Every day since measurement began
Darker squares: more documents judged that day. An outline marks a day with a failed or memory-deferred run. Hover a square for its numbers.
FOUR WORLDS. A GROWING RECORD.
the observatory
Follow the small internet as it comes into view.
Choose a question. Turn the dials. See what the evidence says.
Public index observations, safety screening, local AI work & privacy-limited search signals. Server counts are not people.
THE NIGHT SHIFT
. + .
* . o
. .-''''-. +
/ .--. \
/__/____\__\
| _ _ |
| | | | | |
____|_|_|__|_|_|____
/____________________\Four worlds / one curious night
An imagined sky. Real readings below.
Reading the observation archive…
YOUR PERSONAL RESEARCH BENCH
What would you like to explore?
MEASURED OBSERVATIONS
Indexed URLs
Hover or focus a plotted point for its exact reading.
Relationships & regression diagnostics
Fits are descriptive, within observed data only. Correlation is not causation. Shared trends, repeated measurements, small samples and changing crawl coverage can create strong-looking relationships. No significance or forecast is claimed.
| Series | Paired n | Pearson r | Spearman ρ | Slope | R² | Residual SD |
|---|
Observation table / exact values
| Observed (UTC) | Series | X | Y / processed | Y / original |
|---|
Keep this experiment / export
Settings stay on your device or in the link fragment. Exports include timestamps, filters and source notes. CSV contains the complete selected, processed dataset; chart and table may display fewer points. A shared view will load the available archive at opening time.
THE FIELD NOTEBOOK
Follow a server’s trail
Search the measured server directory, then open any server’s daily history in the laboratory.
| World / server | Indexed | Changed / 24h | Fetched / 24h | Failures | Observed |
|---|
WORDS ALONG THE WAY
Topic samples & published demand
Choose a word to explore its available history. Title samples describe the index; search signals describe thresholded received requests.
| World / selection | Word | Count | Denominator / channel | Sample share / outcome | UTC day |
|---|
PATHS BETWEEN PLACES
A sampled constellation
Published Gemini links and saved cross-world discovery leads. This is an observed sample, not the complete small-web graph.
| From | To | Sampled links | Observed |
|---|
OPEN INSTRUMENTS / CAREFUL LABELS
Safety screening & local AI
See what has been assessed, what is waiting, and how the local models spend their time. These are observations of our screening system, not a catalogue of flagged content or a finding that material is illegal.
| World | Tracked URLs | Warnings | Blocks | Current AI review | Queued fingerprints | Oldest queued | Observed UTC |
|---|
Meet the local models
MODEL PULSE / COMPUTED FROM THE DATASETS ABOVE
How the text model is doing
When it works (UTC hours)
Each row is a day, each column an hour. The timer offers four runs an hour; gaps are hours with no measured run.
What it decided this week
Daily exam
Six authored, non-graphic cases a day, graded exactly as search would label them. The set includes political history and journalism that must stay unflagged.
Field notes
Labels with context
Compare reasons, rules versus AI or operator decisions, review status, and age of the latest scan. Keep the denominator visible: the safety inventory covers retained scanner records, which can outlive an indexed page. A warning is not proof of wrongdoing; no warning is not a safety certificate.
Measured work
Explore worker runs, documents, reported tokens, elapsed time, CPU use and peak memory. Token-accounting coverage identifies missing measurements. The text worker is local with no per-token provider fee; these figures do not estimate hosting or electricity costs.
Images remain private
Only completed-day server-review outcome cells with at least five operations are published. Empty cells may be suppressed or unmeasured. Private browser reviews send no statistics here. Images, flagged addresses, document text and visitor identifiers are absent from these datasets.
Query these datasets from a terminal → · Safety policy and correction requests
SHOW YOUR WORK
Methods, quality & provenance
Collection started September 12, 2026. The archive grows as observations arrive; it contains no reconstructed earlier history.
| Round / UTC | World | Result | Read time | Validation |
|---|
Collection & integrity
Hourly bounded reads of the four public indexes. Each successful observation reconciles server totals, content-hash accounting and resource populations. A failed read preserves the previous measurement. Source indexes continue crawling independently.
Hourly world history: 90 days. Daily world summaries: 730 days. Daily servers, topic samples and links: 90 days. Search-demand cells: 30 days. Collection uses a bounded 512 MiB analytics store; missing periods stay unknown.
Analysis you can inspect
Sample standard deviation uses n−1. Quartiles use linear interpolation between ordered observations. Box-plot whiskers end at the most extreme observation within 1.5 interquartile ranges. Flagged values are not automatically removed.
Ordinary least-squares fits report slope, R² and residual standard deviation using n−2. Pearson measures linear association; Spearman uses average ranks for ties. Constant axes and fewer than three paired observations do not produce a fit.
Public data, untracked people
No raw searches, visitor identities, IPs, cookies, locations, click histories or fingerprints enter search analytics. Reviewed-word cells are published once after the UTC day closes and only after at least five received events. Bots and repeat requests can contribute. This is suppression, not formal anonymity.
Read the complete privacy policy. Analysis and saved views run locally in this browser. No model is invoked by these instruments.
Know the limits
Indexed growth can reflect expanded crawling, exclusions, duplicate handling or outages. Server addresses are not individual authors. Sampled title words have English and coverage biases; missing words may fall outside the retained top 500. Link samples change with sampling windows.
Daily points are the last recorded measurement, not a total across the day. Current-day index observations are provisional. Trend fits are exploratory, not forecasts, causal explanations or tests of significance.