A running log of every research paper, blog post, podcast and field note that shapes how the Castle thinks. Updated as I read.
Three startups this week described the same failure mode: their model gets smarter every quarter, their workflow doesn't. The bottleneck is never the LLM — it's the operator who refuses to write down what just happened. Crystallise or perish.
A 3B parameter model steered with CAST tokens matches a 70B baseline on closed-domain retrieval — at 1/40th the inference cost. The interesting bit isn't the benchmark; it's that the steering vectors are auditable. You can read what the model is being told to attend to.
A pilot that ships without a Proof-of-Thought trail is not a pilot — it's a demo. Demos don't compound. The post catalogs six failure modes I see in enterprise rollouts, and the three the Castle protocol is designed to prevent.
Swyx interviews three operators running enterprise-scale Action Graphs. Best bit at 38:20 — the triage nurse who modelled her hospital's after-hours pattern and reduced misroutes by 41%. Lo-fi. Auditable. Unglamorous. Exactly the loop the Castle is built for.
A model is a snapshot. A graph is a memory. Build the memory.
Briefing, mission, debrief, crystallization, retirement. Five phases that should appear in every AIOps team's weekly cadence. The post argues that omitting retirement is the single largest source of graph rot in enterprise deployments.
Proposes a Merkle-anchored consent log that survives audits across regulatory regimes. The clever piece is the "default deny on missing consent" — agents pause rather than fall back to a permissive policy. This is exactly the Bridge Protocol Allura describes in the charter.
One: the retrieval index hasn't been pruned in 90 days. Two: the highest-frequency node hasn't been touched in 60. Three: nobody on the team can name the most recent retirement. When all three are true, the graph is no longer compounding — it's accumulating. Treat as a brownfield migration, not an incremental update.
Recommended without comment. The opening paragraph reframes the entire AGI debate as a question of execution latency, not capability. Required reading before Cohort III orientation.
Walks through the Proof-of-Thought protocol the Castle uses to preserve agent reasoning across every step of a workflow. Includes a worked example from a real triage pod and the audit trail it produced after a regulatory review.
The world bends to the one that moves — not the one with the best model.
A formal treatment of when to retire nodes from a retrieval graph. The 'staleness × confidence × frequency' policy maps cleanly onto Coran's Maintenance Code. Cohort II should pilot this on Pod 09 next quarter.
The skill is necessary but vastly insufficient. Candidates who pitched prompt-engineering as their core competency uniformly under-delivered when asked to design a retrieval policy, a retirement schedule, or an audit trail. Recalibrate the screen.