The thing that clicked today: a composite score is a great way to feel ready and a terrible way to be ready, because a strong number in one place quietly pays for a weak one somewhere else.
Waypoint is a study and credential workspace. It answers "what should I work on next" from what I have actually recorded, rather than from a plan I wrote optimistically in advance. Today's work was about making it harder to lie to myself.
Built / shipped
A canonical registry for every certification and exam sitting on the path, machine-readable, with each official vendor objective document SHA-256 verified against what the compiler was built from. The vendor objectives are the permanent scope spine. Books and other sources enrich that spine; they cannot redefine it, because the exam does not care what a book thinks the syllabus is.
Content releases are bound to that registry and fail closed on scope or source drift. If a source hash no longer matches, the release does not quietly proceed with the new file, it stops.
Readiness became ten separate evidence gates, never combined into one figure. Objective coverage, recent practice average, open gaps, retention due, hands-on work and the rest each pass or fail on their own terms. The honest current state for the one fully published pack is "building evidence", three gates of ten. That is a much less satisfying number than a percentage and a much more useful one, because it tells you which thing to go and do.
The plan also stopped pretending. It now shows the target date beside the current projection and derives the pace the gap implies, instead of asserting a completion date it cannot know.
Problems & fixes
Only one certification has a real published content pack. The other five had domain scaffolds, which look identical to real content in a progress bar. That is exactly the sort of thing that inflates a composite score, so scaffolds are now explicitly marked and cannot imply learning progress.
Two surfaces hardcoded a daily study target that no longer matched the live weekly goal, so the app was quietly grading me against a number I had changed weeks earlier. Both now derive from the goal, or show an explicit temporary override.
Day boundaries, streaks and schedule windows were each computing "today" slightly differently, which produced a streak that was correct in one panel and wrong in another after about seven in the evening. All of them now route through one clock module, with a regression test that runs after that hour specifically.
The public mirror of the progress page had drifted from the authenticated one, because they were two components that had to be updated in tandem and had not been. One shared component now backs both, so the mirror cannot omit a field the real page gained.
Decisions
Readiness is never composited. If someone wants a single number, the honest answer is the count of gates passed, which is a number that gets worse when you skip things.
The coaching layer retrieves from guides and reviews only. Question prompts, answer explanations and the whole practice bank are stripped before inference, and citations that are not present in the supplied packet are rejected. A study assistant that has read the answer key is not assessing you.
Learned
Most of this session was not adding capability, it was removing places where the app could tell me a comfortable story. The version that reports three of ten is more useful than the version that reported a percentage in the seventies, and they describe the same underlying evidence.
Still open / next
Content packs for the remaining five certifications, under the same vendor-authority and fail-closed rules that produced the first one.