Corpus-Complete Problem Taxonomy

Snapshot: corpus-complete-20260711 (2026-07-11) · source commit 1737631198b184e05ba0a343d67df6fa9e71b938 Scope: 1780 approved schema-v2 extractions; 1754 deduped sources; 991 deduped agentic projects.

Corpus-complete analysis output. Counts are generated from extraction.json; citations link to deterministic report.md views. This is an input to, not the output of, Phase 4 synthesis.

This taxonomy uses an agent-authored rubric. Class names, descriptions, and signals are an explicit judgment layer; assignments are then computed deterministically for reproducibility. This is not embeddings-only clustering.

Assignments are multi-label. Each project keeps one primary class for ranking, while strong cross-cutting classes such as evaluation, safety, retrieval, context, and infrastructure are retained as secondary labels.

Denominator semantics: percentages use 1754 deduplicated sources. Any-label class counts overlap and therefore do not sum to the corpus size; primary counts form the partition.

Fallback primary assignments: 0 projects (0.0% raw-project rate). Near-tied top assignments: 416 projects (23.4%).

Taxonomy Classes

Rubric Rationale

Agentic workflow automation and orchestration

This class captures the agentic-systems core: decomposed tasks, agent handoffs, workflow state, and tool execution.

Signals: agent, workflow, orchestration, orchestrator, multi-agent, handoff, tool, task automation, autonomous

Coding-agent and software delivery automation

Coding-agent cases have distinct constraints: repositories, CI, PRs, test loops, merge conflicts, and developer context.

Signals: coding agent, code, software, developer, github, pull request, pr, ci, test, refactor, copilot, codex

Enterprise knowledge retrieval and RAG

These projects solve knowledge access problems by retrieving, ranking, grounding, and citing internal or external content.

Signals: rag, retrieval, knowledge, search, vector, embedding, semantic, document, question answering, faq

Evaluation, observability, and reliability governance

The problem is quality control under nondeterminism: eval harnesses, traces, judges, metrics, and release gates.

Signals: evaluation, eval, observability, monitor, tracing, metrics, quality, reliability, judge, benchmark, regression

Customer support and conversational service

These cases share user-facing dialogue constraints: escalation, tone, resolution quality, and support-channel integration.

Signals: customer, support, chatbot, chat, conversation, conversational, service, sales, assistant, call center

Domain-specific model adaptation

The central problem is model behavior adaptation, not just application orchestration.

Signals: fine-tun, continued pretraining, domain-specific, domain adapted, distillation, small language model, slm, training, model adaptation

Document processing and structured extraction

These cases are shaped by document layout, schema fidelity, extraction accuracy, and review loops.

Signals: document, pdf, form, contract, invoice, metadata, extraction, structured, ocr, legal

Search, ranking, and recommendation quality

The recurring problem is improving retrieval/relevance/ranking behavior rather than general knowledge access.

Signals: ranking, recommendation, recommend, rerank, retrieval ranking, matching, discovery, commerce search, product search

Safety, compliance, and trust controls

These cases are dominated by trust boundaries and require deterministic controls around probabilistic models.

Signals: safety, compliance, privacy, security, moderation, policy, guardrail, regulated, risk, trust

Enterprise AI platform adoption and governance

The problem is organizational deployment: platforms, governance, enablement, and shared infrastructure.

Signals: enterprise, platform, adoption, governance, internal, organization, company-wide, developer platform, enablement

Context engineering and memory management

These cases make context itself the design object: what to include, retain, compress, or retrieve.

Signals: context, memory, long context, prompt, instructions, compression, compaction, session, state

Data analytics and decision intelligence

The core problem is making data and metrics answerable or actionable through AI interfaces.

Signals: analytics, data analyst, business intelligence, dashboard, sql, forecast, decision, data warehouse, metrics

Infrastructure, cost, and latency optimization

These cases are primarily about production constraints around throughput, cost, latency, and deployment architecture.

Signals: latency, cost, inference, throughput, scale, rps, deployment, gpu, cpu, infrastructure, serving

Multimodal, edge, and voice AI systems

These cases introduce modality and device constraints that change system architecture.

Signals: voice, speech, audio, vision, image, multimodal, edge, mobile, wearable, camera

Content generation, localization, and creative operations

These cases share content-production constraints: brand quality, localization, review, and throughput.

Signals: content, marketing, translation, localization, creative, copy, campaign, product description

General production LLM systems

Kept explicit so low-confidence assignments are visible rather than forced into a false cluster.

Signals: fallback class when no narrower class scores.

Design Gaps

• Multi-label assignment is signal based and intentionally conservative; reviewer sampling should check whether the secondary threshold misses cross-cutting themes.
• Any-label counts are not additive across classes because a single deduped project can appear in several cross-cutting classes.
• 0 fallback projects and 416 near-tied assignments remain explicit for reviewer inspection.
• Problem statements do not carry separate evidence arrays, so taxonomy citations point to project reports and extracted fields rather than direct problem-level quotes.