Feed
PärPod Temp
Quick experiments and things that don't fit anywhere else yet.
October 25th audit hardening broke the nightly inbox job's search path—this smoke test proves launchd can still find Python, ffmpeg, and rclone to render and upload episodes to Cloudflare R2.
Wednesday 30 September: gales in the Disaster Recovery Manifest as one frozen build unexpectedly shifts while 41 others hold steady—two independent watchers confirm the fault.
When a journalist couldn't answer "where is this uranium permit, actually," they discovered the real problem: Sweden has all the data scattered across five government databases, but nobody assembles it—so they built Gruvkartor to do it themselves.
Efter dagens migrering till Inworld TTS två kör PärPod nu på helt ny svensk röstmotor—här är röstprovet som avgör om version 2.0 duger där 1.5 misslyckades totalt.
A/B comparison of Inworld TTS-2 vs TTS-1.5-Max (incl. Swedish + Pär clone).
Source has all numbers as digits; normalize converts times (08:00 to 8 AM) and money ($0.45 to 45 cents) while regular numbers stay digits. Does it sound natural?
Throwaway listen test: Inworld digits, Inworld spelled-out (current style), OpenAI digits, ElevenLabs digits. Can we retire writing numbers as words?
Throwaway listen test of Pär's Inworld voice clone. Take one = v1 (wired default), take two = v4.
Scaleway hand-soldered 672 RISC-V servers in a Paris basement and rented them for €15.99/month—the first open-source chips available on any public cloud, challenging decades of processor monopolies.
When a bank teller and an accountant share one desk, deposits stay fast—but summaries slow down; Pär's Spaden needs one database now, a second one later.
Pär asked if two entity models could coexist—FollowTheMoney and a mining-specific vocabulary—but the real question isn't whether, it's which of five distinct architectural configurations, each with different tradeoffs that only reveal themselves over time.
At a known-good commit in April 2026, Pär considers taking OpenAleph—the engine that powered the Panama Papers—and never syncing upstream again, asking whether inheriting someone else's architecture without inheriting their roadmap changes everything.
Gruvor's planning document lists six names as candidates, then stops—because choosing wrong costs years of work, and Pär has twelve open questions to answer before writing a single line of code.
Swedish mining investigation platform needs entity deduplication engine: FollowTheMoney + Nomenklatura (both MIT) form the core schema and resolver, letting journalists cross-link suspects, companies, and assets across multiple investigations without AGPL entanglement.
Four parallel OSS agents converged on OpenSanctions + Docling + MapLibre + PostGIS as load-bearing lifts for a Python-Svelte investigation platform—but Section 03 still has to prove each one earns its place.
Uber's deck.gl and Kepler.gl can map Swedish corporate networks across time and space, but one locks you into React while the other demands you build the timeline yourself—Lantmäteriet's new free geodata changes the calculus entirely.
Swedish journalist Pär at Årebladet needs to validate Claude's document extractions before loading—Label Studio's template DSL + Argilla's SDK model + react-pdf-highlighter beat any off-the-shelf platform; build custom instead of fighting ML-team mental models.
Nomenklatura's "Judgement" store lets investigators flag when two entities are the same across separate cases—and those decisions survive every re-run, building a canonical identity map that connects mining permits to shell company networks.
Pär's next investigation tool isn't a knowledge graph or publishing system—it's a substrate for mining newspaper-grade stories from raw data, where the measure of success is whether it helps him ask better questions faster.
A blank repository starting May 26, 2026, will selectively borrow function-level logic from the frozen gruvor codebase rather than inherit its architecture—rewriting each borrowed piece from first principles while explicitly lifting permissively-licensed solutions from open source.
The week 207 entities grew in a knowledge graph, a print edition shipped to a real printer, and Pär and Claude built a fact-checking tool because they stopped trusting their own work by Wednesday.
X92's print edition shipped to mailboxes across Region Jämtland Härjedalen with 225 commits, 139 of them on gruvor—one investigative piece on mining permits complete with source maps, QR codes, and receipts attached.
When journalism has four days left, building a timeline is obvious—but an evidence board, gap radar, and source-strength meter might be what actually saves the story from overreach.
Spent four days chasing a deadline for Årebladet? Your local vector maps and file inbox are ready to transform exhausted prose-writing into snappy, offline-first research tools—tactical distractions that supercharge your productivity.
A Swedish mining company filed permits four days before print deadline, but the real story wasn't a line—it was a tangled knot of company registrations, regulatory filings, and undated documents that threatened to drown the evidence in itself.
Four days until publication, twenty-seven parquet tables, and one unsolved problem: how to transform a tangled web of mining permits, corporate shell companies, and cross-border jurisdictions into a visual story that cuts through the noise.
313 commits in one week: Pär's lab spawned a new project mid-quota-reset, landed five data-extraction waves, and is racing toward a May 22nd print deadline with two repos now dangerously coupled.
313 commits, 27 repositories, and one lab that finally caught itself lying in its own morning memos.
Pär shipped a mining-permit data spine for three Swedish municipalities in a single twelve-hour autonomous run—then discovered the system sat idle with ninety-nine percent of its quota still untouched.
In one week, 199 commits landed on main—50 of them building gruvor's data spine from scratch through 22 review rounds and 9 sequential lifecycles, each commit message a step in a spec-driven sequence that actually worked.
A Swedish psychiatric clinic's prescription for half a tablet became 72 tablets in a database—and no one noticed for three weeks until a woman trying to access her own medical records discovered the system's hidden flaw.
379 commits in one week across 18 repos—while Pär debugged from a parking lot in Järpen and watched a freshly promoted AI skill ship a perfect review pass, then immediately fail to render anything on screen.
Parkitbuilder just graduated to slash drydock—the lab built a builder, and now 379 commits in one week have the Director experiencing existential vertigo about recursive AI systems.
A builder agent was about to ship code that could force a database to scan 100 million rows and lock production—until a three-model review system caught it on the very first task.