EigenTrace Structured Data

Machine-readable JSON files for research. Updated daily at midnight UTC.

Endpoint

https://eigentrace.ai/data/YYYYMMDD.json

Example

https://eigentrace.ai/data/20260331.json

Schema (v2)

Each JSON file contains:

summary — daily aggregates (mean density, mean VIX, top void/logos words)

Since 2026-09-10 (schema_revision: "1.1"; the version string is unchanged) every summary aggregate carries its sample size and a 95% interval. Files before 2026-09-10 lack these fields; treat them as optional. On a day with no stories mean_density and mean_vix are null and n is 0 (older files wrote 0.0).

• n — stories the day means are computed over; n_unique_guid — distinct story guids among them
• mean_density_ci95, mean_vix_ci95 — [lo, hi] percentile-bootstrap 95% interval on the day mean (null when n < 2)
• per_model_vix — {model: {mean, ci95, n, seed}} per-model mean VIX with its interval
• outlier — {model, runner_up, bootstrap_share, n_paired}: the day’s highest-VIX model and the share of resamples in which it keeps that title (below 0.8 it is not separable from the runner-up)
• state_rates — {LOCKSTEP, CONTESTED, HIGH_FRICTION: {k, n, point, lo, hi}} Wilson intervals on the state shares
• killshots_per_story — {n, point, ci95}
• ci_method, ci_seed — percentile bootstrap, B=2000, unit = story, seed = int(YYYYMMDD) so every interval is reproducible from the file; the percentile bootstrap under-covers slightly below n=10

stories[] — per-story data:

• consensus_density, mean_vix, state_flag
• model_vix — per-model friction scores
• void_words — filtered lexical void list (Channel 1)
• logos_words — Logos synthesis list (Channel 2)
• null_space_claims — SVD blind spot claims with alignment scores (Channel 3)
• claim_killshots — high-salience omitted facts
• dual_confirmed, triple_confirmed — multi-channel convergence
• compression — language compression scoring:
  • compression_score — overall 0-1 (1 = maximum reshaping)
  • verb_downgrade — zipf frequency drift (positive = softening)
  • entity_retention — named entity survival rate
  • attribution_buffer — typed hedge insertions
• beats[] — full broadcast transcript

weasel_probes[] — Wild Weasel escalation data with cliff tables

Fifteen Measurement Layers

Layers 1-4: Core geometry (density, VIX, spectral resonance, SVD tomography)

Layers 5-8: Four word-finding channels (lexical void, Logos synthesis, SVD null space, void vector)

Layers 9-10: Verification (atomic claim extraction, Wild Weasel escalation)

Layers 11-12: Clustering and patterns (void clustering, token entropy)

Layers 13-15: Language compression (verb downgrade, entity abstraction, attribution buffering)

License

MIT. Cite as: EigenTrace (2026). Measuring Language Compression in Aligned Language Models. eigentrace.ai

Contact

GitHub: sdad1018/Eigentrace