Methodology

This page explains how SIG produces and qualifies every claim. SIGreports; it does not endorse. Every material claim carries four independent epistemic fields, a support glyph, a persistent marker when contested, and an honest gap wherever SIG has not looked.

Evidence tiers and support

Sources are graded by an evidence tier (e.g. a records-request contract outranks a portal aggregator snapshot). The support glyph is a four-step mark whose fill reflects how many independent, method-distinct evidence classes stand behind the winning value, and it always carries a machine-readable downgrade reason when it sits below fully confirmed.

Reconciliation and contradiction

When sources disagree, SIG resolves by a versioned ruleset (e.g. highest tier wins) and preserves the dissent — a contested value is plotted as a range, with each competing claim's source, tier, date, and document link. Two numbers that measure different quantities are labelled as such, not forced into a disagreement.

Absence is data

A missing row is never silently omitted. SIG distinguishes four kinds of absence — not researched, no evidence found (with the sources searched named), evidence of absence, and unresolved — and each is a clickable invitation toresearch it.

What SIG does not do

SIG does not publish a capture–recapture estimate of device population from volunteer mapping and vendor portal reporting — not with a caveat, not with a wide interval. Linkage is impossible from published counts, the samples are not independent, and the estimator's known failure mode is understating the very thing SIG exists to document. SIG publishes counted quantities withnamed denominators and records-derived bounds instead — never a total. See the coverage metrics.

How well SIG resolves entities (SIG measuring its own method)

SIG resolves many overlapping observations into single entities ("this camera seen by three sources is one camera"). A project that does this must publish how well its resolver measures against ground truth — and must say plainly when that ground truth is provisional. These are the current holdout metrics.

What a "resolved site" is. An observation-level record is one source's published row for one camera. A resolved site is a cluster of those records that SIG's camera-site entity resolution judged to describe the same physical device. SIG reports N resolved sites from M observation-level records (N is never more than M) and the dedup ratio 1 − N/M. Only merges from rules whose measured holdout precision clears the published floor are applied automatically; every other candidate merge waits for human review and is not counted. Two rows from the same source are never merged, a republished copy (mirror) of the same upstream dataset is not treated as independent confirmation, and each point shown is one record's own — never coordinates combined from two sources. Deciding which of one record's own values stands (for example its latest coordinate) is notdeduplication and is never counted as a resolved site.

PROVISIONAL — the gold set is LLM-bootstrapped with a small maintainer seed and the auto-write floor is set against provisional labels (D-R6.1-EVAL, OPEN). A claim of "resolved sites" discloses this until the deferral closes.

• Cohen's κ (LLM adjudicator vs maintainer seed)0.714labelled pairs (bar ≥ 0.70 → LLM trusted as gold for the training partition)
• auto-write precision floor0.980the published floor an auto-write tier must clear on the frozen holdout
• pairwise precision / recall / F1 at tier ≤ 3 (auto-write)P 1.000 · R 1.000 · F1 1.000the frozen, human-verified holdout
• B-cubed cluster precision / recall / F1P 1.000 · R 0.952 · F1 0.976the frozen holdout clusters
• auto-write tier-0 holdout precision1.000the frozen holdout (0 tiers demoted below the floor)
• camera sites: Cohen's κ (blind LLM adjudicator vs agent seed)0.669540 double-adjudicated camera-record pairs (below the 0.70 bar, so the LLM is a suggester only and its disagreements go to human review)
• camera sites: shared-upstream-id rule (1g) holdout precision1.000 (70 of 70)the frozen, agent-verified 180-pair holdout (95% lower bound 0.948); auto-writes
• camera sites: coincident-point rule (3g) holdout precision0.986 (69 of 70)the frozen, agent-verified 180-pair holdout (95% lower bound 0.923; 0.800 if scored with the LLM's labels); auto-writes

These measure SIG's method, on a named, frozen holdout — never a device population and never a capture–recapture estimate. The full report and its reproducible generator are committed with the build.

Freshness and coverage

• Data freshness — per source: last successful run, last content change, current status, and stale-entity counts measured against predicate volatility (§32.4).
• Coverage metrics — counted quantities with named denominators, records-derived bounds, reconciliation ratios, and measured survey recall (§32.5).
• Editorial style guide and editorial standards — the register rules that govern every published sentence, including generated text.

Something wrong, or a privacy or safety concern?Dispute or correct this record — one click, no account required.