{ "version": "v2.0", "date": "2026-06-25", "schema_url": "/methodology.v2.schema.json", "license": "CC-BY-4.0", "formula": "agi_score = round((autonomy + multimodality + production) / 15 * 100)", "blend": { "curator_weight": 0.6, "community_weight": 0.4, "community_quorum": 5, "max_per_category": 5, "category_count": 3 }, "rubric": [ { "key": "autonomy", "label_en": "Autonomy", "label_de": "Autonomie", "description_en": "Self-direction, planning, tool-use, self-correction, runtime hours without a human in the loop.", "description_de": "Self-direction, Planung, Tool-Use, Selbstkorrektur, Lauflängen ohne Mensch im Loop.", "levels": [ { "level": 1, "label": "Scripted", "anchor_en": "Single-prompt chains, no planner, no tool selection at runtime.", "anchor_de": "Einzel-Prompt-Ketten, kein Planner, keine Tool-Wahl zur Laufzeit.", "evidence": [ "repo example without planner", "no tool-router module" ] }, { "level": 2, "label": "Reactive", "anchor_en": "ReAct-style tool calls, one-shot retries, no memory across runs.", "anchor_de": "ReAct-Tool-Calls, einmalige Retries, kein Cross-Run-Memory.", "evidence": [ "tool-call loop in core", "retry policy in code" ] }, { "level": 3, "label": "Planning", "anchor_en": "Explicit planner module, subgoal decomposition, bounded self-critique.", "anchor_de": "Expliziter Planner, Subziel-Zerlegung, begrenzte Selbstkritik.", "evidence": [ "planner class/graph", "critique step in docs" ] }, { "level": 4, "label": "Persistent", "anchor_en": "Multi-step persistent state, long-horizon memory, supervised self-correction.", "anchor_de": "Persistenter Zustand, Long-Horizon-Memory, überwachte Selbstkorrektur.", "evidence": [ "persistent memory store", "documented horizon > 30 steps" ] }, { "level": 5, "label": "Open-ended", "anchor_en": "Open-ended task pursuit, reproducible runs > 8h with no human input, published evals.", "anchor_de": "Open-ended Aufgaben, reproduzierbare Runs > 8h ohne Mensch, publizierte Evals.", "evidence": [ "published eval (e.g. GAIA/SWE-bench)", "public run trace > 8h" ] } ] }, { "key": "multimodality", "label_en": "Multimodality", "label_de": "Multimodalität", "description_en": "Text · vision · voice · code · computer-use · world models — how natively the framework reaches in.", "description_de": "Text · Vision · Voice · Code · Computer-Use · World-Models — wie nativ das Framework zugreift.", "levels": [ { "level": 1, "label": "Text-only", "anchor_en": "Text in / text out. No vision, audio, or code execution loops.", "anchor_de": "Text rein / Text raus. Keine Vision, Audio oder Code-Execution.", "evidence": [ "only text adapters in repo" ] }, { "level": 2, "label": "Code-capable", "anchor_en": "Code execution sandbox or first-class code-interpreter integration.", "anchor_de": "Code-Execution-Sandbox oder first-class Code-Interpreter.", "evidence": [ "code-exec module", "sandbox docs" ] }, { "level": 3, "label": "Bi-modal", "anchor_en": "Native vision OR voice in addition to text+code.", "anchor_de": "Native Vision ODER Voice zusätzlich zu Text+Code.", "evidence": [ "vision adapter", "voice adapter" ] }, { "level": 4, "label": "Tri-modal", "anchor_en": "Text + code + vision + voice with documented examples.", "anchor_de": "Text + Code + Vision + Voice mit dokumentierten Beispielen.", "evidence": [ "multi-modal example", "released cookbook" ] }, { "level": 5, "label": "Computer-use", "anchor_en": "Computer-use, GUI control or world-model integration shipped as supported feature.", "anchor_de": "Computer-Use, GUI-Steuerung oder World-Model als unterstütztes Feature.", "evidence": [ "computer-use module", "release notes shipping it" ] } ] }, { "key": "production", "label_en": "Production", "label_de": "Production", "description_en": "Eval harnesses, observability, cost control, scale, deployment paths.", "description_de": "Eval-Harnesses, Observability, Cost-Control, Skalierung, Deployment-Pfade.", "levels": [ { "level": 1, "label": "Demo", "anchor_en": "Notebook-grade. No tests, no observability, no deployment guide.", "anchor_de": "Notebook-Niveau. Keine Tests, keine Observability, keine Deployment-Guide.", "evidence": [ "no CI", "no tracing" ] }, { "level": 2, "label": "Tested", "anchor_en": "CI test suite, basic logging, manual deployment recipe.", "anchor_de": "CI-Tests, Basis-Logging, manuelles Deployment-Rezept.", "evidence": [ "CI workflow", "logging module" ] }, { "level": 3, "label": "Observable", "anchor_en": "Built-in tracing (OTel/LangSmith/W&B), cost telemetry, structured retries.", "anchor_de": "Eingebautes Tracing (OTel/LangSmith/W&B), Cost-Telemetrie, strukturierte Retries.", "evidence": [ "tracing integration", "cost metering hooks" ] }, { "level": 4, "label": "Scaled", "anchor_en": "Async/distributed runtime, queueing, horizontal scaling docs, SLO guidance.", "anchor_de": "Async/Distributed-Runtime, Queueing, Scale-Out-Docs, SLO-Guidance.", "evidence": [ "distributed runner", "scale-out guide" ] }, { "level": 5, "label": "Hardened", "anchor_en": "Reference deployment (k8s/serverless), eval harness gating releases, SOC2-ready logging.", "anchor_de": "Referenz-Deployment (k8s/serverless), Eval-Harness als Release-Gate, SOC2-fähiges Logging.", "evidence": [ "reference k8s/serverless template", "eval gating release notes" ] } ] } ], "changelog": [ { "version": "v2.0", "date": "2026-06-25", "summary_en": "Added 5-level anchor rubric per category, evidence requirements, confidence score, and JSON-Schema for the framework record.", "summary_de": "5-stufige Anker-Rubrik pro Kategorie, Evidenz-Anforderungen, Confidence-Score und JSON-Schema für den Framework-Record ergänzt." }, { "version": "v1.0", "date": "2026-04-01", "summary_en": "Initial release: three categories (autonomy, multimodality, production), 60/40 curator/community blend, 1–5 scale.", "summary_de": "Erstveröffentlichung: Drei Kategorien (Autonomie, Multimodalität, Production), 60/40 Kurator/Community, 1–5-Skala." } ] }