Editorial patterns

The five patterns on this page are the toolkit's repeated structural responses to recurring problems in the discovery space. They are not after-the-fact rationalisations. The patterns sit inside the tool cards as operational logic, and the cards' shape encodes the patterns even where they are not named. This page makes the patterns visible as a system so a reader can recognise the same structural response when it shows up on a new case.

The framing matters because the toolkit is small for what it covers. Sixty-five tools across twenty-two outline subcategories, in a discovery space of more than two hundred and fifty candidates, with a six-country audience whose working conditions differ sharply across the region. A toolkit at that scale either codifies its responses to recurring problems or drifts into ad-hoc judgement. The five patterns below are the codified responses. Each is anchored to a reference card the rest of the toolkit pivots against, and each carries an editorial position the cards make operational instead of declaring it.

The patterns interact. A single case typically triggers two or three of them. A reader who is new to the toolkit should read the patterns once together before opening any individual tool card; a reader who is operational should treat this page as a lookup index for "I have seen this kind of tool before and I recognise the response, but I want to check the editorial framing the toolkit applies."

Pattern 1 – Access-barrier framing (the gradient)

Some tools the toolkit covers are operationally inaccessible to the audience the toolkit is written for. Enterprise licensing in the fifty-thousand to two-hundred-thousand US dollar range per year makes the tool a procurement decision for a regional broadcaster, not a download for a frontline newsroom. IFCN-signatory gating excludes journalists structurally even where they sit inside a fact-check coalition. Institutional CSIRT deployment assumes infrastructure a typical civil-society organisation does not run. The pattern is what the toolkit does in response: it documents the tool, names the access barrier in the same breath, and routes the reader to a use the tool can serve from the other side of the barrier.

The gradient runs through four named cases. Sensity sits at institutional pricing; the card carries the cost band concretely in the Cost section and lists Sensity as primary at 1C.2 with the understanding that the realistic path for a frontline SEA newsroom is partner-mediated, not direct deployment. Graphika sits at the same price band but at escalation-option only; the card adds the frontline-no-access disclaimer across all six focus countries and substitutes a "cite the institutional research output" Quickstart for the conventional steps. Meta Content Library sits at IFCN-signatory gating; the card names journalists as structurally excluded and points to the IFCN-signatory partners (Mafindo, Rappler, Vera Files) who can run the data on a coalition's behalf. DISARM Foundation Navigator + STIX2 sits at institutional CSIRT deployment; the card frames the tool as institutional companion to the public-facing ABCDE framework, not as something a frontline desk operates directly.

The four cases together describe a gradient. Each layer of the gradient is a different kind of barrier: monetary, institutional certification, infrastructure capacity. The editorial response is the same across all four: name the barrier in the cost or how-to-access section concretely, position the tool at the right cell role (escalation-option, or primary with a partner-mediated routing), and redirect the Quickstart toward what the reader on the other side of the barrier can usefully do with the tool's output. That output is almost always cited institutional research, not first-hand deployment.

The position the pattern embeds is that access barriers are not the toolkit's failure to find affordable alternatives. They are structural features of the disinformation-defence landscape, and a toolkit that pretended these tools did not exist would mislead the reader about what institutional capacity looks like. Naming the gradient is more honest than excluding it. The reader gets a clear picture of what sits at each access tier and a clear handle on how to interact with each tier from the access position the reader actually holds.

Pattern 2 – Vendor accuracy wrapping (Sensity 3-leg)

Every detector card in the toolkit pairs the vendor's headline accuracy claim with at least one independent reference point. The pairing is mandatory under Content Policy 2 (vendor accuracy wrapping) and codified at C1 in the selection criteria. The reference instance for the pattern is the Sensity card, where the wrapping carries three legs.

The first leg is the vendor headline. Sensity markets a ninety-eight percent accuracy figure and returned that number on the Doc Willie Ong eye-drop deepfake processed through the Rappler / #FactsFirstPH pipeline, along with seventy-five and a half percent on the AI-object-generation modality and ninety-four percent on the audio modality for the same item. The number is what Sensity does on a vendor-favourable input.

The second leg is the field reading. On the Brawner "Dark Eagle" deepfake, a separate political-impersonation case from the same Philippine actor pool and through the same workflow, Sensity returned seventy-nine and three-tenths percent AI-generated. That number sits roughly twenty percentage points below the headline. The toolkit's editorial position is that the field reading is the more operationally honest reference point, because it is what the tool actually returns on contested regional political content and not on a curated vendor demonstration.

The third leg is the category baseline. The Deepfake-Eval-2024 benchmark (Lyu et al., 2025) places Sensity in the anonymised commercial pool, where individual provider scores are not separable but the best commercial video detector reached only zero point seven eight accuracy and zero point seven nine AUC on two thousand and thirty-six in-the-wild 2024 deepfakes. That number is a class-level upper bound on commercial video detector performance under field conditions. Sensity sits somewhere inside that pool, and reading the ninety-eight percent headline against the zero point seven eight pool ceiling reframes the headline as a vendor-favourable ceiling and not as a field expectation.

A fourth leg sometimes appears. Where no independent SEA-specific benchmark exists, the toolkit writes "No independent SEA-specific benchmark identified as of [date]" in the Independent Accuracy admonition and records the vendor headline for transparency. The sentence is acceptable, defensible, and more honest than improvising a number.

The three legs together produce honesty about a detector tool that a single number could not. Vendor headlines alone mislead by ceiling. Independent benchmarks alone do not reflect deployed performance because the deployment context differs from the test set. Field readings alone are too small a sample to anchor a category-level expectation. Read together, the three legs let the reader place a verdict in operational context: the verdict is in the bottom third of the vendor's claim and the upper half of the category ceiling, which is what a working desk needs to know.

The position the pattern embeds is that the detector class is not unreliable in the abstract. Detector performance is conditional on input. Vendor headlines describe one set of inputs; category baselines describe another; field readings describe a third. The wrapping discipline is what makes the conditionality legible. A detector card without the wrapping pair is editorially incomplete; the toolkit will not ship such a card.

Pattern 3 – Cite-as-external-source redirect (Graphika)

When the access barrier in Pattern 1 is structural and not monetary – a tool whose deployment requires institutional infrastructure the typical reader does not have – a different editorial response applies. The card does not invite the reader to plan deployment. The card redirects the reader toward citing the institutional research output the tool produces, as an external source the reader's own published work can reference.

The reference instance is the Graphika card, where the Quickstart section opens with a parenthetical statement that the conventional steps do not apply, then sets out an alternative four-step routine. The four steps walk the reader through: contacting an institutional partner who has Graphika access for a relevant report's methodology and dataset, confirming the data-licence terms before citing the partner's published work, identifying the Graphika commercial onboarding path for institutional partners who do hold access, and investing the Graphika-equivalent budget into open-source 1C.1 alternatives (CIB Mango Tree, Coordination Network Toolkit, CooRTweet) plus the Gephi visualisation layer and the ABCDE / DISARM codification frameworks if the reader's operation has no partner-mediated path.

The redirect pattern propagates to three other cards. Meta Content Library carries the same redirect for IFCN-signatory gating: journalists work through their IFCN-signatory partners (Mafindo, Rappler, Vera Files) for data access. DISARM Foundation Navigator + STIX2 carries the redirect for institutional CSIRT deployment: frontline readers reach the Navigator's output through partner CSIRTs or government-side reporting, not through direct STIX2 ingestion. Sensity carries a softer version: the Quickstart steps are normal but the first step is "Confirm institutional access; the realistic path for a frontline SEA newsroom is partner-mediated, not direct."

The redirect is not a hedge. The cards retain operational depth: methodology, output type, what the tool produces, what the output is good for, how the reader's published work can cite it. What changes is the action verb. The reader does not "run Graphika" or "ingest STIX2 packages." The reader cites the institutional research output, which is the operationally accessible product even when the underlying tool is not.

The position the pattern embeds is that toolkit relevance does not require deployment feasibility. A tool the reader cannot run can still be operationally useful if the reader can cite its output. Honest framing of the access boundary preserves the tool's relevance without pretending the reader has access the reader does not have. The cite-as-external-source redirect is what the toolkit ships in place of the deployment plan.

Pattern 4 – Operator-identity independence caveat (Sebenarnya)

Some tools the toolkit covers are operated by parties whose independence from regulatory or political authority is structurally questionable. The pattern is what the toolkit does in response: it documents the tool because the information environment the toolkit describes cannot be honestly described without it, embeds the operator-identity caveat descriptively across the relevant sections of the card, and routes the reader's deployment recommendations to civil-society and newsroom partners instead of to the operator.

The reference instance is the Sebenarnya.my AIFA card, where the caveat surfaces in four places. The TL;DR opens with "Government-run multilingual fact-check chatbot (MCMC) – read with independence caveat" so the reader meets the caveat before any operational detail. The Limitations admonition names MCMC as the same body that holds enforcement powers under the Communications and Multimedia Act and the Online Safety Act 2025, locating AIFA inside that institutional context. The Privacy and threat model section sets out the operator-identity threat model directly: a user forwarding content into AIFA is sending content into a government channel, not a private channel or a civil-society editorial workflow, and the consequences in the Malaysian legal environment are documented in the Malaysia country page. The In the toolkit's workflow section closes with the editorial-position statement: "the toolkit documents AIFA because the Malaysian information environment cannot be honestly described without it; the toolkit's deployment recommendations sit with civil-society and newsroom partners."

Sebenarnya is the only state-operated tool in the sixty-five-tool shortlist. The card carries a B1 exception-pass framing in the [selection criteria]: for most tools, English-only UI is operationally sufficient for the toolkit's audience of trained fact-checkers, and local-language UI is a bonus signal and not a selection requirement. AIFA is the explicit exception. Its audience is the Malaysian lay public, not trained fact-checkers, and Tamil, Malay, Mandarin, and English UI is operationally critical for that audience. The chatbot would not reach Malaysia's demographic mix without the four-language coverage. That makes AIFA the rare case where local UI is a selection criterion in its own right, not just a bonus.

The pattern does not apply only to state-operated tools. It applies to any tool whose operator's relationship to the information environment is structurally relevant to how the tool's output should be read. The shape of the pattern is consistent: name the operator in the opening line, locate the operator inside the relevant institutional or legal context, set out the operator-identity threat model in the Privacy section, and state the editorial position on deployment recommendations in the workflow section. The caveat is descriptive, not preachy. It states the structural fact about the operator and lets the reader bring the toolkit's documentation through their own editorial judgement.

The position the pattern embeds is that operator independence is part of the threat model, not separate from it. A fact-check chatbot operated by a regulator that holds enforcement powers under the same speech laws sits inside the regulator's enforcement context whether or not the chatbot's specific outputs are politically directed. Reading AIFA outputs with the same scrutiny a civil-society source-of-record claim would receive is the operational form of editorial honesty. Excluding AIFA from the toolkit on independence grounds would be the easier editorial move, but it would leave a seventy-million-view fact-check channel uncovered in a Malaysia-relevant workflow, which is the dishonest move. The caveat-with-documentation routing is what the toolkit ships in place of either inclusion-without-caveat or exclusion-without-coverage.

Pattern 5 – Honest-gap naming (Decision 7 in operation)

The toolkit ships with named structural absences. The pattern is what the toolkit does when a category of tool that would be operationally useful does not exist for a specific language, country, or use case: it names the absence as a documentary feature of the discovery space instead of masking it through a multilingual fallback or pretending the gap will be closed in some near future.

Three recurring forms surface. The first is asymmetric language coverage. Sinhala has SinLlama, MisinformationCorpusSinhala, and the SLTK toolkit, all referenced in LIRNEasia research; Tamil leans on pan-Tamil X-CLAIM with a Sri-Lanka-specific adaptation gap because X-CLAIM was trained on a pan-Indian corpus instead of a Sri Lankan Tamil corpus. The asymmetry is real. The Sri Lanka country page carries it, the 2B.2 cell carries the pin, and the X-CLAIM card carries the SL-adaptation gap binding caveat. The toolkit does not soften Tamil coverage by claiming it is "covered" through XLM-R fallback. It names the asymmetry and routes the reader to the right mitigation: pan-Tamil tooling with the SL-adaptation gap declared.

The second form is structural absence. There is no Lao-first independent fact-checker; there is no IFCN signatory in Laos; there is no Lao-language voice-clone detector under field conditions. The gap is structural and ongoing. The Laos country page carries the structural framing, and the affected cells (1A.3 audio, 2B.1 tipline, 2B.2 claim extraction) carry the cell-level pin. The toolkit's response is not to fabricate a Lao-capable alternative or to bury Lao under "Southeast Asian language coverage broadly." It is to name the absence and route the reader to workflow alternatives the absence forces: diaspora review, manual verification, regional partner routing through institutional intermediaries.

The third form is class-level unreliability. The audio detector class is broken across SEA languages – the Deepfake-Eval-2024 benchmark and the DW Innovation November 2025 audit on Hiya, Deepfake Total, and DeepFake-O-Meter together establish that. The toolkit's response is to ship the cell deliberately thin (1A.3 carries one detector; 1B.3 ceilings at three detectors plus one institutional reference) with the detector-as-weak-signal framing binding on every entry. Inflating density with vendor-only claims would violate Architectural Anchor 1. The honest move is the thin cell with the structural framing named.

The position the pattern embeds is that the toolkit is more useful when it names what it does not cover than when it pretends to cover everything. A reader who opens the toolkit looking for a Lao voice-clone detector and finds an honest gap with workflow alternatives is better served than a reader who finds an English-only detector with "limited Lao support" claimed by a vendor. The honest-gap policy is paired between Lao and Sinhala / Tamil so that Sri Lankan readers are not held to a higher coverage standard than Lao readers; the asymmetry between Sinhala (asymmetric-strong) and Tamil (pan-Tamil with SL-adaptation gap) is named with the same rigour that the Lao structural absences are named.

How the patterns interact

The five patterns are not mutually exclusive. A single tool often triggers three of them. Sensity triggers Pattern 1 (access barrier at institutional pricing), Pattern 2 (the 3-leg vendor wrapping is anchored on Sensity itself), and a soft form of Pattern 3 (the "confirm institutional access" first step in the Quickstart). Graphika triggers Pattern 1 (access barrier at enterprise-only) and Pattern 3 (the cite-as-external-source Quickstart) and arguably Pattern 5 (no documented frontline deployment in any of the six focus countries is itself a gap the card names). Sebenarnya triggers Pattern 4 (operator-identity caveat) and a corner of Pattern 5 (Malaysian Tamil through a state-operated platform highlights the absence of Sri Lankan Tamil through an independent platform).

The pattern recognition the toolkit asks of the reader is therefore additive. When a new tool enters attention – a new SEA-language deepfake detector launched in 2026, a new tipline platform from a regional civil-society coalition, a new enterprise CIB platform marketed at frontline newsrooms – the reader walks through the five patterns: which access barrier applies, what the vendor wrapping legs are, whether the tool is structurally a cite-as-external-source case, what the operator's relationship to the information environment looks like, and what gaps the tool would or would not close. The result is a structured first read on the tool before any deeper evaluation, and it is the same first read the toolkit's own selection process applies to the sixty-five tools already in the shortlist.

The patterns sit beneath the seven criterion groups of the selection framework, not above them. The criterion groups are the gate; the patterns are how the gate's decisions ship inside the tool card prose. A new tool that clears the criterion gate may still need the access-barrier framing, the vendor wrapping, the cite-as-external-source redirect, the operator-identity caveat, or the honest-gap naming, depending on what kind of tool it is. The patterns are not redundant with the criteria. They are the editorial system the criteria's decisions are rendered through.

A reader who has internalised the five patterns can read any tool card in the toolkit and recognise which patterns are firing on that card. A reader who is operational and on deadline can use this page as a shortcut: identify which pattern the case in hand triggers, follow the reference card link, and read how the toolkit applies the pattern in its anchor instance. The methodology pages exist so that the editorial system is legible as a system, not just as embedded operational logic on sixty-five separate cards.

The remaining methodology pages frame the system from other angles. How we chose tools narrates the selection process that produced the sixty-five cards. [Selection criteria] condenses the seven criterion groups as an evaluative rubric the reader can apply to tools encountered outside the toolkit. Architectural anchors sets out the three anchors that govern the toolkit's framing of the detector class and the signal architecture. Change log records the structural decisions that shaped the toolkit's current form. Read together, the five methodology pages document the toolkit's editorial system at the depth a working partner needs to extend, adapt, or contest it.