Skip to content

2B – Collaborative Verification

2B is where Pillar 2's claim of durable strategic value gets anchored. The cells here run on infrastructure that lives in messaging apps and newsroom backends; the user base does most of the verification work; AI helps the operation scale through claim extraction, multilingual processing, and LLM-assisted drafting. The intended reader is a tipline manager at MAFINDO, a Cofact Thailand coordinator on the LINE side, a Watchdog NLP operator running Dissect on Sinhala intake, a CekFakta partner editor stepping into a Sara Duterte deepfake wave. This section covers the three 2B cells, the toolkit's honest treatment of language-asymmetric coverage across Sinhala, Tamil, Lao, Bahasa, Malay, Filipino, and Thai, the framing around the toolkit's only government-operated tool (Sebenarnya AIFA), the scaling route forward to 2C Large-Scale Intelligence, and the cross-reference back to 1B.2 AI text detection for the text-generation cases that come through tipline intake.

When this tier applies

A WhatsApp Hoax Buster bot in Indonesia routes a forwarded image of a Sri Mulyani impersonation through MAFINDO's Kalimasada tipline at 2 a.m. on a Tuesday. The bot is built on Meedan Check infrastructure; the intake routes into MAFINDO's Yudistira Bahasa-specific claim database. That is 2B.1 in its routine operational form – the case study the Indonesia country page documents and the cleanest worked example of an SEA tipline operation absorbing a deepfake escalation without losing throughput.

Cofact Thailand receives a LINE-side claim about an alleged Anutin Charnvirakul recording. The Cofact backend lives in the LINE ecosystem because LINE is the dominant Thai messaging environment, which is why 2B.1 ships four tools, not one: matching platform to country is part of tipline design.

A surge of incoming claims through the CekFakta coalition turns out to be variations of the same Bahasa-language post about a politician's alleged corruption. Manual processing would mean reading hundreds of near-duplicates by hand. 2B.2 AI claim extraction is what keeps the operation functional at that volume: MAFINDO's Yudistira clusters claims in Bahasa, Watchdog's Dissect handles the Sinhala equivalent, Meedan Alegre runs the multilingual fallback when material crosses language boundaries.

A trainee fact-checker at a CekFakta partner wants an LLM that runs locally and can produce a draft summary of an unfamiliar claim before the verification work begins. 2B.3 LLMs in fact-check workflows routes to SEA-LION, the AI Singapore eleven-plus-language model that now (as of the March 2026 v4 release) explicitly includes Lao – the only LLM in the toolkit shortlist with documented Lao coverage. The trainee runs SEA-LION on the local machine, treats every output as suggestive, and never publishes anything the LLM produced without verifying every claim against 2B.2 claim extraction or 1B.5 source-reliability scoring.

How techniques in this tier connect

The 2B pipeline runs in this order on most cases: a tipline receives intake (2B.1), AI claim extraction processes the intake to surface fact-checkable claims (2B.2), and LLMs assist with drafting, summarisation, or hypothesis generation around those claims (2B.3). The 2B cells are deeply interdependent – tipline volume drives claim-extraction load, claim-extraction output feeds LLM-assisted drafting, LLM hypotheses route back into claim-extraction for verification.

The toolkit's editorial honesty about regional language asymmetry binds across all three cells. Sinhala and Tamil sit on opposite sides of the documented Sinhala/Tamil asymmetry: Sinhala has a more coherent local language-tech stack (LIRNEasia's MisinformationCorpusSinhala at 3,576 annotated documents, the SinLlama LLM, the SLTK tokenizer) that supports Dissect directly, while Tamil leans on pan-Tamil and India-based resources with an explicit Sri Lanka adaptation gap on X-CLAIM. Lao remains absent from claim-extraction tooling entirely – no Lao-language claim-extraction tool exists in the toolkit shortlist – although SEA-LION v4 now closes the Lao LLM gap that the toolkit previously paired with Sinhala under Decision 7. The toolkit names each of these asymmetries directly because softening them would misrepresent the regional information environment to the users who depend on it.

The framing on government-operated tools binds at 2B.1. Sebenarnya AIFA – the Malaysian Communications and Multimedia Commission chatbot covering Malay, English, Mandarin, and Tamil – is the only government-operated tool in the toolkit shortlist (a B1 exception-pass for public-facing lay use, codified in). The Sebenarnya card carries the independence caveat descriptively, in the operator-identity threat model paragraph and the editorial-position statement; the toolkit documents AIFA because the Malaysian information environment cannot be described without it, and the deployment recommendations sit with civil-society and newsroom partners, not with the operator. The same framing carries forward into this section's prose: government operation is one element of the threat model on a tool, named alongside the operational fact that AIFA reaches a wide public audience that civil-society alternatives do not.

The text-detection question that the 1B.2 cell houses connects here. When a tipline-routed claim arrives as text and someone wants to know whether the text itself was AI-generated, the productive answer is to treat any 1B.2 detector output (Pangram, GPTZero) as one weak signal alongside the non-detector signals that drive 2B.2 claim extraction and 1B.5 source reliability. The 1B.2 verdict is never sufficient on its own under Anchor 2; 2B.2 is where the productive verification work happens.

What this tier produces

A processed pipeline of incoming claims, deduplicated and clustered through claim extraction, with the highest-priority items surfaced for human verification at desk-tier work. The output also includes drafted summaries from LLM-assisted work where appropriate, source-reliability annotations against 1B.5 hits, and the operational record of which tipline received which claim through which platform – the routing log that downstream coalition reporting and pattern analysis depend on.

When to escalate, when to stop

Move forward to 2C Large-Scale Intelligence when the case scales from per-tipline volume to cross-platform pattern question – a claim that appears as variations across WhatsApp, Telegram, Facebook Groups, and TikTok in the same week is a 2C.1 monitoring question and a 2C.2 network-analysis question, not a 2B routing question. Move sideways into Pillar 1 verification when 2B work surfaces an artefact whose authenticity is the question – a tipline-routed video that needs 1A.2 / 1B.1 work on the video itself, an audio clip that needs 1B.3 audio forensics.

Stop when the tipline routing closes a claim through existing fact-check work – a Google Fact Check Explorer hit on the same claim, an existing CekFakta debunk, a Cofact Thailand response that has already been distributed through LINE. The T7 tipline routing tree carries the platform-by-country routing; the T5 escalation tree carries the tier-pivot logic when 2B work surfaces a case that needs Pillar 1 escalation or 2C scaling.

The cells in detail

2B.1 – Multilingual tiplines

Tipline platforms are where Pillar 2's claim of durable, strategic value is anchored. The user base does the verification work and AI helps process intake. The cell ships four tools to cover the actual messaging-app ecosystem in the focus countries.

Meedan Check is the global primary – 140,000-plus users, the CekFakta and #FactsFirstPH backbone, 34 languages including the SEA seven, with self-hostable backends for partners that need data control. Cofact Thailand is the LINE-native option for Thailand – the only LINE-side tipline in the toolkit shortlist, which matters because LINE dominates the Thai messaging environment. MAFINDO Kalimasada is the Indonesia WhatsApp Hoax Buster bot built on Meedan Check infrastructure, the reference card for the single-country tipline pattern. Sebenarnya AIFA is Malaysia's government-operated multilingual chatbot covering Malay, English, Mandarin, and Tamil – a B1 exception-pass for public-facing lay use, kept in the toolkit with the operator-identity independence caveat carried in the card.

Organisations selecting a tipline match platform to country and to threat model. WhatsApp plus Bahasa points to Kalimasada. LINE plus Thai points to Cofact. Multi-language coverage plus Tamil in Malaysia points to AIFA, with the independence caveat. Cross-country deployment with self-hostability points to Meedan Check. All four are source-history pillar non-detector platforms under Architectural Anchor 1; under Anchor 2 a tipline match by itself is one signal class and combines with claim extraction (2B.2) and CIB detection (1C.1) in standard institutional pipelines.

The cell-level honest gaps are binding. No Lao-language tipline exists in the toolkit shortlist and the section names this directly – the Laos country page is the structural-gap close that the Decision 7 honest-gap policy commits the toolkit to. Sri Lanka's tipline layer is the real gap on that side (per LIRNEasia research): forms, email, phone, WhatsApp distribution, but no documented Meedan-style purpose-built multilingual verification backend at the institutional level. The T7 tipline routing tree is the operational layer; the bridge from 2A productivity tools into 2B.1 fires when output volume warrants community routing, and the bridge upward to 2B.2 fires when intake scales beyond manual processing.

2B.2 – AI claim extraction

When tipline volume scales beyond manual processing, AI claim extraction is what keeps a fact-check operation functional. The cell ships five tools to cover SEA's language asymmetry honestly.

Dissect (Watchdog Sri Lanka / LIRNEasia) is the only Sri Lanka-specific Sinhala NLP claim tool in the toolkit, the primary entry for Sri Lankan Sinhala work, and the operational expression of the LIRNEasia language-tech stack. Meedan Alegre is the multilingual XLM-R Apache-2.0 alternative covering all six SEA languages, the productive fallback when material crosses language boundaries. Yudistira (MAFINDO) is MAFINDO's Bahasa-specific claim database, paired with Kalimasada in the Indonesian pipeline. ClaimBuster is the English baseline for global outlets working in English-language source material. X-CLAIM is the Tamil pan-Indian baseline with an explicit Sri Lanka adaptation gap noted on the card (per LIRNEasia research).

In workflow terms, Sri Lanka work routes to Dissect first for Sinhala and X-CLAIM second for Tamil (with the SL-adaptation caveat). Indonesia work routes to Yudistira first for Bahasa with Alegre as the fallback for non-Bahasa material. Cross-country work routes to Alegre because XLM-R is the only multilingual non-vendor option that covers Lao at all, with the disclaimer pass on low-resource performance. English-language source material reaches ClaimBuster.

The honest gaps the cell carries are binding. The Tamil gap on Sri Lanka is real: X-CLAIM is pan-Tamil and India-trained, not SL-adapted, and the toolkit names this on the card and at the section level (per LIRNEasia research). The Lao gap is also real – no Lao-language claim-extraction tool exists in the toolkit shortlist, and Alegre's XLM-R multilingual coverage operates at the documented low-resource performance ceiling for Lao text. Per Architectural Anchor 1 all five tools are source-history non-detector signal classes. The bridge in is from 2B.1 tipline as the scaling-up path; the bridge sideways is to 2B.3 LLMs in fact-check workflows for LLM-driven claim verification (a different workflow shape from claim extraction itself); the bridge to 1B.2 is the cautionary cross-reference noted above – text-detection signals are one weak signal alongside the load-bearing claim-extraction work that this cell performs.

2B.3 – LLMs in fact-check workflows

LLMs in fact-checking are useful, hallucinatory, and changing rapidly. The cell ships four tools that span the regional language landscape and the journalist-versus-lay-user axis.

SEA-LION is primary – the AI Singapore eleven-plus-language model, Apache-2.0 in most variants, locally runnable on adequate hardware. The v4 March 2026 release explicitly includes Lao, closing the Lao LLM gap that the toolkit previously paired with the Sinhala-on-major-SEA-LMs gap under Decision 7. AI Fact Checker App is the lay-mobile alternative documented in app-store coverage across Laos, Malaysia, the Philippines, and Thailand – a tool the trainer can point a lay user at, with the binding hallucination caveat applied. Google Pinpoint is the journalist-pipeline LLM option (cross-cell with 2A.1). MaLLaM Mesolitica is the Malay-specific escalation option for Malaysian work where Malay-first phrasing matters.

Journalists handling Bahasa, Filipino, or Thai source material run SEA-LION locally for hallucination-mitigated draft outputs, then verify every claim against 2B.2 extraction or 1B.5 reliability scoring before any publication. Lay users in the four documented countries can be pointed to AI Fact Checker App with the same binding hallucination caveat applied. Malaysian Malay-first work uses MaLLaM in escalation. Per Anchor 1 LLMs are productivity tools, not detection signals. Per Anchor 2 every LLM output is suggestive, never a signal class on its own.

The cell-level honest gap is partial closure. Lao is now covered by SEA-LION v4 – the March 2026 release is the only LLM in the toolkit shortlist with documented Lao support and fits the surveillance-environment routing the rest of the Laos stack already assumes (locally deployable, Apache-2.0). Sinhala remains absent from the major SEA LMs; the Dissect / LIRNEasia stack handles Sinhala via specialised non-LM infrastructure that 2B.2 cross-references, with SinLlama (LIRNEasia research) sitting outside this category as a non-toolkit specialist resource. The bridges in are from 2B.2 claim extraction (use an LLM to verify candidates) and from 2A.1 transcription cross-cell with Pinpoint. The bridge to 1B.2 is the cautionary cross-reference: the same hallucination behaviour that limits LLMs as fact-checkers also limits them as text detectors, so LLMs do not migrate from 2B.3 to 1B.2 as a productive escalation.

Cross-references

Sources

  • MAFINDO. Kalimasada WhatsApp Hoax Buster. MAFINDO Indonesia, 2025. mafindo.or.id. (Kalimasada / Yudistira / Satgas Pemilu Indonesia anchor for 2B.1.)
  • Sebenarnya.my / AIFA. Sebenarnya.my — Malaysia's Official Fact-Check Portal. Malaysian Communications and Multimedia Commission, 2025. sebenarnya.my. (Sebenarnya AIFA Ramadan-aid deepfake case March 2026; operator-independence caveat; B1 exception-pass for public-facing lay use.)
  • Cofact. Cofact Thailand LINE-native Fact-check Platform. 2025. cofact.org/th. (Cofact Thailand LINE-native tipline for 2B.1.)
  • LIRNEasia. MisinformationCorpusSinhala. LIRNEasia, 2025. github.com/LIRNEasia/MisinformationCorpusSinhala. (Sinhala / Tamil asymmetry; SinLlama as non-toolkit specialist resource at 2B.3; Lao tipline gap.)
  • how-we-chose-tools — B1 exception-pass for public-facing lay use applied to Sebenarnya AIFA.
  • Architectural Anchors — Anchors 1 and 2 as operationalised across the 2B cells.