docs: new Compass page + resync Plan-Protocol, Autonomy and Configuration with the code
The wiki was last synced on 2026-07-11 and described two things that are no longer
true: a compiled service contract as the mechanism that keeps a weak model honest
about an API, and a skill ladder promoted "by successful reuse".
Compass (new page). philont's answer to soul.md, and the authorship runs the other
way: not a persona handed to the agent, but the owner's declaration of where their
second mind should point. Documented by what it mechanically does — the [min,max]
clamp on live trait tuning, focus areas seeded as pursuit rows (prompt injection
alone never reaches the background drivers, which run without the main system
prompt), and the prose injected verbatim — including the design choices worth
knowing: lenient parsing so a hand-written file cannot break startup, and no
backdating so declaring five focus areas does not fire five sessions at once.
Honest scope: no hot reload, no learned focus proposals, and it grants no
permission of its own.
Plan-Protocol §7 — the mechanism loop. The page only covered the protocol the model
walks; the state machine that owns a guide-driven turn was undocumented. The section
explains why two weak models reviewing each other is not automatically worth more
than one model trying twice — the first version wasn't, and production oscillated —
and what the mechanism had to remember for it to become a ratchet. Also records why
the compiled spec was switched off: a wrong call gets corrected by the service, a
wrong contract has nothing to correct it, because the contract IS the corrector.
Autonomy — verify-then-condense. The audit that found none of the three learning
sub-loops closing, the paradigm root cause (online per-artifact attribution), the
judge's grounding-tool rule and its shadow gate, Phases 0 and 3a, and the three
independent defects that were each alone enough to cancel the loop — plus the
general lesson that a learning loop is a product of AND-gates, where the symptom of
a silent veto is not an error but nothing happening.
Supporting: Memory-System corrects what "confirmed" used to mean; Configuration
documents AUX_LLM_* (and that leaving it unset removes the plan reviewer),
compass, AGENT_LANGUAGE and the new flags, and drops the stale claim that
VISION_LLM_* is unwired; Why-a-Cheap-Model gains the failure mode the review
ratchet catches; Selfhood-Engineering §3b, Home and _Sidebar link the new page.