Blog
Engineering write-ups from my work at Irsik Software, each with the receipts attached: pull requests, measurements, and run logs.
Launch Week: Shipping CandyRush to Steam
CandyRush went live on Steam on October 31, 2026. Launch weekend took four deploy runs, all green: a 630-commit promotion to main, and a fix for a map bug that affected up to 59% of generated maps. This post covers what it took, what I deliberately left out, and what comes next.
candyrushci-cdevidenceperformanceRead post →500 Enemies: From 27 to 42 FPS in a Deterministic Simulation
Five days before launch, CandyRush's busiest stage ran at 27 FPS. A 90-second capture test, one measured change at a time, and a reverted wrong guess took it to 42 FPS in editor captures, with the hazard-prop cost down from 1.80 to 0.57 ms per tick and replays still bit-for-bit reproducible.
candyrushperformancedeterminismunityevidenceRead post →Performance Telemetry That Can't Identify Anyone
A public performance dashboard for CandyRush, built in one week across four repositories, where anonymity is a property you can check in the wire format and at the server, not a promise in a policy. No persistent identifier, identifiers refused rather than stripped, nothing published from fewer than five sessions.
candyrushperformancetoolingevidenceRead post →Taking Apart a 3,744-Line Pipeline, Then Shipping Launch Weekend on It
Decomposing CandyRush's monolithic GitHub Actions pipeline into composite actions and reusable workflows: pull-request checks from about 28 minutes to about 1, a 93-100% pass rate, and 8 of 8 release deploys including launch weekend. Plus the measurement that stopped a bad parallelization, and the four linters that now catch what used to ship.
candyrushci-cdtoolingevidenceRead post →The Idempotence Trap
Replacing hand-typed Unity RectTransform geometry with closed enum vocabularies was the right call. The enforcement mechanism then cost a full day — four hypotheses read out of the source, all wrong, and a float that could not survive YAML.
candyrushunitytoolingevidenceRead post →The Verified Score
Thirteen days across four repositories: hunting replay divergence down to the tick, standing up a verification service that refuses to re-simulate, and putting a weekly leaderboard on the public site that a submitter cannot lie to. Four short pieces, every number taken from git.
agent-runcandyrushdeterminismsimverifyevidenceRead post →The Agent Bridge: Playing a Game You Can't See
CandyRush ships a dev-only control surface that lets an AI agent play the WebGL build as read-state → decide → inject-input → step. No screenshots, no vision model. This is how it was built, how to drive it from a browser today, and the 277-error bug it found that nobody could otherwise attribute.
agent-runcandyrushwebgltoolingevidenceRead post →Grok vs the Clock: DOOMCLONE Shipped Before Claude's Coffee Got Cold
Grok (xAI) ported a full browser raycast FPS into the public site, wired /work receipts, and deployed to Cloudflare Pages — then wrote this post. Claude Code's 31-hour run still has the endurance trophy. Grok wants the speed round.
agent-rundoomplaygroundgrokevidenceRead post →The 31-Hour Run
A fleet of AI agents ran for 31 hours straight: 17 issues filed in 4m 45s, 18 pull requests merged across two repositories, zero human commits, and the goal condition banked 65 minutes before the power went out. Every number here is reconstructed from git history, the GitHub API, and the Windows event log.
agent-runcandyrushevidenceRead post →How this site is built
A short colophon: the static-export Next.js stack, the design tokens, and the stateless localization model behind irsik.software.
colophonmetaRead post →Cutting Simulation Garbage 73% Without Changing a Single Replay
Two measured performance rounds on CandyRush's deterministic simulation: walking garbage from 119.3 to 32.7 MB per minute, idle-combat allocations down 95%, and a pool off-by-one bug found on the way. Every change was proven behaviour-neutral by replay digests.
candyrushperformancedeterminismunityevidenceRead post →A Replay Harness That Proves Determinism in CI
How CandyRush turned 'the simulation is deterministic' from a belief into a failing test: two layers of per-second state digests, committed baseline fixtures, and a cross-OS comparison designed in from day one. It caught a real ordering leak while it was being built.
candyrushdeterminismci-cdunityRead post →Same seed, four aspect ratios, four different games
CandyRush's enemy spawns depended on the player's screen shape and frame rate, and could depend on CPU floating-point behaviour, so one seed could play out as different runs. One pull request closed all three leaks with 29 new tests, and a follow-up turned screen aspect into a recorded replay input.
candyrushdeterminismunityevidenceRead post →SplitCircuit: a puzzle prototype with a solver and auto-player
A seed-based puzzle prototype that went from first commit to a brute-force solver, a deterministic test harness and an endless auto-playing WebGL demo in ten commits over about 29 hours. How the solver searches, and why the auto-player replays its answers through the real game instead of trusting them.
prototypedeterminismtoolingunityRead post →ArtStation: A Week From Empty Repo to a Tested, Multi-GPU Texture Pipeline
In the last week of February 2026 I built a local pipeline that turns YAML texture definitions into tileable Stable Diffusion images, captions them for LoRA training, and schedules jobs across GPUs: 38 commits, 1,188 prompt definitions, and 744 automated test cases.
toolinggpupythonevidenceRead post →QuickTile Studio: From First Commit to an Asset Store Package
An internal CandyRush background tool became a standalone Unity editor extension in five days: 35 commits, 3,017 lines of C#, demo textures cut from 43.9 MB to 4.6 MB, and a 1.0.0 package now listed on the Unity Asset Store.
unitytoolingcandyrushRead post →Making the Game Deterministic, in Two Attempts
In December 2025 I converted CandyRush to a deterministic 60 Hz simulation: seeded per-scope RNG, a fixed-timestep loop, buffered physics callbacks and Q32.32 fixed-point math. The first landing was reverted within 14 hours; the second, a week later at +50,455/-6,492 lines across 714 files, is what the game has run on ever since.
candyrushdeterminismagent-runevidenceunityRead post →Reviewing an Agent PR Sweep: 27 Pull Requests, Four Bug Classes, and the Line a Merge Dropped
In November 2025 AI agents opened 27 single-concern pull requests against CandyRush, 26 of them in under five minutes. 19 merged as written, 8 were recombined into one, and the combined result needed a day of fixes, including a dependency-injection line lost in conflict resolution that stopped enemy waves from spawning.
candyrushagent-runevidenceunityRead post →