📄
Total Sources?
Total Sources — All documents in the repository across all confidence tiers. Includes peer-reviewed journals, grey literature, reports, datasets, and theses.
0
↑ 2,418 this week
✓
Annex 2 Compliant?
Annex 2 Compliant — Sources that passed all 9 IPCC Procedures Annex 2 criteria. Only these can be cited in AR7 chapters. Currently 94.2% pass rate.
0
↑ 94.2% pass rate
🔬
Case Studies?
Case Studies — Structured capsules generated by the ANALYST agent. Each has objectives, methods, key outcomes, quantitative findings, and limitations ready for chapter authors.
0
↑ 124 new this month
📊
Chapter Coverage?
Chapter Coverage — Percentage of AR7 chapter sections with ≥1 high-confidence African source routed to them. Target: raise from current ~18% to 20–25%.
0%
↑ 3.2pp since last sprint
Ingestion Pipeline
LIVE · 3 sources active
Crawl
✓ Done
SCOUT Agent
Automatically discovers and fetches documents from 6 source tiers — UNFCCC, African journals, UN bodies, national ministries and more. Runs OCR on scanned PDFs.
Automatically discovers and fetches documents from 6 source tiers — UNFCCC, African journals, UN bodies, national ministries and more. Runs OCR on scanned PDFs.
›
Vetter
◉ Active
VETTER Agent
Checks every source against all 9 IPCC Annex 2 criteria. Screens for predatory publishers. Only compliant sources enter the repository — this gate cannot be bypassed.
Checks every source against all 9 IPCC Annex 2 criteria. Screens for predatory publishers. Only compliant sources enter the repository — this gate cannot be bypassed.
›
Structure
○ Queue
STRUCTURER Agent
Builds SDMX data structures, populates the RDF knowledge graph, classifies sources by WG chapter and sector, and generates vector embeddings for semantic search.
Builds SDMX data structures, populates the RDF knowledge graph, classifies sources by WG chapter and sector, and generates vector embeddings for semantic search.
›
Analyse
○ Queue
ANALYST Agent
Scores each source using the 5-component confidence formula, applies IPCC calibrated uncertainty language, and generates structured traceable accounts ready for chapter authors.
Scores each source using the 5-component confidence formula, applies IPCC calibrated uncertainty language, and generates structured traceable accounts ready for chapter authors.
Live crawl
0 docs
UNFCCC / BTRs
OpenAlex Africa
AJOL Journals
AfDB Portal
Recent Activity
Last 24 hours
2,418 sources ingested from UNFCCC transparency corpus
14 min ago
Annex 2 gate passed 94.2% — 142 flagged for human review
1h 22m ago
WGII Ch.9 Africa gap map updated — 31 new case studies added
3h ago
7 sources rejected — predatory publisher detected (Beall criteria)
5h ago
ILK batch of 18 sources routed to Sub-Regional Centre panel
8h ago
AR7 Chapter Coverage Heat Map
African source citation density per chapter
📥 Ingest Hub — What Does This Do?
The Ingest Hub is how African climate literature enters the KIFARU repository. Every document — whether discovered automatically by a web crawler or uploaded manually — passes through here before being processed, vetted against IPCC Annex 2 standards, and made available to AR7 chapter authors. Choose the method that fits your source type.
🕸 Auto Crawler
Automatically searches and downloads from UNFCCC, AJOL, OpenAlex, AfDB, WMO, and more. Best for bulk discovery of new sources. Runs in the background.
📁 Manual Import
Upload PDFs, Word docs, or export your Mendeley/Zotero/EndNote library as a RIS or BibTeX file and drop it here. Best for documents you already have.
🔗 DOI / URL Lookup
Paste a DOI, arXiv ID, or URL and KIFARU fetches the full metadata automatically. Best for individual papers you've found elsewhere.
⏳ Queue & Status
See the status of all documents currently being processed — how far they've got, whether they passed the Annex 2 gate, and what was rejected.
🕸 Auto Crawler
📁 Manual Import
🔗 DOI / URL
⏳ Queue
3
🕸
Configure Crawler
Multi-source automated discovery
UNFCCC Corpus
OpenAlex Africa
AJOL Journals
AfDB Portal
WMO Africa
ICPAC / WASCAL
UN Bodies
Zenodo Africa
Natl. Ministries
GBIF Biodiversity
English
French
Portuguese
Arabic
Swahili
Amharic
Run Annex 2 gate automatically
Auto-generate confidence scores
⚙
Crawl Settings
Rate limits & scheduling
📂
Drag & drop files here
or click to browse
PDFDOCXRISBibTeXCSVXLSXXML
or import from reference manager
📚 Mendeley
📖 Zotero
📝 EndNote
Import Settings
Run Annex 2 check on import
Mark as ILK source
DOI / URL Direct Import
Paste any DOI, arXiv ID, PubMed ID, or URL
Completed
12,418
processed
In Queue
1,243
pending
Rejected
87
failed gate
🌍
UNFCCC_BTR_Africa_Batch_0412.zip
218 files · 42 MB
✓ Done
📰
OpenAlex_Africa_Climate_2023.ris
3,421 refs · 1.2 MB
67%
📑
Manual_upload_GCF_reports.pdf
1 file · 4.1 MB
Waiting
Repository Storage Architecture
Where data lives · How to access it · Software stack
🗄
Primary Database
PostgreSQL 16 + pgvector
Stores all source metadata, confidence scores, Annex 2 status, and SHA-256 hashes. pgvector extension adds 1024-dim embedding columns for semantic search.
🕸
Knowledge Graph
Apache Jena Fuseki · SPARQL 1.1
RDF/OWL triple store for semantic relationships between sources, institutions, geographies, and chapters. Exposed via SPARQL endpoint at /kifaru/sparql.
⚡
Vector Index
FAISS / Qdrant · INT8 quantised
High-speed approximate nearest-neighbour search for semantic retrieval. ~200 MB compressed; synced to device edge cache for offline queries.
📊
SDMX Data Flows
SDMX 3.0 · SDMX-JSON 2.0
Structured statistical metadata mapped to AFRICA_GREY_LIT_FLOW dataflow. Interoperable with UN SDMX registry and IPCC TSU systems.
📱
Edge Cache
SQLite + FAISS (device-local)
Offline-first device cache synced from the cloud DB. Stores last 500 evidence items and full vector index. Auto-syncs when connectivity restored.
🔒
Document Archive
Object storage (S3-compatible)
Original PDFs, abstracts, and reviewer packets stored with SHA-256 integrity hashes. ILK sources in segregated partition with access controls.
High Confidence
5,824
score ≥ 0.80 · Annex 2 ✓
Medium Confidence
4,912
score 0.50–0.79
Low Confidence
1,682
score < 0.50
Showing 12
Loading… total
12 seed sources
0 added this session
Last updated: —
🔬 Evidence Synthesis — What Does This Do?
The Synthesis section is where KIFARU becomes an active research assistant for chapter authors and reviewers. Type a question or topic — the AI searches the entire KIFARU repository, triangulates the strongest African evidence, applies IPCC-calibrated uncertainty language, and returns structured Traceable Accounts that you can copy directly into chapter drafts or use to respond to reviewer comments.
📋 What is a Traceable Account?
A Traceable Account is an IPCC-required formatted statement containing: the key finding, confidence level (e.g. "High confidence"), which sources support it, limitations, and the reviewer access link. It is citation-ready for insertion into chapter text.
📦 What is a Reviewer Packet?
A downloadable ZIP file containing the source PDF, an English abstract, a metadata sheet with author contacts, a data availability statement, and a SHA-256 hash. This is exactly what IPCC requires you to provide when citing grey literature in a chapter.
⚡ Rapid Query (≤30 min)
Returns top 5 relevant sources with confidence scores. Use when you need a quick check on whether African evidence exists for a specific topic. Optimised for mobile and low-bandwidth.
📚 Chapter Sprint (48–72h)
Full evidence package for a chapter section: sources, uncertainty statements, gap analysis, and a clean bibliography. Use when drafting or revising FOD/SOD text.
💬 Review Response
Paste a reviewer comment and your draft paragraph. KIFARU finds additional African evidence to address the comment and suggests calibrated response text.
🗺 Gap Analysis
Shows where African evidence is missing for a topic — by country, sector, or time period — so you know what to prioritise finding before the literature cutoff.
Evidence Query
Rapid Query
Chapter Sprint
Review Response
Gap Analysis
Show
per synthesis · sorted by confidence score
📦 Reviewer Packet — what's included in every download
?
Reviewer Packet — The IPCC requires that any grey literature cited in a chapter must be accessible to reviewers. A reviewer packet bundles everything needed: the document itself, contact details, and a hash to verify nothing has changed.
📄 Source PDF
📝 English abstract
📋 Metadata sheet
📞 Author contacts
🔗 Data availability link
🔐 SHA-256 integrity hash
✓ Annex 2 certificate
🗺 Chapter Router — What Does This Do?
The Chapter Router maps every source in the KIFARU repository to its most relevant AR7 chapter. Click any chapter cell to see all African sources routed to that chapter, their confidence scores, and how many are Annex 2 compliant. Use this to assess coverage gaps before the FOD literature cutoff, find evidence for a specific section, or export a ready-to-use bibliography.
WG I — Physical Science
WG II — Impacts & Adaptation
WG III — Mitigation
Chapter Sources
Select a chapter above to view its sources
📋 What you can do with these sources
📦 Download reviewer packets
📋 Copy traceable accounts
⬇ Export BibTeX / RIS
🔬 Run chapter synthesis
🗺 See evidence gaps
Knowledge Graph Explorer
SPARQL-backed · 48,219 nodes
SourceInstitutionGeography
Click nodes to explore
SDMX Data Structure
Live preview — selected source
{
"id": "KIFARU-GE-048291",
"sdmx_flow": "AFRICA_GREY_LIT_FLOW",
"dimensions": {
"REF_AREA": "KE",
"TIME_PERIOD": "2023",
"IPCC_WG": "WGII",
"SECTOR": "URBAN"
},
"attributes": {
"CONF_STATUS": "HIGH",
"CONF_SCORE": 0.87,
"ANNEX2_PASS": true
}
}
"id": "KIFARU-GE-048291",
"sdmx_flow": "AFRICA_GREY_LIT_FLOW",
"dimensions": {
"REF_AREA": "KE",
"TIME_PERIOD": "2023",
"IPCC_WG": "WGII",
"SECTOR": "URBAN"
},
"attributes": {
"CONF_STATUS": "HIGH",
"CONF_SCORE": 0.87,
"ANNEX2_PASS": true
}
}
Nodes
48.2k
Edges
184k
Countries
54
Orphan
312
Passed
11,694
all 9 criteria met
Conditional
637
awaiting human review
Rejected
87
blocked
Conditional — Awaiting Expert Review
K
nowledge
I
ntegrator for
A
frica's
R
esearch &
U
tilization
KIFARU — Knowledge Integrator for Africa's Research & Utilization — is a multi-agent AI platform purpose-built to convert African grey literature into IPCC AR7-compliant, Annex 2-ready evidence. It serves IPCC AR7 Working Group authors (WG I, II, III), chapter coordinating lead authors (CLAs), contributing authors (CAs), and expert and government reviewers throughout the ten-step IPCC authorship cycle — from Scoping through to SPM Approval.
Africa carries the heaviest climate burden yet remains structurally underrepresented in global assessment reports. AR7 is the window to correct this imbalance. KIFARU delivers curated, traceable, confidence-scored African evidence in the exact format IPCC demands — when chapter teams need it. KIFARU is also the Swahili word for rhinoceros — strength, resilience, and persistence.
Africa carries the heaviest climate burden yet remains structurally underrepresented in global assessment reports. AR7 is the window to correct this imbalance. KIFARU delivers curated, traceable, confidence-scored African evidence in the exact format IPCC demands — when chapter teams need it. KIFARU is also the Swahili word for rhinoceros — strength, resilience, and persistence.
Version 1.0 · April 2026
AR7 Phase: Post-Outline / FOD Prep
Edge + Cloud Deployment
AIMS RIC · couma@aimsric.org
§1Getting Started
Installation · First launch · Platform support
Supported Platforms
Desktop / Laptop
Chrome 115+, Firefox 116+, Edge 115+, Safari 16.4+. Full feature set including cloud LLM tier (Claude Sonnet 4.6).
Tablet
iPadOS 16+, Android 12+. Hybrid mode: local inference (Phi-3-mini) for retrieval, cloud for synthesis.
Mobile Phone
iOS 16+, Android 11+. Offline-capable with pre-downloaded vector index (~200 MB). Gemma-2B-IT for edge inference.
Institutional Server
Docker container with PostgreSQL + pgvector + Apache Jena Fuseki SPARQL. Recommended for Sub-Regional Centres.
First Launch Steps
- Open your browser and navigate to the KIFARU URL provided by your AIMS RIC administrator.
- Click Install when prompted to add KIFARU to your home screen or taskbar for offline access.
- Allow local storage (up to 500 MB) on first launch to enable the offline vector index and document cache.
- Log in with your institutional email credentials. Admins approve or adjust role access after registration.
- Select your role: Author, Reviewer, or Data Manager. This pre-configures the interface and available features.
- Select your primary Working Group(s) and chapter assignment(s). KIFARU pre-filters the repository to your scope.
💡On mobile, select Add to Home Screen in your browser menu. KIFARU runs as a full-screen PWA without browser chrome, maximising your working space on small screens.
§2Mission Control (Dashboard)
Real-time pipeline · Stat cards · Chapter heat map · Activity feed
The dashboard is your landing page. It provides a live overview of the entire KIFARU pipeline at a glance. All four stat cards animate on load to show current totals.
Stat Cards
Total Sources
Count of all curated documents across all confidence tiers in the repository.
Annex 2 Compliant
Documents that passed all nine Annex 2 criteria and are cleared for IPCC citation.
Case Studies
Structured capsules generated by the ANALYST agent, each with objectives, methods, outcomes, and limitations.
Chapter Coverage
Percentage of AR7 chapter sections with at least one high-confidence African source routed to them.
Ingestion Pipeline Strip
Four-stage strip (Crawl → Vetter → Structure → Analyse). A pulsing indicator marks the active stage. The live crawl panel shows per-source document counts in real time. Click Manage → to jump to the Ingest Hub.
Chapter Coverage Heat Map
A compact grid showing African source citation density per chapter. Darker fill = higher coverage. Click any cell to open the full Chapter Router for that chapter filtered to the relevant WG scope.
§3Ingest Hub
Four tabs: Auto Crawler · Manual Import · DOI/URL Lookup · Queue & Status
Auto Crawler Tab
Discovers and ingests documents from KIFARU's six-tier source universe. Configure source tiers, languages, WG filter, date range, and crawl schedule before starting.
Source Tiers
- Tier 1 — UNFCCC: BTRs, NCs, NDCs, NAPs, REDD+ technical annexes, ETF datasets
- Tier 2 — African Intergovernmental: AfDB, WMO Africa, ICPAC, WASCAL, SASSCAL, IGAD, AU, AUDA-NEPAD
- Tier 3 — African Academic: AJOL, OpenAlex (Africa affiliation filter), Sabinet, DOAJ Africa, university institutional repositories
- Tier 4 — UN Bodies: UNEP, UNDP, FAO, WHO, UN-Habitat, GCF, GEF — Africa-scope documents only
- Tier 5 — National: Line ministries, NMHSs, national statistics offices, NGO evaluations (IATI registry)
- Tier 6 — Open Science: Zenodo Africa filter, PANGAEA, GBIF, OSF, Figshare
ℹThe Annex 2 gate runs automatically on all crawled documents unless you untick the toggle. Sources that fail are archived in the rejection log but never added to the active repository.
Starting a Crawl
- Tick the source tiers you want to crawl and set language filters.
- Configure the schedule, max docs per source (default 5,000), and request delay (default 800 ms).
- Click Start Crawl. A live progress panel appears with per-source bars.
- Navigate away freely — the crawl continues in the background.
- Click Stop at any time to halt gracefully. Partial results are saved.
Manual Import Tab
For documents you already hold locally. Two methods are available.
Drag-and-Drop File Zone
Drag files from your file manager directly onto the upload area. Accepted formats: PDF, DOCX, TXT, RIS, BibTeX, CSV, XLSX, XML. Drop entire batches at once. Each file processes individually with a real-time progress bar.
Reference Manager Import
Export your library from Mendeley, Zotero, or EndNote as a RIS or BibTeX file, then drop it into the upload zone. KIFARU parses all entries, deduplicates against the existing repository, and processes each reference individually.
💡Mendeley: File › Export › RIS. Zotero: File › Export Library › BibTeX. EndNote: File › Export › select RIS or XML format.
DOI / URL Lookup Tab
Enter any DOI, arXiv ID, PubMed ID, or direct URL. KIFARU resolves the identifier, fetches full metadata, generates a pre-score, and presents a confirmation card before adding to the repository. The batch text area accepts one identifier per line for bulk operations.
§4Repository
Searchable · Filterable · Confidence-scored · 12,418 curated African sources
Confidence Tier Summary
Three tier cards show the distribution by confidence level: High (≥ 0.80), Medium (0.50–0.79), and Low (< 0.50). Use the filter dropdowns to narrow by document type, Working Group, and African sub-region. The filter count updates live.
Source Cards
Each source shows: document type badge, title, country, year, institution, DOI (if available), a confidence bar with numerical score, WG routing chips, and an Annex 2 badge. Click any card to open the full detail modal, which includes IPCC calibrated uncertainty language, reviewer packet download, and options to add to a synthesis queue or copy the traceable account text.
§5Evidence Synthesis
IPCC-calibrated traceable accounts · Four query modes
Rapid Query
Top 5 relevant traceable evidence items within ~30 minutes. Optimised for low-bandwidth and mobile use.
Chapter Sprint
Full evidence package for a specific chapter section within 48–72 hours, including gap analysis and uncertainty language suggestions.
Review Response
Takes a reviewer comment plus draft text and returns targeted African evidence to support or refine the response.
Gap Analysis
Identifies sectors, regions, and time periods where African evidence is absent for a given chapter or topic.
Traceable Account Cards
Each result shows a circular confidence ring, IPCC calibrated uncertainty language, the key finding, a limitation note, source triangulation count, and action buttons: Copy Account, Reviewer Packet, Add to Chapter.
§6Chapter Router
AR7 WG I · WG II · WG III chapter alignment
Maps the repository to the AR7 chapter structure. Each cell shows chapter name, African source count, and a fill bar for coverage density. Click any cell to load the full source list filtered to that chapter's WG scope.
⚠Literature Cutoff: After the TSU announces a cutoff date for any Working Group, KIFARU enforces an ingestion freeze. Sources arriving after the cutoff are archived but marked INELIGIBLE and cannot enter active chapter bibliographies.
§7Annex 2 Gate
IPCC Procedures compliance · 9-point checklist · Non-bypassable gate
Shows three panels: Passed (all 9 criteria met), Conditional (awaiting human expert review), and Rejected (blocked with audit reason codes). QA officers can review conditional items, assign to specialist panels, approve, or reject.
The Nine Annex 2 Criteria
- Archived copy with stable URL, DOI, or controlled reviewer-accessible link
- Full provenance: title, all authors, institutional affiliation, publisher, date
- Methodological transparency: methods section present and legible
- Data availability: raw or processed data accessible or pathway documented
- English summary: abstract or executive summary in English (min. 200 words)
- Reviewer access: confirmed live URL at time of packet generation
- Contact persons: at least one verifiable institutional email
- Critical appraisal: peer review status verified where claimed
- Uncertainty qualifiers: evidence base assessed and calibrated language attached
⚠The Annex 2 Gate is non-bypassable. No user role — including administrator — can push a failing source into the active repository. Rejections are permanently logged in the audit record.
§8LLM Architecture — The Six Agents
Mixture-of-Agents · Each agent runs a purpose-selected model
KIFARU is a mixture-of-agents system. Six specialised AI agents handle distinct functions. No single LLM runs all six agents — this prevents any one model's limitations from compromising the pipeline and allows independent upgrades.
⊙
KIFARU Core — Orchestrator
Routes all tasks · Synthesises multi-agent outputs · Enforces IPCC language
claude-sonnet-4-6Anthropic200K context
Deployment
Cloud — Tier A (laptop/browser) and Tier B hybrid offload
Role
Receives every user request, decomposes into sub-tasks, dispatches to agents, synthesises results, enforces Annex 2 gate logic and IPCC calibrated language compliance across all outputs
Claude Sonnet 4.6 combines a 200K-token context window, strong multi-step reasoning, and reliable tool use — essential for orchestrating long document contexts. It can hold an entire chapter draft alongside multiple evidence packets simultaneously in context.
1
SCOUT — Web Crawler & Ingest Agent
Discovers, fetches, classifies, and deduplicates from all six source tiers
claude-haiku-4-5-20251001Anthropic200K context
Deployment
Cloud — high-throughput, low-latency bulk processing
Tasks
URL resolution, multilingual OCR routing (EN/FR/PT/AR/SWA/AM/HA), document type classification, geographic tagging (ISO 3166), deduplication, metadata normalisation
SCOUT processes thousands of documents per hour. Haiku 4.5 is the fastest and most cost-efficient model in the Claude family — optimal for high-volume classification where deep reasoning is not required. Its quality fully suffices for metadata extraction and document type identification at scale.
2
VETTER — Annex 2 Gate Agent
Non-bypassable compliance gate · Predatory content screen · Peer review verification
claude-sonnet-4-6Anthropic200K context
Deployment
Cloud — reasoning-intensive, non-bypassable gate logic
Tasks
Applies all 9 Annex 2 criteria, screens for Beall-list predatory publishers, verifies peer review process for African journals (editorial policy + CrossRef DOI + DOAJ listing), generates compliance certificates and rejection records
Annex 2 compliance requires nuanced judgement — distinguishing a documented peer review process from a merely claimed one, evaluating methodological transparency from partial text, assessing contact reachability. Haiku lacks the reasoning depth needed here. Sonnet's balance of reasoning quality and throughput is optimal for this gate.
3
STRUCTURER — Metadata, SDMX & Knowledge Graph Agent
SDMX 3.0 · RDF/OWL · Five-level taxonomy · pgvector embeddings
claude-sonnet-4-6Anthropic200K
multilingual-e5-largeMicrosoft / ONNXEdge1024-dim
Deployment
Cloud for LLM structuring; ONNX embedding inference runs locally on all device tiers
Tasks
SDMX 3.0 data structure generation (AFRICA_GREY_LIT_FLOW), RDF/OWL triple generation (Apache Jena Fuseki SPARQL), L1–L5 taxonomy classification, Wikidata reconciliation, pgvector embedding generation
Why multilingual-e5-large? Evaluated against LABSE, mBERT, and XLM-R on African climate text retrieval. multilingual-E5-large produced the highest MRR for cross-lingual retrieval across EN/FR/PT/SWA. It quantises to INT8 at ~200 MB — deployable on mid-range Android devices.
4
ANALYST — Evidence Synthesis & Uncertainty Agent
Traceable accounts · IPCC calibrated language · Gap maps · Case study capsules
claude-sonnet-4-6Anthropic200K context
Deployment
Cloud — full context window required for multi-source synthesis
Confidence scoring
Five-component weighted formula: Source quality (30%) · Methodological rigour (25%) · Geographic specificity (20%) · Temporal currency (15%) · Cross-source triangulation (10%)
IPCC language map
Score ≥ 0.80 = "High confidence" · 0.50–0.79 = "Medium confidence" · 0.30–0.49 = "Low confidence" · < 0.30 = "Very low confidence"
Synthesis is the most intellectually demanding task. ANALYST must hold multiple source documents in context, identify convergent findings, apply the Mastrandrea et al. 2010 evidence-agreement matrix, and produce text meeting IPCC calibrated language standards without hallucination. The 200K window allows processing an entire chapter's sources in one pass.
5
RESPONDER — Author Interface Agent
Four interaction modes · Cutoff compliance · Bibliography export · Reviewer packets
claude-sonnet-4-6Anthropic200KCloud
Phi-3-mini-128kMicrosoft / ONNX INT4Edge ~2.3 GB
Gemma-2B-ITGoogle / ONNX INT4Edge ~1.4 GB
Mode selection
Auto-selects Claude Sonnet when connected; falls back to Phi-3-mini (tablet) or Gemma-2B (phone) when offline
Interaction modes
Mode A: Rapid query (≤30 min) · Mode B: Chapter sprint (48–72h) · Mode C: Review response · Mode D: Cutoff compliance check
Why Phi-3-mini on edge? Evaluated against Mistral-7B, Llama-3-8B, and Gemma-7B. Phi-3-mini achieved comparable output quality at one-third the size, deployable on mid-range Android tablets (4 GB RAM). Its 128K context window — exceptional for its size — holds multi-source evidence packages without truncation.
6
LEARNER — Adaptive Learning & Feedback Agent
RAG optimisation · Federated learning · Drift detection · ILK validation routing
multilingual-e5-large (fine-tuned)Quarterly retrain
MiniLM-L-12-v2 cross-encoderHuggingFace
Federated learning
Flower (flwr) framework — privacy-preserving gradient aggregation across Sub-Regional Centres. Model updates share only gradients, never raw documents.
Drift detection
Evidently AI monitors retrieval precision@5, Annex 2 pass rate, confidence calibration error (ECE). Alert threshold: >5% degradation triggers retraining flag.
ILK handling
ILK sources are excluded from algorithmic scoring and federated pipelines entirely. Routed to human expert validation via IPBES credibility-legitimacy-relevance framework.
§9LLM Quick-Reference
All models deployed across the six agents at a glance
| Agent | Model | Provider | Context | Selection rationale |
|---|---|---|---|---|
| Core Orchestrator | claude-sonnet-4-6 | Anthropic | 200K | Multi-step reasoning, tool use, large context. Routes all tasks and synthesises outputs. |
| SCOUT | claude-haiku-4-5-20251001 | Anthropic | 200K | Fastest inference, lowest cost per token. Ideal for high-volume document classification (thousands per hour). |
| VETTER | claude-sonnet-4-6 | Anthropic | 200K | Nuanced compliance judgement required. Haiku insufficient; Opus over-specified. |
| STRUCTURER (LLM) | claude-sonnet-4-6 | Anthropic | 200K | Structured output for SDMX and RDF triples requires reliable format adherence. |
| STRUCTURER (embed) | multilingual-e5-large | Microsoft / ONNX | 512 tok | Best MRR for cross-lingual African climate text retrieval. INT8-quantisable to ~200 MB. |
| ANALYST | claude-sonnet-4-6 | Anthropic | 200K | Deep synthesis and uncertainty calibration demand full reasoning capability. |
| RESPONDER (cloud) | claude-sonnet-4-6 | Anthropic | 200K | Final author-facing formatting and review response generation. |
| RESPONDER (edge) | Phi-3-mini-128k-instruct | Microsoft / ONNX INT4 | 128K | Best quality-to-size ratio tested. Deployable on 4 GB RAM tablets. 128K context exceptional for its size. |
| RESPONDER (alt edge) | Gemma-2B-IT | Google / ONNX INT4 | 8K | Ultra-light fallback for <3 GB RAM devices. ~1.4 GB footprint. |
| LEARNER (retrieval) | multilingual-e5-large (fine-tuned) | Microsoft / Flower | 512 tok | Quarterly fine-tuning on KIFARU author feedback via federated learning. |
| LEARNER (re-rank) | MiniLM-L-12-v2 cross-encoder | HuggingFace | — | IPCC-domain precision re-ranking of retrieved candidates. |
ℹModel upgrade policy: Quarterly review cycle aligned to AR7 calendar. Upgrades require improvement on ≥ 2 of 3 metrics without degrading the third. Model upgrades are frozen during SOD and Final Draft phases to ensure output consistency across the review cycle.
🔒Privacy: All API calls to Anthropic use zero data retention (ZDR) endpoints — query content is not stored beyond the inference request. ILK sources are never transmitted to cloud LLM endpoints without explicit community consent.
§10Exporting Data
SDMX · Reviewer packets · Bibliography · Traceable account text blocks
SDMX Export
Top-bar button → select SDMX-JSON 2.0 or SDMX-ML 3.0, scope, and options. Downloads as ZIP with manifest and SHA-256 checksums for every source.
Reviewer Packet
Source PDF + English abstract + metadata sheet + contact details + data availability statement + SHA-256 hash. Download from any source card or synthesis result.
Bibliography
BibTeX, RIS, or IPCC-format list. IPCC format appends confidence score and reviewer URL to each entry. Export from Chapter Router.
Traceable Account
Formatted text block (source ID, key finding, confidence statement, limitations, reviewer link) ready for direct chapter draft insertion. Click Copy Account on any synthesis card.
§11Glossary
Key terms used throughout KIFARU and the IPCC AR7 process
| Annex 2 | IPCC Procedures Annex 2 — nine compliance criteria for non-peer-reviewed literature to be cited in assessment reports. |
| Calibrated language | IPCC-mandated vocabulary for uncertainty: "High confidence", "Very likely", etc. Defined in Mastrandrea et al. (2010). |
| Case study capsule | Structured KIFARU output with objectives, methods, key outcomes, quantitative findings, and limitations. |
| CLA | Coordinating Lead Author — senior author responsible for a chapter section in an IPCC Working Group report. |
| Confidence score | KIFARU's numerical (0.00–1.00) weighted evidence quality assessment, mapped to IPCC calibrated language tiers. |
| FOD | First-Order Draft — the initial chapter draft reviewed by invited experts. |
| ILK | Indigenous and Local Knowledge — subject to special data sovereignty handling. Never algorithmically scored in KIFARU. |
| KIFARU | Knowledge Integrator for Africa's Reviewable Understanding. Also Swahili for rhinoceros — strength and resilience. |
| LLM | Large Language Model — a neural network trained on large text corpora for natural language understanding and generation. |
| MoA | Mixture-of-Agents — AI architecture where multiple specialised models handle distinct sub-tasks with an orchestrator synthesising outputs. |
| ONNX | Open Neural Network Exchange — open format enabling ML model deployment across hardware and browser environments. |
| Provenance package | Full documentation set required by Annex 2 for a grey literature source. |
| PWA | Progressive Web Application — installable web app capable of offline operation. |
| RAG | Retrieval-Augmented Generation — retrieves relevant documents from a database to improve LLM answer accuracy. |
| SDMX | Statistical Data and Metadata eXchange — ISO standard for sharing structured statistical data, adopted by the UN system. |
| SOD | Second-Order Draft — revised chapter draft reviewed by both experts and governments. |
| SPM | Summary for Policymakers — the approved high-level summary of each Working Group report. |
| Traceable account | IPCC term for a fully cited, source-identified statement traceable back to its underlying evidence. |
| TSU | Technical Support Unit — manages logistical and technical support for each IPCC Working Group. |
| Vector index | Data structure storing embedding vectors for all documents, enabling fast semantic similarity search. |
| WASM | WebAssembly — binary instruction format enabling near-native-speed execution in browsers; used for on-device OCR and LLM inference. |
🔗 Blockchain Approval Layer — How It Works
Every source submitted by any user is placed into a cryptographically signed approval chain before it enters the main repository. Each pending source is a block containing: a SHA-256 hash of the source content + the submitter's email + a timestamp. That block is chained to the previous block's hash, making the sequence tamper-evident. Admins approve blocks. Once a block reaches the required approval threshold it is automatically committed to the repository. No single person — including SuperAdmins — can add a source without the chain recording who did it and when.
🔐 Tamper-evident
Changing any block invalidates all hashes after it. Any tampering is immediately detectable by recomputing the chain.
👥 Multi-approver
Each source requires a configurable number of approvals (default: 2). Single-person approval is not possible for non-Admin submissions.
📋 Full audit trail
Every approval, rejection, and state change is recorded with the approver's email and timestamp. The audit log cannot be deleted.
⚡ Auto-commit
Once a block reaches its approval threshold, it moves from "Pending" to the main repository automatically — no manual step needed.
🔗
Approval Chain
Pending · Approved · Rejected
👥
User Registry
Roles · Permissions · Emails
⚙
Configuration
Database · Auth · Roles