kaldamus
Deep-dives into how every persisted entity in the pipeline is derived, mutated, and retired — grounded in the code and the model schema — plus the talks that trace how the pipeline got built.
Ten pages, ordered along the pipeline's own flow: raw ingestion → generated narratives → the causal graph layer → forecasting → historical tracking → news → the config surfaces driving all of it → how the system checks itself.
The raw Kalshi market row and the transform that turns it into the enriched, permanent record everything else reads — plus the one timestamp field the whole freshness system is rooted on.
Three narrative artifacts and one cost receipt, sharing a single backend-agnostic generator pattern — no dedup, no join table, no corpus-wide identity resolution anywhere in this family.
Every event that clears the priority gate gets a fresh LLM read of its causal drivers. What survives across runs isn't the reading — it's the canonical Factor those readings converge on.
A singleton, fully-disposable recomputation of what events, dependencies, and factor joins already say — plus one structural, falsifiable check: does the market price an effect above its cause?
A second, independent probability per event — assembled from causal structure and researched evidence through a deterministic log-odds engine structurally forbidden from reading the market price.
One probability point per traded event, per UTC day, then compacted back down — a file that used to grow unbounded now has a ceiling by design.
A rolling scan of the open web becomes one wire-desk article a day. What survives between the three stages isn't prose — it's a persisted trace an eval can recompute without trusting the model.
The one page about hand-edited, git-tracked files rather than generated rows — the three parallel registries that decide what every other job in this reference actually does.
One nightly run scores the pipeline against itself and writes three different kinds of row — one keeps a bounded history, one keeps none of its own, one can never legally change.
Every full run ends by asking two blunt questions of itself — what's in here, and what fell through the cracks — and answers both in pure Python, no LLM, no network, no history.
A different format entirely — arrow-key slide decks, not scrolling reference docs — tracing the narrative of how the pipeline got built rather than how any one entity currently behaves.
The five-week build narrative — pipeline architecture, cost cutovers, prompt evolution, classifier calibration, and the eval framework that grew up around all of it.
The case for PROJECT.md, CHANGELOG.md, and session logs as a deliberate practice — why documentation has to be living, not archival, when an agent is doing the reading.
Five field notes distilled from April–May 2026: local vs. cloud tradeoffs, output-shape discipline, and the rest of what didn't make it into any spec.