The Council

High-stakes decisions have a common failure mode: you already know what you want to do, and you think about the decision in ways that confirm it. A good deliberation system breaks that pattern — not by being contrarian, but by forcing you to genuinely inhabit perspectives you wouldn't naturally reach. Karpathy's methodology does this well. The challenge was that the original requires multiple frontier models in parallel — API costs put it out of reach for personal decision-making. A single-model adaptation needed to recreate the deliberation quality without the multi-model architecture.

What I adapted from Karpathy: the core structure (multiple named advisors, independent deliberation, peer review, chairman synthesis), the isolation mechanic (each advisor reasons without seeing others' outputs until peer review), and the peer-review protocol (advisors evaluate argument quality rather than conclusions, preventing convergence on whoever argued most confidently). What I added or changed: Two original advisor profiles. Stakeholder explicitly models counterparties whose buy-in is load-bearing — decisions don't happen in isolation. Pathos surfaces emotional and ethical dimensions, licensed to split its answer (the emotional read and ethical read can point in different directions, and naming that tension is more useful than collapsing it). Dynamic seating with wildcards — the bench is assembled per session based on the decision type, with wildcard seats that can be filled by whatever lens the decision warrants. A two-phase Outsider mechanic — the Outsider runs twice: before deliberation begins (what does someone with no context see?) and after all advisors have argued (what is the council missing?). A chairman protocol with confidence rating — the verdict includes an explicit High/Medium/Low rating and a single recommended next step; Medium is not a failure, it's an honest read of genuine ambiguity. An HTML report artifact — chairman's verdict, advisor chips, agreement/clash summary, advisor highlights; readable standalone and shareable.

• Decision 1 — SGWS Sales Consultant promotion: Medium confidence. Two advisors independently identified that the SGWS employment-agreement check was a blocking condition. Verdict: resolve that prerequisite first, then re-council. Accurate.
• Decision 2 — HELIOS project scope (knowledge system vs. life hub): High confidence with a reframe. The decision pitted a long-horizon vision (grow the personal knowledge system into a full life hub — finance, calendar, task execution) against a finisher's instinct to complete a focused tool. Five of six advisors independently rejected the framing itself — the real choice wasn't "which scope" but which underlying need the project serves (proof-of-capability vs. daily life-utility), each pointing to a different answer. The Historian's anti-survivorship discipline surfaced the representative outcome — a graveyard of abandoned half-built "life-OS" projects, not the rare loud successes — and peer review caught that the project's own "demonstrable evidence" principle, applied back to the decision, independently selected the same path. Verdict: finish it as a focused knowledge system, ship it, then expand only from real need. I adopted the recommendation and narrowed the scope. The clearest demonstration to date of the system's purpose — it changed its own author's mind against a path he was drawn to.
• Decision 3 — Whether to write this case study: The Brief ran this decision and routed to "answer normally" before The Council was convened. The Council never ran — that's correct; the deliberation system shouldn't run on questions that don't need it.

Honest limitations. Small N — three full council sessions on real decisions, still untested on actively misleading framing or decisions with dominant emotional components. Single-user validation — all sessions ran on my decisions, calibrated against one person's thinking patterns. Isolation is simulated, not structural — in Karpathy's original, isolation is enforced architecturally (separate models literally can't see each other's outputs); in a single-model implementation it depends on prompt engineering, and a sufficiently long context could leak earlier advisor positions into later ones. No automated evaluation framework — whether the verdict was good is assessed retrospectively by real-world outcomes, which takes time. Attribution lineage requires active management — every public surface credits Karpathy explicitly; the discipline of separating adaptation from original work is the control mechanism. What's next: more test runs across decision types the council hasn't seen, a Council-lite mode (faster three-lens deliberation for lower-stakes decisions), an automated evaluation rubric for session quality, cross-session memory (if the same decision is re-councilled later, surface what changed), and public packaging as installable skill files. Technical details. Platform: Claude.ai Pro. Format: markdown-based skill file, installed locally. Trigger: natural language or direct slash command (/council). Output: in-conversation deliberation + single-file HTML report artifact. Integration: receives structured input from The Brief when available; runs standalone otherwise.