Methodology

How evidence, reading signals and ranking work

This page discloses the evidence rules of the Intelligence site, exactly what the reading telemetry does and does not collect, the invalid-traffic filters, and the heat formula that is not yet active. Any change is versioned here.

Version v0-preview · Dynamic ranking: not active

1. Evidence ledger

Every claim comes from a published briefing and records the quote, source link, publisher, date, evidence type and confidence. Extraction runs on InsightBridge's own compute with a local model and is independently re-checked by a second model; links and quotes must match the original text verbatim or the claim does not enter the ledger. Disputed or superseded entries stay visible and are labelled, never deleted.

Topic briefs are drafted by a model and every paragraph must cite ledger entries; a brief is published only after editorial sign-off, and all versions are retained.

2. Source levels

L0InsightBridge original content
L1Alliance members / licensed contributions
L2Public full-text feeds that do not forbid ai-input
L3Public web pages: metadata and short excerpt only
L4Crawling forbidden: title and link cited manually

We honour robots.txt, Content Signals, RSL and TDM declarations. Our own declarations: robots.txt · llms.txt

3. Reading signal: what is and is not collected

Collected

engaged_seconds
max_scroll_depth
content_length
session_id (random, per tab session)
country (edge header)
network ASN (edge header)
user-agent class
Turnstile result

Never collected

mouse coordinates
text selection
clipboard
raw IP address (only a salted hash)
cross-site identifiers

Events are deleted automatically after 180 days. Nothing is collected when the browser sends Do-Not-Track or Global Privacy Control.

4. Invalid-traffic filter (GIVT)

Known crawler / automation user-agent
Reading speed above 120 characters per second with near-complete scroll
More than 60 events from one session within 10 minutes
Scrolled to the end in under 3 seconds
More than 120 events from one network address within 10 minutes

Filtered events are kept for filter-rate statistics but never enter reading metrics. Sophisticated invalid-traffic (SIVT) detection follows once enough data has accumulated.

5. Heat formula (preview, not active)

deep_read: engaged_seconds / expected_seconds(content_length, lang) >= theta_c AND max_scroll_depth >= 0.7
p_lb: Wilson 95% lower bound of deep_reads / valid_views
base: p_lb * ln(1 + valid_views)
heat: base / (age_hours + 2) ^ g_c
exploration_slots: 10-20% of each dynamic block reserved for new items
status: not active — thresholds (theta_c, g_c) are calibrated after 4-6 weeks of data

Thresholds are calibrated after 4–6 weeks of data; this page is updated before the formula goes live.

6. Sponsorship and commercial content

Paid placements are always labelled as sponsored, are never mixed into evidence ledgers or briefs, and never influence ranking or search.

← Back to topics