Cadence

[ architecture ] · § fig. 2, inked

[ decisions ] · § as filed

[ validation ] · § the receipts

ci rows link the public run · repo pins are the exact commits verified 2026-08.

[ corrections ] · § the register

• erratum · 2026-07-30

  The IDOR count was stale at seven. An eighth of exactly the same class — TagService reading and deleting by id alone — was found and fixed on 2026-07-27 at cadence @ 75180a3, three days after the other seven were filed, and this file went on saying seven. It now says eight across nine endpoints, and the eighth carries its own pin rather than moving the pin the first seven were read at: a receipt names the commit its finding was made at, and re-pointing CADENCE_SHA forward would silently restate seven older findings as if they had been re-verified somewhere nobody re-read them. The eighth also brought the sharpest finding in the set, and it is the reason this correction is worth more than the number: the regression test that was supposed to cover that route asserted WHERE id = $1 with the id as its only parameter. It was pinning the vulnerable query in place and reporting green. A test can certify a bug. This very page learned the same lesson from the other end on the same day — a Playwright spec here was requiring the page to repeat a false sentence about CI, and retracting the sentence broke the spec that was defending it.

• erratum · 2026-07-30

  Three statements on this page were false, and every one of them was an UNDER-claim — this file was disowning work it had actually done. It said the 11 isolation tests “do not run in ordinary CI”, that they “are not in CI”, and that “the repo’s own CI is red on main right now”. Checked against the run logs rather than against memory: the workflow at cadence @ 54c79e0 provisions a postgres:16 service and sets RLS_TEST_PG_ADMIN_URL, so the skip guard never fires there, and run 30133037462 records ✓ lib/__tests__/rls.postgres.test.ts (11 tests) 232ms on 2026-07-24. Main has been green since 2026-07-23. The likely origin is a copy: the identical caveat is TRUE of Applied, whose backend workflow really does provision no database, and it appears to have been carried across to a repo whose CI had since gained one. A false disclaimer is the same broken receipt as a false boast, and arguably the worse one, because a reader has no reason to doubt a claim that costs its author something. What has NOT changed is the standing that matters: the migration is still hand-run, still not applied in production, and receipt 05 still says so.

• note · 2026-07-26

  The repository was renamed yadava5/taskflow-calendar ⟶ yadava5/cadence, and this file now carries two pins on purpose. The ledger above and receipts 04–10 name cadence @ 54c79e0, the current public head — where the code lives and where the isolation work landed. Receipts 01–03 stay at taskflow-calendar @ 69a59e7, because that is the commit the 1,145-test count was measured at, and re-pinning a number to a commit nobody re-ran it at turns a measurement into a guess. GitHub redirects the old paths; both resolve.

• erratum · 2026-07-26

  The workspace outcome row linked taskflow-calendar-ashy.vercel.app. That deploy still answers, so nothing 404’d — but it is not the app any more, and the file was printing one host in its meta ledger and another one two sections down. The row now links usecadenceapp.vercel.app, the same host as the rest of the file.

• note · 2026-07-30

  Receipts 01–03 now read cadence @ 69a59e7 rather than taskflow-calendar @ 69a59e7. The pin is unchanged and that is the whole point of the 2026-07-26 note above: 69a59e7 is still the commit the 1,145-test count was measured at, and nothing has been re-pinned. What changed is only the repository’s name in the label, because the old name now survives on a GitHub redirect and a redirect is not a permalink — it stops resolving the moment any repo called taskflow-calendar exists under this account again. A receipt on a page arguing that every claim terminates at an artifact you can open should not depend on that. The canonical path was fetched and returned 200 before the label was changed.

• erratum · 2026-08-02

  The note above was not true when it was written. Receipts 01–03, and receipts 11 and 12 with them, went on reading “taskflow-calendar @ 69a59e7” for three more days — the built page rendered the retired name eighteen times — while the register told the reader the labels had already been changed. They read cadence now. This is recorded as an erratum rather than repaired quietly because the error was not a stale name: it was this register claiming a repair it had not made, which is the one failure that costs a corrections register the thing it exists for.

• erratum · 2026-08-02

  The receipt itself now reads 1,159 — see the erratum below, which corrects this entry's own number. The 2026-07-31 note above already reported that re-run — 635 frontend + 533 backend at 932625e — but only the note was written: the receipt, the architecture summary, the diagram’s own node label and the /evidence index all went on saying 1,145 for another two days, so the page argued with itself and the reader had to find the erratum to learn which number was current. Re-measured independently on 2026-08-02 and confirmed identical: 1,168 passing, 11 skipped. The number and its commit move together, so the row is re-pinned to 932625e; receipts 02 and 03 keep 69a59e7, which is still where those source audits were done. This is the second erratum in this register about a correction that was recorded without being applied, and the pair is why the audit now ends by widening the drift gate instead of only fixing the values.

• erratum · 2026-08-02

  1,168 became 1,159, and the reason is a mistake in the correction above rather than in the code. That entry re-pinned the count to 932625e — which is an UNPUSHED local commit. The source URL 404’d, and the number, though correctly measured, could not be reproduced by anyone: the tree it was taken at does not exist publicly. Re-measured at the public head 8eee84e instead — 635 frontend + 524 backend = 1,159 passing, 11 skipped. The nine-test difference IS 932625e, the fix for nine endpoints that never authenticated; those tests land here the moment that commit is pushed, and the number goes back up on its own. Recorded rather than quietly repaired because the rule this file states — a count and its commit are one fact — has a second half it did not say out loud: the commit has to be one a reader can open.

• erratum · 2026-08-07

  This is the third drift of the same number, and this time it shipped as three different values at once. The architecture figure’s node label read 1,159 — twice on the built page — the receipt beside it read 1,185, and the /evidence index read 1,179. Every gate was green, and the erratum above was itself stale: it says “the receipt itself now reads 1,159” while the receipt had moved to 1,185 and left the figure behind. Re-measured at cadence main dbabc74, CI run 31222343049, all five jobs green: 635 frontend across 58 files + 550 backend across 25 = 1,185 passing, 0 skipped. All three surfaces now read that, and the two pins — this file’s CADENCE_SUITE_SHA and proofManifest’s — moved to dbabc74 with it. The register predicted this: the 2026-08-02 pair ends by saying the audit “widens the drift gate instead of only fixing the values”, and the widening was only half done. check-figures.mjs’s Cadence · suite entry had no manifest: key, so it could not see /evidence at all, and it never bound the node label — the same two holes that entry’s Applied twin had already documented and closed for itself. Both are closed for Cadence now, so the next drift is a red gate rather than a reader’s discovery.

• erratum · 2026-08-08

  1,185 became 1,186 within a day of the entry above, and the one extra test is worth more than the four surfaces it moved. Exercising the deployed app — signed in as the demo account, against the real API — found GET /api/tags answering 500 for every user: TagService selected t."createdAt" and t."updatedAt" from a table that has five columns and never had those two. git log dates the over-select to the commit that added the service, so tags have been unreachable through the API for the life of the feature. Its 28 unit tests passed throughout, and still pass with the bug deliberately reinstated, because their fixtures are hand-written objects that never reach Postgres — the same shape as the tagged-task failure this register already carries. The fix ships with a regression test that calls the service against the real schema in a real Postgres, which is the +1. Recorded here rather than folded into the entry above because the number the page published yesterday is not the number it publishes today, and a register that quietly restates itself is the failure it exists to prevent.

[ appendix ] · § plates & artifacts

fig. 3 — Production calendar, demo account session.source: production interior — usecadenceapp.vercel.app, demo account · 2026-08 · 2026-08 · open to inspect ⟶

real-screenshot

Production calendar, demo account session

open original ↗ ✕

source: production interior — usecadenceapp.vercel.app, demo account · 2026-08 · date: 2026-08 · boundary: demo account state — no real user data