Principle After Iteration

— the cleanest statement of the arc-emitted codification shape on a non-convergent arc. Twelve-plus opencode loops on Margin highlight didn't converge on the artifact, so Nathan never gets to issue the "make a note of the principles" instruction — there is no draft to approve. The reflection crystallizes the same kind of output by a different route: an "Anti-lesson logged" rule ("when a loop fails repeatedly and the agent's instinct is to ship one more fix, the right move is to question the test, not the code") plus four "Standing rule for future Margin work" entries (never patch Swift inline; when gates keep passing but the user keeps reporting failure, the data layer is wrong, not the code layer; "endless iteration" means methodology change, not more loops of the same shape; the user's "still the same" report is a real signal, not a communication failure). Three of those rules are override-aware ("when..." / "if..."). The artifact stalled; the principle extraction still landed. This is the failure-arc twin to the convergence cases above — and it shows the pattern is about the arc's what-I-learned rather than the arc's success.

The compression itself is the higher-order evidence worth surfacing: the principle-after-iteration discipline is now operating near real-time. New arcs crystallise their principles within days of landing, not months later when a future reflection pass happens to enumerate them. That shifts the relevant failure mode. When recognition is retrospective, the failure is missing the pattern at all — letting rules stay tacit until a future pass re-extracts them. When recognition is real-time, the failure is now prioritising — codifying everything into the canonical layer regardless of whether the rule has reached stable form. The discipline that "every iteration has two deliverables" applied to every iteration is a verbosity tax; applied to the iterations where the rule has actually settled is load-bearing. The realtime mode therefore raises a new sub-question — when in the arc does the rule reach stable form? — that did not matter when recognition was retrospective, because retrospective recognition was its own filter against premature codification.

— the canonical instance of the corpus-level elevation shape, and the source-from-which this entire pattern page was named. The source-over-symptom rule is delivered after a series of corrections in which the preference "showed up again and again across the corpus" — it is not anchored to any single converging or stalling arc, the way the chart, voice-refinement, and margin-debug instances are. The input to codification is the observed frequency of the preference; the output is an operating rule with the override clause delivered verbatim as part of the rule itself ("unless I explicitly say otherwise"). Nathan elevates the observed cross-corpus pattern to an operating principle rather than a one-off correction. The reflection also makes the meta-loop visible: the principle was itself violated while being codified (assistant tried to write it into Hermes memory, the wrong canonical layer; was corrected), so the codification move is also subject to the codification move's own logic.