Method — a library of good procedure

How to run hard, agent-driven projects without drifting. Process is craft, not science — these pages are living playbooks of warning signs and moves, accreting one hard-won lesson at a time. Distinct from career/ (job-search is one downstream consumer of this, not the parent) and research/ (topic exploration, not how-we-work).

Links: Praxis: Agent Teams (the economic theory of why agent teams work), Transpilation as a Grounding Strategy, The LLM Grounding Problem

Entries

• The Anchor Method — the spine: drift is the enemy, the gate is the anti-drift anchor that proves grounding; install the strongest anchor your current grounding allows and upgrade it toward a hard gate as grounding crystallizes; the director owns recalibration (a convergence gate, via divergent-approach tournaments). Born from the NA1 decompiler; tested against the projects that drifted.
• Specimen & Thesis — The Two-Altitude Ledger — the developed form of the Anchor Method’s “seam discipline”: vault work lives at two altitudes (the dated, source-anchored specimen vs the portable thesis it feeds), kept in separate pages and joined by a two-way ledger. The downward half (thesis→its specimens) rots because hubs predate their evidence; the cures are promote, don’t trap and memory is not the artifact. Discovered auditing the debate corpus; tested against Monopoly/MoO/MoM (all fail).
• Author Web, Derive Hierarchy — the organizing law for generative work: exploration discovers structure you can’t pre-file, so keep the substrate write-optimized (web) and derive the read-optimized hierarchy (indexes, ledgers, audit) with tooling, after the shape emerges. A tree severs the lateral edges a sprawling project radiates; the web holds them. The difficulty of a big vault is the cost of compounding, paid down by deriving — not by suppressing exploration. Specs the derive-the-ledger generator.

• The Rosetta Stone Method — how to learn unfamiliar hardware: don’t read the chip manuals first — find a portable program you already understand running on it and let the known semantics decode the unknown silicon (find the interpreter, anchor the opcode table byte-for-byte, follow a syscall into the chip). Inverts learn-hardware-first; the Anchor Method applied to learning — the known-semantics crib is the anchor, and a cross-platform sibling is a free oracle. Depends on someone else’s portability discipline (a portable VM, not hand-ported native code); KOEI is the exemplar, proven across 6502/65C816. Generalizes to any cross-platform engine (SCUMM, Z-machine, AGI).

• Building Swarms — Token Rules as Architecture — the construction end: Nate B. Jones’s 15 token rules are single-agent hygiene, but at fan-out every per-agent overhead is paid N times (55k of tool definitions × 12 agents = 660k before any work), which promotes habit into architecture. Rule-by-rule translation to swarm form; rule 3 becomes the definition of an agent boundary, rule 4 (carry the answer, not the argument) is load-bearing, rule 10 (scope tools per agent) is the highest-leverage change in a wide fan-out. Diagnoses the whole-context-copied-to-every-agent anti-pattern as the symptom of having no substrate — with nothing durable to point at you must ship the history, making the swarm maximally expensive and maximally opaque at once. Its governance twin is The Substrate Is the Governing Mechanism.

• Watching the Solver Learn — the observation end of the anti-drift problem: when you iterate a solver, convergence speed is set by whether you can see what it is currently doing, because reward/objective bugs are behavioural — they crash nothing, fail no test, and yield a confidently-useless agent while every metric climbs. Prescribes the layer the vault’s solver projects lack (a live view of the artifact acting + training telemetry: smoothed and raw objective, a confidence signal, a bounded cost metric, an event log that doesn’t interrupt the run), plus fixed-size perception, action rate as a planning-horizon choice, and run the curriculum with a from-scratch control. Carries one evaluation law: a benchmark the agent can shrink is not measuring the difficulty on its label. Worked from Pezzza’s PPO gladiator (C++/SFML).

Where this sits in the web

This is part of the larger “how computer science & engineering is changing under these tools” theme — the same neighborhood as LLM grounding, principled LLM code, the hierarchy-vs-web question, and the project case studies. In a web vault a node’s identity is its links, not its folder — so reach these from wherever the thread leads.

Tags

agent-teams · methodology