flan
8cefa325a6
Custody: rows+chain commit atomically per batch; audit gains reverse check (unchained content fails), opens read-only, refuses missing stores. Gate: tie-averaged Spearman (order-invariant; old impl could flip the gate by alphabetization), both raters required, peeks labeled + unservable + never pass. Manifests: corpus_hash covers full records (manifest_version 2). Embedding cache self-invalidating (text-hash + model@revision key, chunked reads, one implementation). serve: --config, health survives bad timestamps, routes generated from artifacts.KINDS. cycle command owns window policy; compose ships the collector. artifacts.py pins byte-determinism (utf-8, sorted keys) for every writer. congress.py (dead custody-bypassing draft) deleted. stats.py + hardening regression suite; 60 tests.
5.6 KiB
Glossary
Plain definitions of every technical term used in this project's docs. If you hit a term anywhere in tiltmeter that isn't here, that's a bug — file it.
RSS feed
A machine-readable list an outlet publishes of its latest articles: headline, link, timestamp. It's how tiltmeter knows what each outlet published without scraping their whole site.
Corpus
The full collection of articles tiltmeter has gathered. Scores are always computed from a stated slice of it, never from "the web" loosely.
Snapshot
A frozen slice of the corpus: all articles from a stated time window, pinned by their fingerprints. Ratings are computed from snapshots so results can be checked and re-run later.
Fingerprint (hash)
A short code (SHA-256) computed from an article's exact text. Change one character of the article and the code changes completely. Fingerprints prove that a rating was computed from exactly these articles and no others.
Manifest
The published list of what's in a snapshot: each article's link, headline, outlet, timestamp, and fingerprint — but not its full text (which stays with the outlet). Anyone can re-download the articles and verify the fingerprints.
Embedding
A way of turning a piece of text into a list of numbers so that similar texts get similar numbers. tiltmeter uses embeddings only to spot when two articles cover the same story — never to judge them.
Clustering
Automatically grouping similar things. Here: grouping all outlets' articles about the same news event into one "story."
Story cluster
One news event and all the articles about it, across outlets. The unit of tiltmeter's analysis: for each story, which outlets covered it?
Coverage matrix
A big grid: one row per outlet, one column per story, marking who covered what. All of tiltmeter's scoring reads off this grid.
Correspondence analysis
A standard statistical method that turns a grid like the coverage matrix into a map: things with similar patterns end up near each other. The "seating chart" in the plain-language docs.
Principal axis
The single strongest direction of difference in that map — the line along which outlets' coverage choices differ most. tiltmeter checks whether that line matches political lean, rather than assuming it.
Anchor
The external reference that tells the map which end is politically left and which is right. tiltmeter's anchor is language from congressional speeches — public records, not anyone's opinion of an outlet.
DW-NOMINATE
A long-running political science measure that places members of Congress on a left–right scale using their actual voting records. Public data, published at voteview.com.
Bootstrap
A standard way to measure how sure we are of a number: recompute it many times on random re-samples of the data and watch how much it wobbles.
Confidence interval
The honest error range around a score, produced by the bootstrap. "−0.3 ± 0.1" means: don't read anything into differences smaller than that.
Spearman rank correlation
A statistic (written ρ, "rho") measuring how similarly two lists rank the same items, from −1 to +1. We use it to compare our outlet ordering with the commercial raters' ordering.
Sensitivity analysis
Re-running the pipeline with the judgment-call settings changed (time window, grouping threshold) and publishing how much the scores move. If a small setting change swings the results, that's disclosed, not hidden.
Tunable
A pipeline setting that is a practical judgment call rather than something the research literature dictates. Every tunable is labeled as such in METHODOLOGY.md and covered by the sensitivity analysis.
ADR (decision record)
A short numbered document recording one methodology decision: what was decided, why, what the alternatives were, and what could go wrong. Decisions are never edited in place — a new record supersedes an old one, so the method's history stays visible.
SemVer (semantic versioning)
Version numbers with meaning (MAJOR.MINOR.PATCH). Any change that could alter published ratings bumps the version, so every ratings file names the exact method that produced it.
Reproducibility
The property that anyone can re-run the pipeline on the same snapshot and get the same output, byte for byte. The project's core promise.
Ideal-point estimation
The political science technique of placing actors on a scale from their observed choices — classically, placing legislators by their votes. tiltmeter places outlets by their coverage choices: outlets are the legislators, stories are the votes.
Lede
The opening sentences of a news article, summarizing the story. tiltmeter embeds headline + lede to group articles into story clusters.
Content-addressed storage
Storing each piece of text under its own fingerprint. The fingerprint is both the address and the integrity check: if the text changes, it no longer matches its address. Identical texts are stored exactly once.
Custody chain
The dataset's tamper-evident logbook. Every batch of newly collected items
adds a line, and each line is locked to the line before it by a fingerprint.
Changing or removing any old line breaks every line after it — so the
collection's history cannot be rewritten quietly. tiltmeter audit checks
the whole chain.
Append-only
Data is added, never edited or deleted. Collected articles and speeches are append-only in tiltmeter; only derived caches (like embeddings) may be rebuilt.
Byline
The author credit on an article, stored exactly as the outlet published it. Collected because it cannot be backfilled later; not yet used by any signal.