A field guide to
four connected worlds_

01 / OPEN THE NOTEBOOK

Your first reading

Send a GET request to https://braigetori.org/api/observatory. The default is seven days of hourly observations for all four worlds. The archive began September 12, 2026; choosing a longer window does not create earlier measurements.

curl --fail --get 'https://braigetori.org/api/observatory' \
  --data-urlencode 'view=worlds' \
  --data-urlencode 'world=all' \
  --data-urlencode 'days=7'

Choose “Take a reading” to query the public archive.

02 / PICK AN INSTRUMENT

Eleven datasets, different questions

Always inspect truncated. A limit is a limit, not evidence that the underlying world has no more records. Directory paging is the way to retrieve the full latest server directory.

03 / TURN THE DIALS

The query vocabulary

Supply each parameter once. Unknown parameters, invalid values and repeated parameters return HTTP 400. URL-encode values. All dates and archive windows use UTC.

Parameters outside their dataset’s scope have no effect. For example, resolution does not turn daily server history into hourly history. A history request requires both an exact host and a single world.

04 / READ THE MARKINGS

Response anatomy

{
  "version": 1,
  "generated": 1789225200,
  "filters": {"view": "worlds", "world": "all", "days": 7, "…": "…"},
  "status": {"last_success": 1789224900, "failed_worlds": [], "…": "…"},
  "metrics": {"indexed": ["Indexed URLs", "Definition…"], "…": "…"},
  "rows": [{"at": 1789224900, "world": "gemini", "indexed": 1234, "…": "…"}],
  "count": 1,
  "truncated": false,
  "retention": {"hourly_days": 90, "daily_world_days": 730, "…": "…"},
  "notes": ["Dataset-specific methods and limitations…"]
}

Illustrative, abbreviated structure above; 1,234 is an example, not a live count. The live reader shows actual values.

generated vs. at

Unix seconds: report generation time versus observation time. Cached reports can have an older generation time. Use row at for analysis.

status

Last attempt/success, failed worlds, synthetic marker, daily store-check time and outcome. A null integrity outcome means unknown. Latest successful data can remain available after a failed round.

indexed, distinct, duplicates, unknown_hashes

Searchable URL count; distinct known content hashes within a world; additional copies of those hashes; URLs without hashes. The last three reconcile to indexed URLs. They are not counts of authors.

destinations, known, queue

Indexed server addresses; known resources including unindexed ones; pending Gemini queue entries or due resources in other worlds. Queue populations differ across protocols.

fetched_24h, changed_24h, errors

Resources successfully fetched within 24 hours; indexed resources whose last observed content change falls in 24 hours; known resources currently carrying a failure. These are not visit counts or publication dates.

duplicate_pct, hash_coverage_pct, error_pct, fresh_pct

Duplicate/indexed, known-hash/indexed, failures/known and fetched/known percentages. A zero denominator yields null, not zero.

World rows

at, world, group, method_version plus defined metrics. Older method 1 observations precede the added reconciliation checks; method 2 records validated collection.

Server rows

at, world, host, group and available count metrics. The JSON metric dictionary defines measures; a field may be absent or null when unavailable.

Topic rows

at, day, world, word, n, sample, rate, group. Count n is title-document occurrences; sample is the sampled title denominator; rate is 100 × n/sample.

Link rows

at, day, world, source, target, n, group. Source is a host; target includes the destination protocol. Count n is sampled links, not traffic.

Demand rows

day, selection, channel, word, outcome, n, at, group. Selections can overlap. Bots and repeats can contribute; do not sum cells into people.

Quality rows

at, world, ok, seconds, reason. Failed rounds retain previous measurements. Reasons are bounded public categories.

05 / SMALL EXPEDITIONS

Questions worth taking outside

Download a full server directory with Python

This standard-library example reads all pages and checks their collection round and keys. If collection changes during paging, start again after the round completes.

import json
from urllib.parse import urlencode
from urllib.request import urlopen

base = "https://braigetori.org/api/observatory?"
rows, rounds, totals = [], set(), set()
for page in range(1000):
    query = urlencode(dict(view="places", order="host", page=page))
    with urlopen(base + query, timeout=20) as response:
        report = json.load(response)
    rows.extend(report["rows"])
    rounds.add(report["status"]["last_success"])
    totals.add(report["total"])
    if len(rows) >= report["total"]:
        break
assert len(rounds) == len(totals) == 1, "Collection changed; retry"
assert len({(r["world"], r["host"]) for r in rows}) == len(rows)
assert len(rows) == totals.pop(), "Directory incomplete"
with open("braigetori-servers.json", "w") as output:
    json.dump(rows, output, indent=2)

For analysis, join on world and observation time, preserve missing values, and inspect collection gaps before calculating differences. The observatory laboratory offers regression, Pearson/Spearman correlation, smoothing and exports. Fits describe the observations; they do not establish causation or predict community growth.

07 / A CLEARER VIEW OF SCREENING

Safety and local AI from your terminal

These instruments measure the screening system without publishing flagged documents. No flagged URLs, hosts, snippets, hashes, images, private notes, prompts or visitor identifiers appear in these four datasets. Labels are estimates and policy decisions, not findings of illegality.

curl --fail --get 'https://braigetori.org/api/observatory' \
  --data-urlencode 'view=safety' \
  --data-urlencode 'world=gemini' \
  --data-urlencode 'level=warn' \
  --data-urlencode 'group_by=reason' \
  --data-urlencode 'days=30' \
  --data-urlencode 'resolution=day' > warnings.json

curl --fail 'https://braigetori.org/api/observatory?view=pipeline&days=7' > review-queue.json
curl --fail 'https://braigetori.org/api/observatory?view=models&model=qwen35&days=30' > local-ai.json
curl --fail 'https://braigetori.org/api/observatory?view=models&model=mobilenet&days=30' > image-review.json
curl --fail 'https://braigetori.org/api/observatory?view=actions&days=90' > safety-actions.json

# Text model by the hour, its verdicts by day, and its daily exam (site-wide; world=all)
curl --fail 'https://braigetori.org/api/observatory?view=cadence&days=7' > model-hours.json
curl --fail 'https://braigetori.org/api/observatory?view=verdicts&days=30' > model-verdicts.json
curl --fail 'https://braigetori.org/api/observatory?view=exams&days=90' > model-exams.json

cadence has one row per UTC hour with a measured worker run (90 days). verdicts counts what the text model concluded about documents it newly judged each UTC day: safety label before rules and precedence, usefulness, library shelf and time per document (180 days). exams gives each day's score on six authored, non-graphic cases graded as search would label them, with per-case results in exam_cases (730 days). These views contain labels and counts only: no addresses, titles or text.

Keep the units attached

Safety n and share_pct describe matching retained URL records; population is all tracked safety URLs in that world. It is not a count of people, a census of the index, internet prevalence or model accuracy. Hourly and daily points are the last observed snapshots, not sums. A zero category is meaningful only within a successfully observed population; missing snapshots stay unknown.

Pipeline fields include review_coverage_pct, model_uncertain_urls, stale_urls, review_queue, queue_oldest_hours and queued_characters. The queue is bounded to 2,000 waiting fingerprints, excludes in-flight batches and expires old entries; its size is not the full review backlog.

Text model rows contain additive daily run, document, error, inference and reported-token totals, plus elapsed/CPU time. Take the maximum of peak_model_mib across outcomes, never its sum. token_coverage_pct reports how many attempted calls supplied token accounting; absent accounting is null, not invented tokens. models describes installed models, runtime, role, version, licence, schedule, configured resource limits and known limitations. Usage and action rows are site-wide and require world=all; protocol attribution is not inferred.

Image counts cover only server review operations. Completed UTC outcome cells are frozen once and published only at five or more operations, without totals exposing hidden remainders. There is no private-browser telemetry. Empty image reports can mean suppressed, opted-out, unavailable or not-yet-collected data. New text usage instrumentation does not reconstruct earlier work.

Every response includes metric definitions, filter values, notes and collection status. Preserve those beside exported rows. Use the Safety & AI workspace to chart, transform, compare and export the same records. The repository also includes a dependency-free deploy/observatory-cli.py wrapper for JSON or CSV:

python3 deploy/observatory-cli.py --view safety --level warn --group-by reason --resolution day --days 30
python3 deploy/observatory-cli.py --view models --model qwen35 --days 30 --format csv > model-work.csv

06 / KNOW YOUR INSTRUMENT

Limits that belong beside the data

• Hourly world observations: 90 days. Daily world summaries: 730 days. Daily servers, topics and links: 90 days. Published demand: 30 days. Daily points are last measured values, not daily sums; today is provisional.
• Worlds are collected separately, not in one simultaneous source-database snapshot. A report has a consistent analytics read, but separate requests can span collection rounds.
• Titles use a nonrandom, known-hash-deduplicated sample of at most 5,120 titles/world from 16 ID windows. The collector retains the top 500 words/day; the unfiltered public topic report returns the top 100. An absent word is unknown, not zero.
• Links are sampled from active Gemini links or saved cross-world leads. They are not the complete graph and do not measure clicks.
• Search signals are frozen after each UTC day closes, for reviewed-word cells with at least five received events. Empty reports do not mean nobody searched. No raw queries, visitor identities or private totals are published.
• Each response is limited to 4 MiB. SQLite reads have a two-second execution deadline; at most two reports execute concurrently. An eight-report cache lasts 30 seconds. Reuse downloaded data and avoid rapid polling.

The endpoint supports public GET reports, not arbitrary SQL or administrative commands. Browser access is same-origin; no cross-origin browser permission is promised. Scripts and notebooks can use ordinary HTTPS. Save the returned filters, metric definitions and notes alongside your rows so the evidence stays interpretable.

Privacy policy · Return to the instruments · Questions / corrections