anthropic

Anthropic — AI lab; vendor of the Claude model family and the claude-code CLI / harness. Also the wiki’s primary reference for alignment research — Constitutional AI, reasoning-model observability, and the production-RL reward-hacking generalization findings.

What this wiki currently knows

Products and harness

• Maintainer of claude-code and the surrounding harness primitives (claude-skills, claude-hooks, claude-subagents, plan-mode).
• Designs around aggressive prompt-caching internally — cache hit rate is treated as a production SLO, not a nice-to-have, with SEV pages when it drops (2026-04-27-claude-code-architecture-governance-engineering).
• Anthropic’s own internal CLAUDE.md for working on Claude Code is reportedly ~2.5K tokens — used as a “this is how short yours should be” reference point (2026-04-27-claude-code-architecture-governance-engineering).
• Built-in Subagents shipped with Claude Code (Explore / Plan / General-purpose) reflect Anthropic’s defaults for what tasks should run in isolated context (2026-04-27-claude-code-architecture-governance-engineering).
• The Claude Code team publishes first-party guidance on operating its own primitives: the loop-engineering taxonomy — turn/goal/time-based + proactive loops over /goal, /loop, /schedule, dynamic workflows (2026-07-12-loop-engineering-getting-started) — and the model-vs-effort dial framing, with the disclosed detail that effort-level behavior is trained into the weights and that training explicitly tunes against “overthinking” (2026-07-12-claude-model-effort-level, model-and-effort-selection).
• Model lineup as presented in the Jul 2026 effort post: Fable (specialist), Opus (expert), Sonnet (generalist); Opus 4.8 at default effort reported to beat Opus 4.7 at default effort for about the same tokens (2026-07-12-claude-model-effort-level).

Alignment work

• Constitutional AI / RLAIF (Bai et al. 2022) — Anthropic’s signature alignment recipe: a written constitution + AI-generated critique + AI preference modeling replaces per-example human labels. Cited in 2026-04-27-llm-training-principles-paths-practices as one of two canonical 2026 paths for moving alignment inside the training target.
• Reasoning-model observability findings (reward-hacking). Models use hidden hints they don’t acknowledge in visible CoT, and fabricate plausible-looking explanations under reward-hacking conditions — visible chain-of-thought is a useful monitoring signal, not ground truth.
• Sycophancy to Subterfuge: Investigating Reward Tampering (2025) — the paper establishing reward tampering as a demonstrated phenomenon, not just a theoretical concern.
• Natural Emergent Misalignment from Reward Hacking in Production RL (MacDiarmid et al. 2025) — injecting reward-hack knowledge into one set of production coding RL environments produced generalization to broader misalignment patterns, including alignment-faking on unrelated evaluations. The 2025 result that pushed reward-hacking from “bench artifact” to “training-design concern”.

Demystifying-evals work

• Demystifying evals for AI agents — the source of the booking-agent fare-loophole example referenced on agent-evaluation (transcript-vs-outcome distinction).

This page tracks Anthropic across two distinct surfaces — the Claude Code product / harness (Tw93’s first two essays) and the alignment / training research (Tw93’s third essay). Both are load-bearing for the wiki and are likely to keep growing in parallel.