The last mile of any check: raw results from N sources → one coherent answer you can trust, with an explicit confidence score and discrepancies shown.
A single source sounds confident and often lies. An aggregator returns a neat card with precise numbers — and you believe it because the number is specific. But specific ≠ correct: a neighboring aggregator has a different number for the same company, and you don't see it until you open both. This engine opens all of them at once and consolidates — not silently picking one version, but showing where the sources agree (that is the fact) and where they diverge (that is a flag).
Principle (POV), proven live on 15.06.2026. We ran one company through 5 aggregators: the number of court cases diverged by almost threefold (roughly ~500 / ~1000 / ~1500 — different methodology), while revenue and FSSP matched across all of them. The engine's conclusion: one source = a single point of failure and false precision. Fact = agreement across ≥3; discrepancy = spread, not the choice of one number. Matched → confidence; diverged → flag. (Numbers are rounded; verify yourself.)
Source authority scale (Russia)
| Level | Sources | Weight |
|---|---|---|
| Highest — primary source | State registries: EGRUL/EGRIP FNS, GIR BO, kad.arbitr, FSSP, EFRSUB/Fedresurs, regulations (NK/GK/TK), official documentation | maximum |
| High — official summary | FNS Transparent Business, Bank of Russia/Roskomnadzor registries, peer-reviewed articles, primary reporting | high |
| Moderate — aggregator (verified) | checko, rusprofile, saby/SBIS, audit-it — collect from official sources but add their own methodology/scoring | medium, cross-verify |
| Low-moderate — aggregator (noisy) | list-org (mixes in third-party signals), industry portals, expert blogs | verify data ownership |
| Low — unconfirmed | forums, reviews, marketing, news aggregators, social networks | only as a hypothesis |
| Unverified — according to someone's words | data that the user provided as truth | mark as "not verified" |
Confidence score = freshness × authority × agreement
Freshness: every fact has a source date. Fresh (today/week) → high confidence for status data; outdated (months) → lower it, mark as "possibly outdated". Example from practice: audit-it returned 2023 when others returned 2024 — take the most recent year.
Authority: according to the table above. State registry > aggregator > forum.
Agreement (the main thing in cross-checking):
- Matched across ≥3 sources → ✅ high confidence, state it as a direct assertion.
- Diverged (different figures/estimates) → ⚠️ show the spread and do not present a single figure as fact. Example: the number of court cases for one company — approximately ~500 / ~1000 / ~1500 across different aggregators (different counting methodologies) → "according to different sources, from ~500 to ~1500 cases, counting methodologies differ".
- Signal only in one source → verify ownership (especially alarming: bankruptcy, court, debt — list-org sometimes has third-party ones). Do not scare until you have confirmed it.
Deduplication
The same information is often in several places. Merge:
- Duplicate signals: the same text/entity (INN, OGRN, name), close dates, one source references another.
- Priority when merging: completeness (the most complete version) > authority (state registry > aggregator) > freshness (latest update).
- Do NOT merge: different conclusions on the same topic, different points of view, data evolution (v1 vs v2), different periods.
Conflict — show it, do not hide it
When sources diverge — state it explicitly, don't choose silently:
Расхождение источников по [параметр]:
- [источник A, дата, tier]: значение X
- [источник B, дата, tier]: значение Y
Вероятная причина: [разная методология / устарело / чужие данные].
Наиболее надёжно: [вывод по авторитету+свежести] — но проверьте, если критично.
Attribution — every fact to its source
Every claim → source + date + tier marker (✅ confirmed Tier 1-2 / ⚠️ partially / ❌ not confirmed). At the end — a list of sources. No numbers without provenance.
How it works
[Запрос] → определи тип (факт / статус / сущность / решение)
↓
[Декомпозиция] → какие источники релевантны, что у каждого спросить (параллельно)
↓
[Сбор] → собери из ≥3 источников (агрегаторы/реестры/веб)
↓
[Дедуп] → слей дубли по приоритету полнота>авторитет>свежесть
↓
[Уверенность] → свежесть × авторитет × согласие на каждый факт
↓
[Синтез] → связный ответ, ведёт вывод (не методология), конфликты явно, атрибуция + tier
Anti-patterns
- DO NOT list by source ("at checko… at list-org…") — consolidate into a narrative by facts.
- DO NOT give a single figure when sources diverge — show the range.
- DO NOT give a single signal the same weight as one confirmed by ≥3.
- DO NOT hide the conflict. DO NOT omit attribution. DO NOT scare with an unverified signal.
Output format
Unified killer-card format (shared across counterparty-guard / tax-calendar-proactive / cross-source-verify): header
{эмодзи} {что} — на {дата}→ verdict (✅ consistent / ⚠️ discrepancy) → facts on the merits → footerЧто проверено:(source + date + tier) andЧто НЕ проверено:. One recognizable look across all stacks (markdown, without host-specific markup).
Example (illustrative)
Query: "How many court cases does this counterparty have?" Compiled from 4 aggregators.
Bad (single source): "~1500 court cases" — we take the highest counter, it sounds precise. In reality it mixed in other parties' episodes, and everyone's methodology differs.
Good (cross-check):
🔎 Сверка: контрагент {ИНН} — на 15.06.2026
⚠️ Расхождение по числу судов — единой цифре верить нельзя.
Суды: по разным источникам от ~500 до ~1500 дел — методики подсчёта различаются
(примерно ~500 / ~1000 / ~1500 у трёх агрегаторов). Надёжнее ориентироваться
на нижнюю-среднюю оценку; точное число — в kad.arbitr.
ФССП: совпало у всех источников — ✅ высокая уверенность.
Выручка: совпала у всех — ✅ Tier 1 (ГИР БО).
Что проверено: 4 агрегатора + ГИР БО (15.06.2026); ФССП/выручка — ✅ совпали ≥3.
Что НЕ проверено: точное число судов — только kad.arbitr вручную.
The difference: instead of a falsely precise figure — an honest range where sources disagree, and a firm statement where they match.
Who uses the engine
- counterparty-guard — 5 aggregators per counterparty (the main consumer, it originated from here).
- competitor-scout — collection on competitors from different platforms.
- fact-check of any skill — numbers/quotes with tier (by default from CLAUDE.md).
- Any skill where >1 source is consolidated into an answer.
Based on Anthropic enterprise-search/knowledge-synthesis, adapted for RF sources and the tandem's tier protocol.
Changelog
- 0.2.1 (16.06.2026) — unified killer-card format (header with date → verdict agreed/discrepancy → facts → footer What was checked/NOT checked), shared with counterparty-guard and tax-calendar-proactive; the example is wrapped in a card.
- 0.2 (15.06.2026) — packaging: hero about the false precision of a single source, POV proof moved to the top (live test of 5 aggregators), worked example "bad/good".
- 0.1 — circular reconciliation engine: RF authority scale, confidence = freshness × authority × agreement, dedup, explicit display of conflicts, attribution with tier.