Soft → Hard Migration

Essay 8.4 — From Apprentice to Architect, Part 4 of 9.

Essay 8.3 closed the brain inventory at three months — a small brain over a large knowledge layer with a narrow memory. That shape is the outcome. The mechanism that produces it is the migration of behavioral controls from soft form to hard form within the plugins themselves. This essay opens that mechanism. ^ⓘ

The graduation through the stages of the job-maturation arc (Essay 8.2) — deep single-cycle, to multi-cycle .md plan, to multi-cycle .yaml plan, to plugin form of a job — is only one of the patterns by which the seed agent grows. The other — equally important — is the migration of behavioral controls from soft form to hard form within the plugins themselves. ^ⓘ

The cleanest concrete migration in the prototype’s history is the multiplier sentinel — a behavioral control that started as a coaching voice, was measured to be failing under load, and earned hardening into a PreToolUse block (a Claude Code hook event that fires before any tool is called, capable of refusing the call). ^ⓘ

A lawyer’s seed running matter-intake jobs will see the same migration shape applied to its own domain: a [CONFLICT-CHECK] voice fires probabilistically at the start of each new-matter cycle, the operator notices it being skipped under pressure, the operator (with the seed) hardens it into a PreToolUse block that refuses tool calls until the check returns clear. Same arc; different concern. The lawyer’s seed earns hardening evidence the same way the prototype did — measurement first, then friction. ^ⓘ

The Multiplier Sentinel — A Worked Migration

In the early cycles, every phase entry shipped a single coaching voice (entry-set-multiplier) asking the agent to please set the phase’s multiplier before starting work. The voice fired probabilistically into the LLM’s context. The agent read it and, under cognitive load, often ignored it. Several cycles in, the data was clear: the ratio of voice-fires to actual-multiplier-sets had stopped converging. The voice was not holding. ^ⓘ

The hardening landed during a dedicated maintenance cycle. Every phase’s entry now initializes the multiplier to literal zero — a sentinel value. A PreToolUse guard reads the multiplier on every tool call; if it is zero, every tool is blocked. The agent’s only available move is to call <phase>.sh set-multiplier <job> <value> with a value in a small allowed set. The coaching voice retired. In its place, two new voices took the load: entry-set-multiplier-pre-set informs the agent before the lock; multiplier-zero-block refuses tool calls when the gate fires. What was once probabilistic judgment is now mechanism. ^ⓘ

The Cost Ladder — Voice, Hook, Plugin, Template

The pattern beyond this case — what this essay calls the cost ladder — is consistent with the Lock-13 soft-vs-hard discipline that Essay 7.3 names the over-engineering veto. In this prototype, new behavioral concerns start as voice — soft, probabilistic, ignorable. If measurement shows the voice failing to hold, the operator climbs to hook in an existing plugin — a PreToolUse guard inside the plugin whose concern the pattern belongs to. If the pattern needs its own state or crosses an existing plugin’s boundary, it earns a new plugin. Your seed can enter the ladder at any tier if prior evidence already justifies the cost — the order is a default, not a mandate. The prototype codifies this restraint as a named lock discipline (the over-engineering veto): no new hard gate hardens before measured cycles demonstrate the soft form is failing. ^ⓘ

The deepest migration is the meta-pattern fossilizing into the kit itself. The multiplier sentinel didn’t just become a hook in phasic_system; it became part of the template every new phase plugin inherits. Brain_guard’s cycle-1 universal disciplines (verify-100%-before-bonus, subagent-spot-check, condense-not-deletion, job.sh create-only in multi-cycle) made the same trip — from one cycle’s lesson to a rule every cycle now obeys, codified into .claude/knowledge/opevc/ and inherited by every new job. ^ⓘ

Two Axes, Same Shape

This soft-to-hard migration mirrors the job-maturation arc one level up. Each stage of the maturation arc maps to a hardening tier. The collaborative learning of an early-stage job becomes the codified .md plan once the work repeats; the plan itself migrates into a .yaml injected at phase entry once its shape stabilizes; the most-repeated patterns harden further into plugin form. Every step is voluntary, and every step trades inspectability for friction. The brain grows along both axes simultaneously: jobs mature upward; controls migrate inward. The limit on both axes is the same: friction plus evidence, not mathematical impossibility — a careless operator can ship a hook without tests; a missing hook registration is silently inert. The architecture makes the careful path easier than the careless one; it does not refuse the careless path. ^ⓘ

A pattern travels from voice to hook to template, and the cost ladder keeps the brain honest about which controls have earned the cost of hardening. The next essay opens the inverse question: which limits look hard but are actually CONDENSE discipline, and which earn their gates.

Essay 8.4 — From Apprentice to Architect, Part 4 of 9.

Previous: Essay 8.3 — What Lives in the Brain After Three Months — the prototype as ground-truth inventory. Next: Essay 8.5 — What’s Enforced vs What’s Discipline — the honest accounting of size caps.