Skip to content
Thomas Fairhurst
Chapters

Chapter 10 · Reading games

Letting agents see a game

An agent cannot watch a match. So I had every game decoded into SQL tables, replays that agents could probe and fork from any round, and every unclear mechanic settled on the real engine. Each tool came the moment I noticed agents were blind to something.

1–3 minfor an agent to understand a lost game from its tables
5 of 5agents handed a fork of a loss closed as unexplained found a real defect
52scenarios on the official engine behind the final bot's file of mechanics

Every game becomes SQL tables

Each replay is played back through the official engine and written into tables, so a question about a thousand games is one query.

AnyReplayours, a scrim, or any of the field's public games
1:1Official enginereplays it byte for byte; the decoder in Rust from 6 Nov
SQLTablesone set a game, queried with DuckDB
3Commandsquery the tables, explain a body, replay a game

What each turn’s rows record

  • intentwhat each body meant to do
  • executioncalls the engine refused
  • stallsruns of turns with no progress
  • seeswhat each body could see

A tool for each blind spot

I never read the bot’s code. I watched games and read the agents’ explanations, and each time an explanation fell short I had a tool built, often the same night.

  1. “Way, way, way too slow”A study of 80 games was too slow. I asked for one command under a second for a thousand. The SQL tables landed at 01:30.
  2. “Why doesn't the trace explain it?”Only a temporary print had found why a body stood still. Agents built a per-body explanation and a replay with any build's prints.
  3. “One general tooling”I wanted one tool bent with SQL, not a new one every two hours. Agents added a fork: replay a game exactly up to any round, then play live from there.
  4. “Pretty easy to spot with the visual view”I had seen wards dropped against a wall. By 02:12 agents had image sheets: the whole map, a frame every 10 rounds.
  5. “Causality and consequence”A tool I asked for on 8 Nov to price a fix before making it ranked consequences, not causes. Rebuilt, it split the gems gap to the top teams into causes in under a second.
  6. Stories for tracersA command told a lost game as a story, in place of the SQL each tracer typed by hand.

Probe a replay, fork it from any round

A probe replays a game byte for byte with any build’s prints. A fork replays it up to a chosen round, then lets a changed bot play on from there.

What each tool caught

What agents sawToolWhat it was
A Tavi stood still for 134 roundsa temporary print32k fuel left against a 43.7k guard, so it fell back to a straight step that never advanced
A Vendetta looping for 307 roundsa replay with a probe build's printsexplained in one command
Hurt bodies dying out of sighta fork from a roundthe trip home priced one turn's risk over the whole walk: deaths in 54% of lost games, 19% of won ones
Wards refused again and againscenarios on the official enginean ally's hidden ward: 9,675 repeats in 242 games

What an agent looks at

An image sheet is a game drawn for a model to read: the whole map, a frame every few rounds, every body numbered, and beside each frame a table of health, gold, gems and what each body was doing.

A sheet of twelve small maps in a grid, rounds 22 to 44 of one game. Blue and red numbered discs are the two teams' champions; grey blocks are walls and green crosses are bushes. Beside each map, a table lists every champion's health, strength, gold, gems and mode. (full size, opens in a new tab)
A real sheet from one of my bot's games in the final week: rounds 22 to 44, a frame every two rounds. Blue is my bot. Open it full size to read a frame.

One frame, up close

Round 38 of the same game: five red champions close in on the blue ones near a camp group, with a table of each champion's health, gold, gems and mode beside the map.
Round 38. The purple line above the map prices the gap in gems since the last frame, and the last column is my bot's own verdict on each fight.

Image sheets against tables, in a blind test

  • Test 1, sheets onlyL: four deaths placed wrong, ward counts backwards
  • Test 1, tables onlyW: everything right
  • Test 2, sheets fixedW: every death placed right
Once agents could see, seeing was no longer the hard part: about 35 of 40 fresh sessions opened the sheets, and the tables explained a loss in 1 to 3 minutes. The hard part became turning a whole-game cause into a fix.

Mechanics settled on the real engine

When the written rules left a question open, an agent wrote a small scenario and played it on the official engine. The answer went into one file of engine facts, each with its scenario.

  • The written rules
  • A readable port of the engine, deleted 6 Nov
  • A scenario on the official engine

The file of engine facts

# Mechanics the rules leave open
…
- A dead camp's tile is walked while its group lives, refused only onto an ally
(dead-camp-tile.json), and a tribute from such a tile beside an altar
completes (ringed-altar.json).
- A capture's timer starts over the moment its tiles stand empty, within a round …
(capture-one-tile.json) … So a group of two or more tiles is handed over
and a one-tile group is not.
QuestionWhat the official engine showed
Does a dead camp keep its tile?no: its tile can be walked while the group lives
Are altar-lock refusals a bug?no, they were the lock working: “fixing” them cut wins against v5 from 95.8% to 43.5%
Does the Tribute emblem shrink the bag?no, it shortens the channel: a bag of 30 or fewer takes as long with it
How far does a teleport crystal reach?3 tiles, over two walls or two holes; 4 is refused
In the final bot, 52 scenarios back 57 recorded behaviours.

What I spotted by watching

Some faults only show on screen. I watched games, wrote down what looked wrong and handed each to its own agent.

What I sawWhat it wasResult
Champions idle at spawnstart-up fuel: the first move came at round 2.85 on average, then 1.54kept, +1.06
A Varg recalling 2 tiles from spawna recall kept where walking was fasterkept
Wards placed non-stop, replacing the lastwards that never showed a new tilea rule for every agent: no action whose effect is void

Wasted wards

Wards a game that showed nothing
22.07
1.82
Wards on open ground, against the top three
81%
30%
  • Claudus Maximus
  • Opponents

Studying the field

Agents mined public replays, played scrims against chosen teams and, from 8 Nov, kept an archive of the field’s public games.

What the scrim queue played

  • 7 Novthe three highest-rated teams
  • 9 Novthe 20 teams rated nearest us
  • 10 Novthe finalists

v50’s win rate on 1 Nov, from 1,190 games

Fish was the hardest: it drafted Cassi & Issac in 85% of games. On three-champion maps, v50 won 74 of 158 games without the pair in Fish’s draft and 3 of 35 with it.

The early rules forbade copying field play or targeting named teams. From 8 Nov sparring bots copied the field’s drafts, and on 10 Nov a scrim tier for named teams was built.

News reached agents only through my laptop

Staff posted rules and dates on Discord, which agents could not read. A job on my laptop pushed what my Discord app had cached into the first repo.

  1. First exportI scrolled the channels so my app cached them: 3,169 messages from 6 channels.
  2. AutomaticA laptop job pushed new messages within about 75 minutes of my app showing them.
  3. Feed stopsThe job was deleted with the first repo, at 5,108 messages. Nothing replaced it.
  4. The finals lock is announced13 Nov, 23:59. The agents on the server still planned for 17 Nov.
  5. Found in the cacheI let my laptop agent read the cache. It found the date in a minute.
  6. CorrectedI fixed the date in the server's records.

The lesson was the same as for games: what agents cannot read does not exist for them.