Governance › Citations
Citations
Every source behind this portal, in one place. Internal sources are engagement and program records; published sources are the guidance the guidelines were validated against; platform documentation grounds the overlays and externally validates specific principles. Where a principle on the Decision Register page has an external validation, the register is stronger for it, the reasoning is traceable, not asserted.
Internal sources
Confidential, internal architecture references, not for external distribution. Drop this page from client-facing copies, along with the discovery chips.
| Source | Origin | What it informed | |
|---|---|---|---|
| I1 | Internship Wrap meeting notes | Point One Zero internal, Aug 19 2026 | Seed of the 36 architecture guidelines and the determination rule; the Smart Pipe antipattern was caught live in this review. |
| I2 | PE Platform Interactive Architecture Guide | Point One Zero internal, July architecture review | As-built evidence base. Contributed the decision-register mechanism, classification inheritance, restricted-store handling, scoped index lifecycle, source authority weighting, the canonical record, and identity passthrough. Its eleven layers map into the seven here without loss. |
| I3 | IC Memo Generation Platform, architecture overview diagram | Point One Zero internal | Contributed the deterministic rendering boundary (immutable templates, dynamic data only), the source catalog, source-of-truth-stays-in-the-estate, disaster recovery, and the append-only audit-trail statement. |
| I4 | Architecture Portal v1 | This program; superseded by this portal | The 28-step greenfield setup sequence in Environment & Operations was restored from it. |
| I5 | Architecture Guidelines Validation workbook | Companion Excel asset, Aug 19 2026 | The 36 guidelines validated line by line against S1 to S7 below; ten additional guidelines adopted from those sources. |
Published engineering guidance
The validation base for the guidelines, the SSRA file grammar, and the pattern library.
| Source | Origin | What it informed | |
|---|---|---|---|
| S1 | Building Effective Agents | Anthropic Engineering, Dec 2024 | The determination rule (workflows are predefined code paths), the simplicity principle, and the routing and orchestrator-workers shapes behind patterns P1 and P6. |
| S2 | Claude Code Best Practices | Anthropic docs, originally authored by Boris Cherny, Apr 2025 | The SSRA file grammar: constitutions, skills, commands, hooks as deterministic guardrails, permissions, verification loops. |
| S3 | How We Built Our Multi-Agent Research System | Anthropic Engineering, Jun 2025 | P6 economics (the 15x rule and effort-scaling dispatch), the four-part delegation spec in agent definitions, bounded subagent summary contracts, end-state evaluation with LLM-as-judge, checkpoint-resume and rainbow deployments in the harness inventory, and the SSRA authoring guidance. |
| S4 | Writing Effective Tools for Agents | Anthropic Engineering, 2025 | The tool-design standard for the south servers: consolidation, namespacing, high-signal returns, pagination and truncation defaults. |
| S5 | Effective Context Engineering for AI Agents | Anthropic Engineering, 2025 | Context packs and just-in-time retrieval, compaction policy, system-prompt altitude, subagent context isolation. |
| S6 | Building Claude Code with Boris Cherny | The Pragmatic Engineer interview | Promotion-on-repetition (the build-twice rule), agentic search over static indexes, plan-then-build discipline. |
| S7 | Boris Cherny's 13 Claude Code Tips | Boris Cherny, Jan 2026 (community transcription) | Shared team constitutions via git, the verification feedback loop, subagents for verification, hooks for auto-format. |
Industry platform documentation
External validation of specific principles, and the sources behind the cloud overlays.
| Source | Origin | What it informed | |
|---|---|---|---|
| X1 | Palantir Foundry: Markings · inherited markings · removal guidance | Palantir documentation | Validates classification inheritance as default platform behavior: markings propagate to every derived dataset; removal is an explicit, permissioned exception. |
| X2 | Microsoft Purview: Sensitivity labels · best practices | Microsoft documentation | The same principle shipped as label inheritance: sensitivity follows data into derived assets. |
| X3 | AWS Bedrock AgentCore | AWS documentation | The AWS overlay in Environment & Operations and the 28-step greenfield setup mapping. |
| X4 | Azure AI Foundry | Microsoft documentation | The Azure overlay in Environment & Operations. |
| X5 | Google Cloud Vertex AI | Google Cloud documentation | The GCP overlay in Environment & Operations. |
| X6 | Model Context Protocol specification | modelcontextprotocol.io | The protocol beneath both boundaries: MCP is the bridge, never the business logic, the platform is a server northward and a client southward. |
| X7 | Matthew Coombs, Enterprise AI Architecture infographic | LinkedIn, 2026 | The external reference architecture the diagram was contrasted against; the gaps identified fed the v2 model. The comparison view has since been retired from the portal so a single layer vocabulary appears; this citation stands as the record of what was borrowed. |
| X8 | Claude Managed Agents overview | Anthropic platform documentation | The managed-agent-runtime concept and the buy-versus-build open item on the Decision Register; terminology alignment across the portal (subagents, skills, slash commands, project memory). |
| X9 | AWS: S3 Vectors with Bedrock Knowledge Bases · S3 Vectors GA announcement | AWS documentation | The vector-store decision guidance on the AWS overlay (OpenSearch for latency and throughput, S3 Vectors for massive cost-optimized collections, Knowledge Bases backing onto either), added while reconciling the AgentCore setup walkthrough. |
Engineering Standards sources
Added Aug 21 2026 with the Engineering Standards section and the SSRA correction. These ground the skill anatomy, the project skeleton, the eval and judge controls, the observability conventions, the caching economics and the model-customisation routes.
| Source | Origin | What it informed | |
|---|---|---|---|
| S8 | Agent Skills overview | Anthropic platform documentation | The corrected SSRA expansion, Skills, Scripts, References, Assets, the three-level progressive disclosure model, the rule that script source never enters context, and the on-demand reference-loading behaviour behind the Skill anatomy page. |
| S9 | Prompt caching | Anthropic platform documentation | Cache breakpoint placement, prefix ordering (tools, system, messages), TTL behaviour and the read-versus-write pricing structure behind the Token & cost engineering page. |
| S10 | 12-Factor Agents | Dex Horthy / HumanLayer | The engineering posture on the Project skeleton standard page: own your prompts, context window and control flow; structured output as the module contract; small focused agents; humans reached by tool call; errors compacted rather than looped on. |
| S11 | GenAI observability with OpenTelemetry · Langfuse OTEL endpoint | OpenTelemetry & Langfuse documentation | The vendor-neutral instrumentation standard: GenAI semantic conventions for span naming, model and token attributes, tool and agent spans, and content-capture opt-in. The basis for treating the observability backend as swappable. |
| S12 | Golden dataset evaluation | Langfuse engineering guidance | Golden-set sizing (start at twenty to fifty, grow to a few hundred), the fast CI subset, dataset versioning, holdout partitions and monthly refresh from production failures. |
| S13 | LLM-as-a-judge bias and calibration literature | MT-Bench and successor studies; 2026 judge-reliability surveys | The judge control table: human calibration before promotion, cross-family judging against self-preference, order rotation against position bias, rubric decomposition, rationale requirements, verbosity normalisation and pinning the judge contract. |
| X10 | Bedrock model customisation · RFT for open-weight models | AWS documentation | The managed supervised and reinforcement fine-tuning routes, reward-function grading with rule-based or AI judges, intermediate checkpoints, and custom model import for open-weight architectures. |
| X11 | Fireworks AI supervised fine-tuning | Fireworks AI documentation | The hosted tuning route: JSONL in standard chat-completion format, LoRA adapters trained elsewhere and uploaded for serving, and the vendor-boundary caution on the Model customisation page. |
| X12 | Multi-LoRA serving on SageMaker and Bedrock | AWS documentation | Adapter-based customisation economics: frozen base weights, adapters swapped per request, several tuned variants sharing one GPU. |