Every Android location tracker is unreliable for the same reason — OEM battery managers kill background processes you don't own. The fix was ~150 lines of Kotlin: a daily WorkManager job that POSTs a city-only fix to my own endpoint.

Adding film posters to the telemetry page, I tried to reconstruct Letterboxd's CDN URLs from the film id and slug. Three of four worked; Memories of Murder 403'd at every size. The reason: TMDB-imported posters live under a totally different path — /sm/upload/{hash}/ with no id and no slug — so they're unconstructable. The honest source is the film page's JSON-LD image field.

The PLACES tab needed a map of ~119 cities that recolors with the site's four themes. Leaflet and Mapbox can't do that — raster tiles ignore your CSS. So I rendered the map as plain SVG with d3-geo at build time: one <path> for all land, a <circle> per city, colored entirely by --field-ink / --field-bg. It themes for free because it's just ink on a background.

The telemetry page wants to show what I'm watching on YouTube, live. The Data API has no watch-history endpoint — the watchHistory playlist has returned an empty placeholder since ~2016 and activities is deprecated. So I stopped asking Google and read the data where it already exists: the watch page. A ~40-line MV3 content script beacons each video to a Redis key.

Added MultiheadAttention across the 4 commodities before the LSTM in panel-VLSTM. Mean Sharpe Δ −0.30. Crude_Oil specifically lost 1.17.

Phase A flagged Silver Kalman cointegration as a -0.51 NEGATIVE result. The negative result was a one-line bug — fillna(0) corrupted β at early NaN dates by inventing a $1 fictitious gold price.

Reduced min_train 2500→1500 expecting marginal change. NBEATSx/NHITS/PatchTST shrugged. TFT lifted Crude_Oil Sharpe 1.20→3.27.

Picking 14 pulsars from 3924 candidates via greedy geometric optimization — five strategies tested, GDOP wins on positional spread.

Cron scans last week's session JSONLs for repeating patterns. SessionStart hook nudges me to review proposals before any session starts. Threshold + rejection ledger keep noise out.

Period divided by hydrogen's 21cm spin-flip frequency, integer to binary, binary to tick marks.

Claude Code, Codex, OpenCode, Kimi all spell the same model differently. One alias map collapses them.

Hashing source:session:provider:model:day into a primary key means re-running the sync is a no-op.

ascii-shader-tsx is markdown reference docs, not a library. The agent reads, emits one inlined .tsx file. Distribution unit shifts from npm to instruction set.

Day/night/rain/snow state machine driving ANSI colors on Nerd Font glyphs — using actual sunrise/sunset, not hardcoded hours.

Async LibraryManager handles I/O; a dedicated mpsc thread handles playback so Tokio's scheduler jitter never reaches the audio pipeline.

compute_alpha_features() winsorized once, the CV loop winsorized again, inference winsorized once. The model trained on a different distribution than it predicted on.

Signum's risk check was reading a weekly-stale equity curve. A 15% mid-week drawdown was invisible until the next Wednesday's rebalance.

fetch_sp500_tickers() scraped today's index and used it as the historical universe. Two years of training data, zero defaults, zero acquisitions, zero scandals.

Two compounding bugs killed the CLI-based daemon: Windows publishes only the socket filename not the path, and the CLI doesn't expose foreground process names anyway.

wslhost.exe blocks PID inspection from Windows. Workaround: zsh preexec emits OSC 1337 user-vars, WezTerm Lua writes JSON, Rust daemon updates Discord.

I have a crude-oil arb running across Hyperliquid, NYMEX, and India's MCX, and the obvious next move was to point it at gold. The backtest printed a clean +0.4%-a-morning edge at 70% hit. It was real for about an hour — until I checked whether you could fill at the price the edge depended on. You couldn't. The teardown: a research gap that wasn't an alpha gap, a look-ahead bug that wasn't the killer, a stale opening print that was, and the moment I realised my live strategy fills on the exact same field.

cpal::Stream is !Send. Pulling headphones on a Windows WASAPI setup silently kills the audio stream and the obvious fix — spawn a thread that rebuilds it — is forbidden at the type level. The fix that works: an AtomicBool, the existing 250ms UI tick, and no new thread.

A commodity ML strategy started Phase A with an oracle Sharpe of 2.43. A hostile external critique took it to an honest 1.94 — then a second multi-agent audit found a bug inside one of the honesty fixes itself, landing it at 1.66. The trajectory of how every 0.1 Sharpe got found, traced, and given up — twice.

Close to two years in. The films I watched while I couldn't write directly, and what they kept saying back. Kurosawa, Resnais, Lou Ye, Yamakawa, Rohrwacher, Oshiyama — and a line from Dazai.