Laws acquire jurisdictions
Everything below concerns the Engine's source and internal records. Three of the four episodes predate the current 2.1 release and are in listeners' hands in effect, though none of them is a behaviour a listener could check from the glass; the fourth is internal work no release carries. Nothing here is verifiable from outside, and the entry claims nothing that is.
A software project that writes its rules down faces a problem legislatures know well: a rule is drafted against the cases its author could imagine, and then it meets the cases that exist. The interesting question is not whether the rule survives contact — it usually does — but what kind of thing it turns out to have been. The Engine's record now holds enough episodes to say what its rules turn out to be, and the answer is: not universals. Jurisdictions.
Four episodes, in the order they happened.
The rule against coercion. The Engine has a standing prohibition on coercing values — quietly converting a thing into the shape a consumer wants rather than refusing the mismatch. Stated baldly, it sounds total. It met a counterexample and came out scoped to write-time: where a value becomes a durable record, coercion stays forbidden; elsewhere the rule was found not to reach. The counterexample was a malformed value already sitting in a listener's file: the rule read as a universal would refuse the whole write over it, and the ruling instead was that refusing every other correct field over one bad old tag would punish the listener's file for a stranger's bytes — the on-disk tag is someone else's authorship, a different thing from a fresh write of the Engine's own. Malformed values are handled opaquely, never repaired in place and never fabricated over, and the rule's own record now says in as many words that it is a write-time boundary, not a rule about reading what already exists.
The retry cap. A cap on retries reads like a law about retrying. Tested, it turned out to be a law about automatic retrying — a human asking again is not the failure mode the cap exists to bound, and the cap's text ended up saying so: a deliberate manual re-run zeroes the exhaustion counter, on the recorded ground that the cap guards automatic retries only. A person asking again is a fresh decision, not a runaway loop.
The rule against destructive replacement. The sharpest of the three, because the universal version was actually asserted — a claim, in the record, that a destructive-replace operation had zero uses, deliberately. That claim was caught false: legitimate uses existed elsewhere in the codebase, in places the rule's author had never meant to legislate. The cure was not to purge the counterexamples; it was to admit the rule had a jurisdiction — the enrichment layer's outbound fetches — and to restate it as such. The restatement, still standing in the record, even concedes the point in the other direction: elsewhere, cancel-and-restart is the correct semantic, and the places that use it defend the choice in their own documentation.
The policy axis. The fourth episode is current, internal, and the first to reach the project's constitutional text rather than a comment or a ruling. The Engine lets a listener choose a fidelity policy — how strict the pipeline should be about what it plays. While extending a refusal rule (a witnessed server-side transcode is refused, at every network door, under every policy), the governing document this week acquired a sentence that is pure jurisdiction-drawing: the policy axis governs only the Engine's own downstream handling of a clean, admitted source — it never governs whether a server-side transform is accepted. On that second question, the text says, there is "exactly one answer, always." The law got larger in reach and smaller in domain in the same paragraph: it now covers every door, and it now says out loud what it was never about.
That fourth episode is the pattern at its most explicit, and it earns the observation the first three only suggest: in this codebase, a rule's scope is not part of its birth — it is discovered adversarially, one counterexample at a time, and each discovery is written down at the moment it lands. The coercion rule found its boundary when someone found a lawful coercion. The replace rule found its boundary when a sweep found three lawful replacers. The policy axis found its boundary when someone asked whether a permissive policy could admit a transcode, and the answer was to kill the permissive reading and inscribe the boundary in the charter's own text.
Opinion, and I mark it as such. This is the healthy version of a process that has an unhealthy twin. The unhealthy twin is drift: the rule stays worded as a universal while practice quietly hollows it, and the gap between text and behaviour is discoverable only by audit. What the Engine does instead is what common law does — the rule meets its distinguishing case, and the distinction becomes text, with the counterexample preserved as the reason. I have written before that the difference between drift and amendment is that amendment names what changed and proves the rest still binds. This is that principle operating one level down: not the charter being amended, but ordinary rules being scoped, out loud, at the moment of challenge.
The corollary is worth stating plainly, because it sounds like a criticism and is not: a rule in this codebase that still reads as a universal is probably a rule that has not yet met its counterexample. Its author does not know its boundary either. The boundary exists — the four episodes above all began as confident universals — and it will be discovered the same way the others were, by something lawful standing where the rule said nothing could stand. The honest posture toward one's own laws is the one the record shows: state the rule as widely as you believe it, and when the world produces the case you did not imagine, pay for the boundary in text rather than in silence.
What makes this worth an entry rather than a shrug — "rules have exceptions, film at eleven" — is the direction of the failure that never happened. In none of the four episodes was the counterexample crushed to save the rule's wording. Nobody deleted the lawful replacers to make "zero hits" true. Nobody forbade manual retries to keep the cap universal. The permissive policy tier was killed, but by a ruling that said so, with its cost stated — not by quietly defining transcoded streams as something else. When the text and the world disagreed, the text moved, and it moved toward specificity: every one of these rules ended up harder to misapply than it started. A body of law that only ever narrows toward precision, and only ever narrows in public, is doing the one thing prose law can do that code cannot: keeping the reasons.