Echo Undertow

Bilingual (RO/EN, diacritics-folded) behavioral lexicon classified on two axes: temporal category — intention ("voi anula", "will cancel"), action in progress ("anulez", "cancelling"), completed action ("am anulat", "cancelled") — and direction class — cancel, switch, boycott (exit family) vs purchase, renew, recommend, share, defend, join (advocacy family). Epistemic certainty ("clar", "definitely", "categoric") is a separate stream: it amplifies a behavioral verb in the same clause and is otherwise never counted as commitment. Each clause passes heuristic guards before a marker is counted: negation (window before the verb: "not cancelling", "nu voi renunta", "nobody is cancelling"), conditional ("daca / if / unless"), quoted & reported speech ("zice ca", "they said", text in quotation marks) and hedged intention ("poate ma mut" is deliberation, not commitment). Negative-form exit phrases with built-in negation ("nu mai cumpar", "never again") are matched first and are not double-negated. These are clause-level heuristics, explicitly not a dependency parse; excluded detections are displayed so the exclusions are auditable.

Rates are per-1,000-token with add-0.5 smoothing on counts, so a near-zero baseline can never produce an unbounded multiple. If the baseline contains fewer than 3 behavioral instances, the relative multiple is declared not stably estimable and the stage is decided on absolute recent rates instead — stated in the output. No transition verdict is issued unless baseline + recent windows jointly contain at least 25 counted behavioral instances, the corpus holds at least 600 tokens and 3 time buckets, and at least 60% of documents are in a supported language (RO/EN): otherwise the verdict is INSUFFICIENT EVIDENCE, with the reason.

The pressure index and each proxy carry a 95% stratified-bootstrap interval (B=200; resampling clusters — author if supplied, else day if dated, else documents — within time buckets). Stage and direction are tested by block permutation at the same cluster level (commitment drift, hedging decline, direction balance) under Benjamini–Hochberg control at q=0.05. A passing test means: the lexical drift is unlikely under exchangeability of these clusters, under this model. It does not certify that people will act, that action follows in any specific window, that the marker sets are complete, or that the sample represents the audience. Onset is a CUSUM change-point on the committed-vs-hedging differential, with a bootstrap uncertainty window. Fragility is a greedy removal stress test: the verdict overturns after removing k top-commitment documents — an upper bound on the minimal overturning set, not a proven minimum. Bootstrap CI = sampling error; de-duplicated identification floor = manufactured signal; they answer different questions and are never conflated.

The linguistic trace is many-to-one in the latent state: a transition to action language is compatible both with genuine collective decision and with coordinated injection. The provenance screen computes near-duplicate share (3-gram Jaccard), burst concentration, author concentration (when ids are supplied) and an organic-only lower bound for the index via de-duplication. When contamination thresholds fire, the verdict is CONTESTED TRACE — indeterminate, per the non-identification result — and the fired condition is named with its values. The screen attributes nothing; it reports what the trace cannot exclude.

All analysis runs client-side; nothing is uploaded. Pasted-text analysis works fully offline. Three conveniences require network or bundled assets and degrade gracefully without them: CSV/XLSX import (SheetJS from CDN), webfonts (Google Fonts; system monospace fallback) and the PowerPoint exporter (vendor libraries). Baseline lock and run history use this browser's local storage only — nothing leaves the machine.