D182 - The compatibility record is the other half of the title file

decided · 2026-08-21

Layer::Repo has always existed in orbistoun-overrides, documented as "our shipped compatibility knowledge", and nothing ever wrote it. The type system described a compatibility database that did not exist.

A title file said what orbistoun sets for a title. Nothing said what orbistoun got.

One file, because two would disagree within a week

The measured half lives in the same file as the settings half. They are keyed by the same title, edited in the same session, and produced by the same run; two files would immediately disagree about which was current, and nothing could say which to believe.

Deliberately not merged. Resolved::merge layers settings per key, which is right for configuration and meaningless for a measurement - there is no sense in which a user's run "overrides" the repository's recorded one. They are facts about two different runs, and comparing them is the useful operation. merge iterates settings and compat explicitly, so the new field is excluded by construction rather than by remembering.

Derived, never typed

Every field comes off the trace. A compatibility database whose grades are hand-written drifts the moment somebody is optimistic, and the drift is invisible because there is nothing to check a hand-written grade against.

That also decided the vocabulary. "Playable", "in-game", "intro" would be aspirational fiction here; every rung is a phase the loader already distinguishes, so a grade is a transcription rather than an opinion.

The record is protected from the reward hack

Status::beats refuses to rank two results produced under different stub policies, and compat record refuses outright to record a run where unimplemented functions reported success - including the first one. That last clause was found by recording an entry: a contaminated first entry has nothing to be compared against, becomes the baseline, and then no honest run can ever beat it. A database carrying a best-ever nobody can reproduce is worse than an empty one.

The refusal is symmetric. It says "these cannot be compared", not "the bigger number wins", so a contaminated entry can be displaced by an honest one rather than only by another contaminated one.

The ladder stops at Entered, and the corpus is why

A Sustained rung for "ran to the limit without faulting" seemed obviously worth having. Populating the record disproved it in one table:

PPSA04263-app0   sustained    4 imports 91455278 calls   ran to the time limit
PPSA28061-app0   entered     47 imports      933 calls   image+0x43c4

Ninety-one million calls of four functions sorted above the run that reached forty-seven imports - the least informative result in the corpus at the top of the frontier list. Surviving is an outcome, not a distance, and it was already recorded as one. This is the same argument that had already been accepted for a Stopped rung and then not applied here; the corpus caught what the reasoning missed.

Within Entered, distance is distinct imports and then calls. Calls rank last because a guest spinning on one unimplemented function accumulates them without learning anything - which is precisely what PPSA04263 does.

What it says today

title	reach	imports	calls	standing	outcome
PPSA28061	entered	47	933	85%	image+0x43c4
PPSA02664	entered	19	53	53%	the guest called abort
PPSA03416	entered	19	53	53%	the guest called abort
PPSA25872	entered	14	1735	100%	image+0x7b591e
PPSA21564	entered	13	131	91%	image+0x70a932c
PPSA04263	entered	4	91455278	100%	ran to the time limit

The twin abort at exactly 19 imports and 53 calls is visible at a glance, which no report about a single title could show.

Titles cannot travel; everything about them can

titles/ holds guest material and is never tracked. compat/ holds identifiers, settings and numbers, and is always tracked. That split is what lets a finding travel when the title cannot - a contributor's result is checkable against yours without either of you having the other's material, provided the entry carries the build and the policy that produced it. Which is why it does.