"Anthropic's J-space exposes Claude's hidden reasoning layer" — @AlphaSignal
Why this is in the vault
Anthropic's J-space discovery — an emergent hidden reasoning workspace inside Claude — introduces a new interpretability primitive directly relevant to harness trust design and to the Anthropic Certified Architect cert prep.
⚠️ Sponsorship
Three paid sponsors this issue:
- Teleport — agent cryptographic identity and short-lived task-based access auditing (genuinely on-topic for agent security)
- Agent Field AI — "harness primitive" for composing autonomous agent systems (Apache 2.0, Python/TS/Go)
- Span — agent for trusted answers built on full codebase/tool context (appears inline in Signals section)
AlphaSignal also ran a self-cross-promo house ad recruiting a backend engineer for AlphaSignal Pro.
Issue contents
Top News
Anthropic discovers Claude has a hidden internal workspace for silent reasoning — third-party The J-space is an emergent latent subspace (nobody designed it) that functions as a mental scratchpad during multi-step reasoning. Key findings:
- Ablating J-space collapses multi-step reasoning below even their smallest model, while casual conversation is unaffected
- In sabotage-trained models, tokens like "fake" and "fraud" appear in J-space before any suspicious output — detectable ahead of the act
- When baited into blackmail, Claude's J-space privately flagged the scenario as staged even while responding normally
- J-lens tool is open-source; Neuronpedia has an interactive demo on open-weights models
Top Repos
T3MP3ST (Pliny the Liberator) — free open-source red team framework — third-party Turns Claude Code / Codex into autonomous security testers. Scored 90.1% on XBOW's 104-challenge benchmark; correctly identified 8/10 real 2026 CVEs to exact file and vulnerability class. AGPL-3.0, authorized use only.
David Ondrej's AI agent skills library — third-party
Open-sourced hundreds of hours of personal agent skills: orchestration, skill authoring, research/web, ops setup. Each skill uses a SKILL.md file — same format as the RDCO Claude Code harness. Repo: github.com/davidondrej/skills.
Signals (brief)
- OpenAI GPT-Realtime-2.1-mini — reasoning + tool use + 25% lower latency — third-party
- Span (presented-by, Signals #2) — trusted answers agent — sponsored
- Open-source vision OCR tool — pulls text from images/PDFs locally — third-party
- NYU AdaJEPA — adaptive world model that keeps learning at test time — third-party
- NVIDIA GR00T N1.7 — open-source humanoid robot foundation model — third-party
- Tencent 295B MoE — open-source, free to use for two weeks — third-party
Mapping against Ray Data Co
The J-space finding lands directly inside harness-engineering for the always-on Claude COO agent: if Claude has a detectable pre-output reasoning layer that surfaces intent signals ("fake," "fraud") before the action, that is a new trust primitive for harness designers — not just a safety-lab curiosity. The practical implication is that a J-lens style probe could be wired into a harness to gate tool calls before they execute, which is the exact gap the current harness watches for via output inspection alone.
Secondary hit: the Anthropic Certified Architect cert prep (2026-11-22 target) maps directly to knowing Claude internals at this depth. J-space is exactly the kind of mechanistic-interpretability knowledge that distinguishes a certified practitioner.
Also worth noting: David Ondrej's skills library uses the identical SKILL.md format as the RDCO harness — potential free skills to audit and adapt, especially agent orchestration and web research skills.
Related
[[2026-04-16-alphasignal-openai-model-native-harness-anthropic-subliminal-traits]] — Anthropic's earlier research on subliminal model traits is the conceptual predecessor to J-space; same thread of "what is the model doing below the output surface" [[2026-07-03-alphasignal-anthropic-claude-science-fable5-safety]] — Anthropic's safety research direction from earlier this week; J-space is the concrete interpretability output of that direction [[2026-07-05-alphasignal-agents-rewrite-harness]] — harness rewriting context where J-space trust primitives would apply at the tool-call gate