Knowledge map

All connections

• R001 and R002 (weak): Both argue that agent capability depends on the system around the model rather than the model alone, and both turn on a consistency tradeoff. R001 relaxes Git refs to causal consistency; R002 leaves contradictory facts in semantic memory as an open problem.
• R002 and R004: R002 says memory fails when maintenance (dedupe, reflection, garbage collection) is missing. R004's 5-month wipe is that garbage collection done by hand, all at once, without the eval gate R002's commenters ask for.
• R005 and R002: R005's second brain answers one of R002's open problems: contradictory facts are kept side by side with dates and sources, never overwritten.
• R005 and R004: Both say harness config is a bet that expires with model upgrades. R004 wipes it by hand once; R005 schedules a quarterly re-audit and puts evals in front of it, which is the gate R004 lacks.
• R005 and R006: Both say an agent is the model plus a system around it, and their parts overlap: R005's loop, gate and harness layers are R006's agent loop, guardrails and permissions, and tool calling. R005 makes evals a layer of its own, which is the seventh term R006's commenters say is missing.
• R002 and R006: R006's context management and checkpointing are R002's working and episodic memory seen from the harness. R002's author comments on R006 that context management is the key factor.
• R004 and R006: R004 configures the harness; R006 describes what it does. R004's "default to Auto" is only safe when the harness enforces R006's permissions by blast radius (read, reversible write, irreversible write).
• R010 and R002: A retrieval design for semantic memory; "similarity is not relevance" is R002's failed read step, returning the near but wrong fact.
• R010 and R006: Retrieval as tool calling and context management.
• R010 and R005: Retrieval as the context layer: "what it sees".
• R012 and R005: AMIE's zero safety stops come from a physician watching every chat live, which is R005's gate layer at full cost. Scaling a crash project means replacing that observer with evals and risk-tiered review.
• R012 and R006: AMIE's live physician is R006's human-in-the-loop guardrail at full cost; the open problem is replacing it with evals and risk-tiered review.
• R012 and R004: R004's Goal + Context prompting assumes the user can supply context. The financial-advice paper measures the cost when they can't, so the product should ask for the missing context itself.
• R012 and R014: R014 tells beginners to iterate on output and to verify; R012's RCT shows why that matters, since tutor-style use left lasting gains while AI-written work left nothing. R014 teaches the habit, R012 supplies the evidence.
• R005 and R014: Both are free courses structured as a learning path. R014 targets AI users; R005 targets agent builders. Together they bracket a curriculum.
• R006 and R014: R006's harness vocabulary is what R014's dictionary stops short of, and the natural next lesson after 'what is an LLM'.
• R004 and R014 (weak): R004's goal-plus-context prompting is the adult version of R014's iterate-on-the-menu exercise.