Change log¶
This page tracks the structural decisions that shaped the toolkit's current form. Each entry names the decision, the reason behind it, and where the decision lands operationally inside the cards, trees, country pages, or methodology layer. A reader who finds the toolkit's framing has shifted in later use can come to this page to see what shifted and why.
The log is a living artefact. It is not a commit-message archive — the audit trail at that granularity lives in the git history. It is not a version-rationale dump either — the rationale for any single decision sits in the relevant methodology pages. The log is the chronological record of the eight structural decisions, three architectural anchors, seven content policies, three reference cards, five editorial patterns, and the source-protection aggregation layer that together define the toolkit's shape.
Outline shape locked¶
The two-pillar / three-tier outline was locked early in the toolkit's structuring work. Pillar 1 covers detecting and triaging AI-powered content; Pillar 2 covers using AI to counter disinformation. Each pillar carries three tiers: First-Line Triage, Professional Verification, Institutional-Level Analysis. The technique-level subheadings are the primary labels. An earlier user-level beginner / intermediate / advanced labelling proposal was dropped. The reason: tier progression is carried operationally by the technique-level subheadings and by cell ordering within each pillar (1A → 1B → 1C, 2A → 2B → 2C). A user-level label would imply that a reader stays at one tier throughout the toolkit, when in operational practice a single reader moves across tiers depending on the case in hand. Where it lands: the pillar-1 narrative, the pillar-2 narrative, and the tool-card pivot view all use the technique-level subheadings as primary labels.
Three architectural anchors locked¶
The three anchors were decided after a strategic review of a draft tool inventory containing substantial detector-class density. Anchor 1: the toolkit is not a directory of AI detectors, the workflow runs on four pillars. Anchor 2: at least two non-detector signals are required for any strong public claim. Anchor 3: detector plus detector is one signal class. The reason: a toolkit that lists detectors as the operational answer to AI-powered disinformation will be used as if they were the operational answer, and the field evidence on detector reliability does not support that use. Where they land: the architectural-anchors page sets them out as embedded operational logic; every tool card identifies pillar service and signal class in YAML frontmatter and in the workflow section; the decision trees' publish nodes enforce Anchor 2 at terminal states; the T5 escalation tree carries the Anchor-3-reset branch.
Eight foundational decisions locked¶
The first six decisions were locked alongside the architectural anchors. Decision 1: canonical five-thirty-one-hundred-twenty-minute time-band ladder. Decision 2: ABCDE / DISARM split between public-facing surface vocabulary and institutional annex. Decision 3: inter-tool conflict resolution lives in per-card fields plus a dedicated T5 node. Decision 4: regional suffix scheme for countries and platforms. Decision 5: consolidated gap catalogue as single gap source. Decision 6: detector skepticism hard-coded at architectural, editorial-policy, and operational-override levels. Decisions 7 and 8 were added to address audience-fit and density questions that the first six did not resolve. Where they land: the how-we-chose-tools page and the selection-criteria page carry the operational expression; the tool-card pivot view carries the cell-level density rules; every tool card carries the regional suffix scheme in the Country and platform applicability section.
Decision 7 added¶
The blanket "tool must support a SEA language" rule that earlier scoping work had initially proposed was replaced with Decision 7. Language coverage applies conditionally to what the tool processes: UI accessibility is English-sufficient for trained fact-checkers (with one named exception, Sebenarnya AIFA, for public-facing lay use); content-language coverage is mandatory only for NLP-class tools whose core function is to process natural language; image, video, provenance, network, and metadata tools are language-agnostic; documented regional use is a bonus, not a gate; the honest-gap policy applies to content the reader needs to verify, not to tools' UI. The reason: a trained SEA fact-checker reads English as a professional language and the rule's blanket form was excluding language-agnostic tools (CIB Mango Tree, ExifTool, GeoSpy) without operational justification. Where it lands: the selection-criteria page Group B; the editorial-patterns page Pattern 5; the cell-level honest-gap pins at 1A.3, 1B.3, 2B.1, 2B.2, 2B.3.
Decision 8 added¶
The "12-15 tools total" figure that had appeared in earlier scoping references was superseded by Decision 8: density target of two to three tools per outline subcategory by default, three to four where the landscape is rich, one to two where the landscape is thin. Resulting target range of sixty-five to eighty-five tools across the twenty-two cells. The reason: the earlier figure was an inertial carry-over from a much smaller inventory; the inventory grew to two hundred and fifty-nine candidates, which made the smaller figure operationally indefensible. The density target by cell-level richness is what Decision 8 specifies in its place. Where it lands: the resulting shortlist of sixty-five tools across the twenty-two cells; the tool-card pivot view carries the cell-level density per subcategory.
Five content policies locked¶
Five content policies were derived from the candidate inventory: honest-gap policy for Lao and Sinhala / Tamil; vendor accuracy wrapping rule (any accuracy claim appears as a pair); limitation faithfulness rule (verbatim from source); SEA-language coverage mandatory disclaimer (paired with Decision 7's conditionality); reliable-triad emphasis (provenance, tipline, manual verification carry the durable load, not detectors). The reason: earlier scoping work surfaced recurring honesty failures in vendor documentation (unwrapped accuracy claims, softened limitations) that would propagate into the toolkit if not codified as policies. Where they land: every tool card's Independent Accuracy admonition for C1; every tool card's Limitations admonition for C2; the pillar-1 narrative and pillar-2 narrative for the reliable-triad emphasis.
Three reference cards established¶
Tool-card writing locked the voice register against three reference cards. InVID-WeVerify is the multi-cell, multi-pillar deployable tool exemplar; the card carries the source-history pillar primary role plus the cautious-detector pillar secondary role across three cells. MAFINDO Kalimasada is the single-country tipline exemplar at 2B.1; the card carries the country-anchored tipline pattern used by the four 2B.1 cards. TRIED Benchmark (WITNESS) carries the institutional-reference pattern where the entry is a framework or benchmark applied to other tools, not a deployable tool itself. The reason: sixty-five cards is documentation at scale, and without locked diff targets the voice register drifts. Where they land: the three reference cards are the canonical diff targets for voice register.
Five editorial patterns codified¶
Five editorial patterns the cards apply on a recurring basis are codified explicitly in the methodology layer. Access-barrier framing (Sensity → Graphika → Meta Content Library → DISARM Foundation Navigator gradient); vendor accuracy wrapping (Sensity 3-leg as anchor); cite-as-external-source redirect (Graphika as anchor); operator-identity independence caveat (Sebenarnya AIFA as anchor); honest-gap naming (Decision 7 in operation). The patterns were not declared in advance; they emerged from recurrence across the cards and were codified once the methodology layer was scoped. Where they land: the editorial-patterns page carries the codification; every card that triggers a pattern carries the pattern in operation.
Source-protection layer (S1-S10) aggregated¶
The ten S-classes (S1 source-identifying upload risk, S2 state-linked / legally sensitive investigation, S3 graphic / sexual / child-safety material, S4 doxxing / harassment / vulnerable-community targeting, S5 private / encrypted group collection, S6 detector-only accusation, S7 malware / APK / phishing / ID-harvesting, S8 liar's-dividend risk, S9 cross-border data transfer / vendor retention, S10 staff safety and trauma) were aggregated across the T6 decision tree work and the Digital Safety section. The work consolidated these into one editorial system. The reason: source protection is editorial policy, not procedural advice, and a toolkit that distributes the system across forty-plus tool cards without an aggregation layer fails to make the editorial position legible. Where it lands: the T6 source-protection tree for in-case routing; the source-protection-aggregation page for the pedagogical depth; tool cards render the relevant S-class admonition where the tool's default use would trigger the class.
Decision 7 honest-gap policy applied – country pages¶
The honest-gap policy was applied concretely in the country pages on the Sinhala / Tamil asymmetry and the Lao structural absences. The Sri Lanka page carries the asymmetric framing: Sinhala has SinLlama, MisinformationCorpusSinhala, and the SLTK toolkit (asymmetric-strong); Tamil leans on pan-Tamil X-CLAIM with the Sri-Lanka-specific adaptation gap. The Laos page carries the structural-absence framing: no Lao-first independent fact-checker, no IFCN signatory, no Lao-language voice-clone detector under field conditions. The reason: country pages are where the honest-gap framing is operationally tested; a country page that buried the gaps under "broad Southeast Asian coverage" would have failed the policy. Where it lands: country pages, cell-level honest-gap pins, X-CLAIM card SL-adaptation gap binding caveat.
Methodology layer¶
The five methodology pages – editorial-patterns, how-we-chose-tools, selection-criteria, architectural-anchors, and this change log – surface the editorial system as a system. The pages do not duplicate substance from the operational sections; they make the editorial system visible as a system, complementing the operational logic the tool cards and decision trees embed. The reason: a toolkit whose editorial system is invisible to the reader is harder to extend, contest, or adapt to a different regional context; the methodology layer provides the operational handle for those uses. Where it lands: the methodology index routes external readers to the right starting point.
How to add to this log¶
New entries follow the same shape as the existing entries: decision named, reason, and where it lands operationally. Where-it-lands references include cross-links to the affected cards, trees, country pages, or methodology files. A decision is recorded as it is taken, not reconstructed after the fact.