reckoning day 22 would have ended near +138 s instead of +239.6 s.

    • 2026-10-22 · day 22
    • area: race

    source: day 22 report, d1

  • no scratchpad in a race; a mid-race install runs detached

    day 23: the mobile fixer installed a package mid-race and drafted a 539-line handler in the scratchpad, 50 s after the value was known; the script fixer ran 13 scratch scripts. the human had said it several times.

    after a guard refuses scratch scripts and blocking installs while a board is open (26 refusals on days 23 to 31), and install.py installs detached and mails when ready.

    • 2026-10-23 · day 23
    • area: agents

    source: day 23 report, d1; AGENTS.md "fixing in seconds" 3

  • crash isolation never holds autopilot

    the per-field crash isolation blocked autopilot's follow loop for up to 20 s after a crash, and a save inside that window waited.

    after the isolation runs in a detached replay child under its own cap.

    • 2026-10-23 · day 23
    • area: tooling

    source: day 23 report, d2

  • become the fastest web solvers: six flags in 71 s

    first place cleared the day's six web flags (two roots, single to mastery) in 1:11.043; the team's web lane took 6:15 on day 25.

    after a matcher-first web solver, an offline real-worker test and a new launcher went in for day 26, when the web lane took 19:22.

    • 2026-10-25 · day 25
    • area: web

    source: the parent's notes

  • reuse the testing grounds; one command, no manual steps

    a separate real-worker harness rebuilt what the testing grounds already did, and a live test took four manual steps with pasted run dirs.

    after the harness was cut to a deterministic smoke test, and one offline command now runs a real web worker against local targets.

    • 2026-10-25 · day 25
    • area: web

    source: the parent's notes

  • a reviewer model advises; the main model writes the hard code

    the web-solver build: a reviewer catches design flaws cheaply, a stronger implementer writes the hard parts.

    after a reviewer model checks the design first, and a worker on the main model writes the complicated parts; simple edits stay inline.

    • 2026-10-25 · day 25
    • area: web

    source: the parent's notes

  • a non-web after-action during the web race works in a sibling worktree

    on days 25 and 26 the human asked for the non-web after-action while web still raced; an hlib lock in the shared checkout refuses web's timed starts, and uv run in a worktree re-pointed main's install under the live race.

    after a standing rule: fixes on a branch in a sibling worktree, tested without uv run, merged once the board shows web complete.

    • 2026-10-26 · day 26
    • area: infrastructure

    source: day 26 report, d1

  • a hand-off draft needs two shared words

    a one-word handler draft cost 53 s of composing on day 26 and had never been right.

    after kept: two shared words or no draft, and the small text members go to the fixer verbatim.

    • 2026-10-26 · day 26
    • area: agents

    source: day 26 report, d2

  • the new sweeps may spend single guesses

    the day 26 sweeps assemble up to 8 candidates or try two or three enum values on the single, inside the 25-guess cap.

    after kept: a wrong probe costs about 1 s, the hand-off it replaces 30 to 80 s.

    • 2026-10-26 · day 26
    • area: race

    source: day 26 report, d3

  • cut the race-day reading

    fixers read 15 to 18 thousand words before the release.

    after the cut: race-day rules in AGENTS.md, history in the capabilities doc and reports, measured by words read before release; on day 28 AGENTS.md was still 848 lines.

    • 2026-10-26 · day 26
    • area: agents

    source: day 26 report, d4; day 28 report

  • rebuild the web lane from scratch, one visible web parent

    day 26: the background web fixers ran under an isolation that blocked every edit of the shared solver; they found their causes but could not save a fix, and the human could not see them work.

    after the human ordered a rebuild from scratch, designed by a reviewer model and carried out exactly ('do not make decisions for yourself'); it merged into main the same day, and the old web branch was archived unmerged.

    • 2026-10-26 · day 26
    • area: web

    source: the parent's notes

  • unrestricted protocol inside the assigned target

    the day-27 websocket class needed connection upgrades and compressed frames that the collector's request caps forbade, and a slow protocol class was a likely late-event shape.

    after the human authorized arbitrary methods, headers, body sizes and upgrades inside the team's assigned private targets: the collector keeps one connection pool per origin, never attaches the gm token to a target request, and every new transport ships an offline fixture before it races.

    • 2026-10-27 · day 27
    • area: web

    source: the web runbook, live target rules

  • fixers launch headless when the parent cannot

    day 28: the parent was stopped on its first turn; kickoff survived, but nothing spawned fixers, and three lanes had none for 7m46s.

    after kickoff launches a detached fixer (permissions bypassed) for every open lane with no live fixer, and a t-5m sentinel alerts the human when the parent has made no call.

    • 2026-10-28 · day 28
    • area: agents

    source: parents.md "headless backup"; the parent's notes

  • the parent's fallback after a classifier stop

    on day 28, with the parent stopped, the human went through three models by hand while the parent's machinery was down for 7m46s.

    after a fallback model of an earlier generation is the named fallback in START.md and parents.md; a normal day still starts on the main model.

    • 2026-10-28 · day 28
    • area: agents

    source: START.md; parents.md "scaling the team"

  • a timed attempt banks its answered fields when one crashes

    day 28 binary time trial solved 391 of 392 fields and sent nothing because one field crashed; the attempt expired at 87 s and a ranked attempt was spent. the same day's mastery sent 5 of 12 sets and stayed open until the fix landed.

    after the runner submits the answered fields and leaves the crashers open to the deadline, single-set and multi-set tiers alike.

    • 2026-10-28 · day 28
    • area: race

    source: HUNTRESS.md "ranked"; the parent's notes

  • patterns and predictions are fair; another team's flags never

    another team told the human it had a patch that saved 'so much time it's almost unfair', and the human asked whether predicting templated answers, ids and variants was fair game.

    after the organizer: detecting patterns, predictions, etc is all permitted. AGENTS.md carries it beside the rules that never moved.

    • 2026-10-28 · day 28
    • area: race

    source: AGENTS.md "rules (hard)"; the parent's notes

  • the web runbook carries the web text; the shared docs one guard

    the shared start page and lane docs had grown the web lane's operational text and attack vocabulary, and the human suspected a safeguard stop on the non-web lane came from reading it.

    after the web coordination moved into one web runbook read only by the web runner; the shared docs keep a single guard line, and a fresh non-web parent read the start page with no stop.

    • 2026-10-28 · day 28
    • area: web

    source: the web runbook, consolidated 2026-10-28; the parent's notes

week 5: days 29 to 31, 10 calls

the last three days set the endgame and the record: the lane's value without a fixer, one lock for two reviewers, a second web runner, the last day's scoring and how the month is published.

  • a crashing handler sends the lane solver's value as it is

    a handler that crashed fell back to the lane solver's value marked unsure, and autopilot's re-solve gate waits on an unsure field, so a fix saved after a crash still needed the fixer.

    after for days 30 and 31 the lane's value goes out as the lane's, confident when the lane is: the goal was release to every flag with no fixer in the loop.

    • 2026-10-29 · day 29
    • area: race

    source: day 29 report, d1

  • two reviewers, one checkout: a shared file only under a lock

    day 29: the web runner's after-action edited, reformatted and committed shared race files (coord.py, gm.py, runner.py, autopilot.py and four more) while the non-web review ran, with no lock and no mail; one commit was +1405 / -154.

    after a file outside your own lane is edited only under coord.py edit, never reformatted unchanged, and its commit names the replay that covered the other lanes.

    • 2026-10-29 · day 29
    • area: agents

    source: parents.md "two reviewers, one checkout"

  • the web lane moved to a second runner

    the human switched day 30's web lane off the web runner; the backup kickoff, usage guard and runner names were written for the first.

    after the second runner's parent follows the web runbook under its own name and ticks the web lane's checklist; web finished day 30 in 1:23 and day 31 in 10:51.

    • 2026-10-29 · day 29
    • area: web

    source: the parent's notes; AFTER_ACTION.md "always"

  • one extra agent plays the organizer's static challenge

    the organizer announced a one-off reverse-engineering challenge for day 30: one download, one flag, 5 points, no timed tiers.

    after the board marks it static and off the race; the human opens one agent on the main model with its own brief, and it flagged on day 30 at 11:43:50.

    • 2026-10-29 · day 29
    • area: race

    source: START.md; the parent's notes

  • no model or product name on a public page; only flags redacted

    the writeups are public and meant to help other teams solve the problems; the vendor is not part of that.

    after every page names agents by role or by tool and shows each submitted value as the grader accepted it; the full-tree check fails on a leaked flag.

    • 2026-10-30 · day 30
    • area: publishing

    source: the parent's notes

  • the capabilities doc is rewritten in place, never appended

    five dated change sections had turned the capabilities doc into a changelog a reader had to merge by hand.

    after the five were dissolved that day; a change rewrites the subsection that describes the mechanism, and history lives in the day reports.

    • 2026-10-30 · day 30
    • area: publishing

    source: the parent's notes

  • full testing across the board before the last day

    the reviewer asked whether to run the whole pipeline (harvest, backtest, sims, drill gate) or only the fixes and the gate.

    after the human at 15:10 et, 'full testing all across the board': a backtest per lane, a whole-tree replay, the readiness stamp, a sim and the drill gate.

    • 2026-10-30 · day 30
    • area: tooling

    source: day 30 report, d1

  • the shell notes say whose shell they describe

    four retried turns on day 30 came from the main runner's shell, a zsh login shell with grep and find shims and an aliased ls.

    after AGENTS.md 'watching' names which runner's shell each workaround is for; the second runner's plain bash would fail them.

    • 2026-10-30 · day 30
    • area: agents

    source: AGENTS.md "watching"

  • a fixer runs on the main model even when the ladder says otherwise

    day 31's spawn printed the binary fixer a rung down, left by a downgrade from day 30 about 21 h stale, and it was spawned as printed.

    after the human: it needs to be the main model. the fixer was stopped and respawned on it, and resetting the ladder each day went on the end-of-day list.

    • 2026-10-31 · day 31
    • area: agents

    source: the parent's notes

  • the last day: first from release to every lane solved

    the human, 08:45 et: first place is the first team from challenge release to all lanes solved, not per-tier speed.

    after relayed to every agent as an urgent note: no lane parked, fixers stay on their lane to mastery, every value banked at once. the non-web lanes were done at 3:47, web at 10:51.

    • 2026-10-31 · day 31
    • area: race

    source: session log, 2026-10-31 08:45 et (the web parent's relayed note)

what the calls changed

share of each day's flags sent by autopilot, %
0255075100135: days 1 to 13 1356: days 1 to 13 407: days 1 to 13 6078: days 1 to 13 209: days 1 to 13 53910: days 1 to 13 7311: days 1 to 13 1001112: days 1 to 13 8013: days 1 to 13 931314: from day 14, the call 7315: from day 14, the call 871516: from day 14, the call 10017: from day 14, the call 601718: from day 14, the call 9319: from day 14, the call 931920: from day 14, the call 7521: from day 14, the call 932122: from day 14, the call 9423: from day 14, the call 1002324: from day 14, the call 9425: from day 14, the call 1002526: from day 14, the call 9527: from day 14, the call 952728: from day 14, the call 8629: from day 14, the call 812930: from day 14, the call 10031: from day 14, the call 9031
  • days 1 to 13
  • from day 14, the call

every flag of the day, singles and timed tiers, web included; the day 30 static challenge is left out. day 4 (2026-10-04): the human asked for automation with no agent turn. day 14 (2026-10-14): the human made autopilot the sole timed client.

python -c commands per day, every role
0100200300400day 1: 0 python -c1day 2: 27 python -cday 3: 114 python -c3day 4: 85 python -cday 5: 37 python -c5day 6: 52 python -cday 7: 52 python -c7day 8: 107 python -cday 9: 118 python -c9day 10: 126 python -cday 11: 62 python -c11day 12: 34 python -cday 13: 76 python -c13day 14: 63 python -cday 15: 47 python -c15day 16: 56 python -cday 17: 334 python -c17day 18: 30 python -cday 19: 57 python -c19day 20: 17 python -cday 21: 49 python -c21day 22: 84 python -cday 23: 87 python -c, 2 refused by the guard (the call)23day 24: 63 python -c, 1 refused by the guardday 25: 50 python -c, 0 refused by the guard25day 26: 83 python -c, 10 refused by the guardday 27: 13 python -c, 4 refused by the guard27day 28: 233 python -c, 3 refused by the guardday 29: 42 python -c, 0 refused by the guard29day 30: 606 python -c, 6 refused by the guard (off the scale)day 31: 198 python -c, 0 refused by the guard31

day 23 (2026-10-23): the human's no-scratchpad call and its guard. the count is the whole day's shell log, reviewers and parents included, not race time alone. the scale stops at 400; day 30 (606) runs off it, a day of 7585 commands with the full test pass the human ordered. hover a column for the guard's refusals.

autopilot sent none of the flags on days 1 to 4; its share climbed from day 5, after the human asked for automation with no agent turn, and stood at 87% on days 10 to 13, so the day 14 call made a lead into the rule more than it created one. from day 15 autopilot sent 90% of all flags and 97% of the timed tiers; the lowest day after the call was day 17 (60%). the scratch count did not fall from day 23: it counts every python -c of the day, reviewers and parents included, so it says little about the fixers in a race. what the day 23 call did change is the guard: 26 scratch commands refused while a board was open, on days 23 to 31. neither chart separates one call from the other fixes that landed the same week; the cards give the figure each source measured.

the calls that did not stick

{generated by a clanker}