Where the evidence comes from.

PharoSec connects accessible public observations and keeps a record of their source, date and limitations. This directory distinguishes running connectors from research and access requirements.

Source review: 8 October 2026. Current feed status · Download the source inventory

A public webpage is not automatically an open dataset. No service can ingest every public post or news report: sources have different rights, access controls, languages, deletion rules and geographic coverage. The useful measure is traceable evidence and disclosed gaps.

Reporting

Social

Satellite

Context

Maritime

Verification

Territory

Territorial control

From a report to an assessment

A source record says what was observed or reported. A claim groups records that refer to the same proposed event. An assessment records what a reviewer can establish, what remains disputed and when the assessment expires. Collection and corroboration are separate steps.

1. Preserve provenance. Record provider IDs, original URLs, capture/publication/collection times, language, licence and retrieval health. Keep original text separate from translations; preserve lawful originals privately with checksums and a revision history.
2. Check source independence. Separate original witnesses, wire syndication, reposts and copied narratives. Thirty-two posts can originate from one video. Counts should expose record count, publisher groups and what is known about underlying evidence independence.
3. Verify media context. Check lawful originals for capture metadata, earlier reuse, keyframes, language and visible contextual clues. Compare dates and geographic hypotheses against maps, weather and satellite acquisition records. Similarity and geolocation suggestions require human verification.
4. Review change. Co-register comparable observations, mask clouds/shadows and compare suitable baselines. Optical colour differences, radar geometry, vegetation, flooding, construction and lighting can all resemble damage. A future change detector should produce review candidates with quality checks, never an automatic combat label.
5. Publish dated assessments. City-level control assessments need at least two distinct publisher/source groups, a review note, supporting and opposing evidence, an as-of date and expiry. Current publication applies a minimum 48-hour delay. Unreviewed headline leads remain visible separately and do not assign control.

The map currently provides image comparisons, thermal observations and catalogue records. It does not run automatic damage classification, image geolocation or military tracking. Reported city assessments are linked to their underlying sources. Missing coverage is explicit; no plausible-looking territorial boundaries are generated. Two publisher groups can share underlying witnesses or agency reporting and are not guaranteed independent confirmations.

Where Laya fits

The supplied Laya model is Apache-2.0 text classification. It is a candidate for routing reports by topic, identifying statements that need review, and comparing an attributed claim with an attributed denial. It cannot inspect an image or verify an event against reality.

Its model card documents substantial zero-shot limitations. Before activation, build an analyst-labelled evaluation set for Arabic, Amharic, Tigrinya, Somali, Oromo, English and French, including negation, hearsay, historical reporting, transliteration, satire and syndicated copies. Split by event and source so near-duplicates cannot leak across train and test sets. Measure class-level precision/recall, confidence calibration and abstention; publish version and error rates. A model score describes a classification, not a probability that a village was captured.

Start with a private review assistant and a pinned model/runtime. Evaluate latency and memory on a separate worker before enabling it. No Laya inference is running on the public service yet. Training, evaluation and inference all require appropriate rights to the input data.

What the next version requires

1. Approved source access. Obtain a ReliefWeb V2 appname, resolve documented social-platform access, agree publisher permissions and secure commercial redistribution rights where required. Record each grant, rate budget and deletion obligation in the source registry.
2. Evidence storage and review. Add a database of observations, claims, independent source groups and assessments; a private reviewer workspace; multilingual entity resolution; correction logs; and moderated contributor submissions. Keep personal identity and source-protection data outside the public API.
3. Remote-sensing evaluation. Add quality-masked spectral products and terrain-corrected radar. Benchmark change proposals against reviewed historical cases and non-conflict changes. Store model/version, baseline, thresholds and analyst outcome. Higher-resolution purchases need an archive/tasking and distribution budget.
4. Public intelligence delivery. Publish evidence-linked, dated, coarse geographic assessments and corrections. Add user-selected time windows, place/category search, uncertainty filters and a timeline. Publish coverage and lag per source instead of promising continuous visibility.
5. Commercial API operations. Implement authenticated access, source-specific response permissions, pagination, versioned schemas, quotas, webhooks, audit logs, availability monitoring and a service-level agreement. A purchased feed’s raw records cannot automatically be resold through an API.

A claim response should include claim_id, claim_text, place_id, event_time_range, status, record_count, independent_evidence_groups, supporting_evidence, opposing_evidence, reviewed_at, as_of, expires_at and revision. Each evidence item needs its own timestamp, source, reuse/independence group, method and rights. These fields describe the target claim schema. The current public control endpoint contains public-source city assessments, history, source groups, synthesis dates, expiry and unreviewed headline leads. A private reviewer workspace and contributor submissions are planned.

Rights and privacy

A free public map and a paid API can have different redistribution rights. Review copyright, database rights, platform contracts, privacy, defamation, sanctions/export restrictions and consumer representations for the operator’s jurisdiction and intended markets. Global users do not identify where the company is established.

Public social posts can still contain personal data or sensitive political information. Where UK/EU rules apply, assess a lawful basis, necessity, transparency, minimisation, retention, deletion requests, international transfers and whether a DPIA or special-category condition is required. A journalism exemption cannot simply be assumed for investor analytics. See the ICO’s legitimate-interests guidance and the applicable regulator’s guidance.

Maintain a correction and removal route, protect witnesses and contributors, and avoid publishing personal location trails or operational military details. State-source claims and opposition claims should be attributed consistently; source quantity and automated classification cannot guarantee freedom from propaganda. This source review is product research, not a determination of legal compliance in every country.

Faction polygons for Sudan, Ethiopia and Yemen use attributed community sources, separately labelled from city assessments. The revision date does not establish current ground truth throughout the map. Cartographic method, rights and limitations.

Read the implementation plan and this release’s delivered features.