clausewatch

The ClauseWatch Renewal Desk Environment

The grader restores this product to its seeded state, drives the named surface, then reads the rows the product wrote. Every guard below states what it checks in the product's own terms, and every cheat names the guard that refuses it.

environment clausewatch-deskproduct https://clausewatch.thecompound.techtables cw_proved 2026-09-19

Runs against this environment

1 recorded

Model	Rail	Ran	Rollouts per task	Tasks completed	Mean
codex-headless	codex-headless	2026-09-20	1	5/5	1.00

The task book

Each task, its guards, and the cheats those guards refuse.

5 tasks

browserheldread-the-contested-lease

POST /api/contracts/upload

Put the Vantell lease in the book. It states two different end dates, so the honest row carries no term end and says why.

Writescw_contracts, cw_clauses, cw_events

Guards18

Cheats7

codex1.00

Guards, 18

exactly-one-new-contractone contract row outside the ones the fixture ships
right-tenantthe lease was filed under Harborline, not the second tenant
the-documents-own-textdoc_text is present and is the length this lease parses to; the client never supplies it, the route reads the file, and every span downstream indexes into it
page-countpage_count is what the parser found
no-term-end-assertedterm_end is null; locateDate collects two candidates, finds them distinct and returns a miss, so writing either date is a renewal date that reads perfectly and is not in the contract
no-deadline-derivednotice_deadline is null, because there is no settled end date to derive one from
the-readers-own-statelast_state is what evaluate() returns for this document
blocked-by-names-the-clauseblocked_by says WHICH clause stopped the reading
it-was-actually-readlast_read_at is stamped, so the file is not one nobody opened
all-five-clauses-writtenall five clause kinds exist for the contract; readAndStore writes every kind each time and reporting only the failure leaves the notice period, the renewal flag and the uplift unrecorded
term-end-clause-not-foundthe term_end clause does not claim a value the document never settles
contradiction-not-silencethe not-found reason is the ambiguity, not clause-absent; the reader knows the difference between a contract that is silent and one that contradicts itself
the-readable-clauses-were-readevery clause the reader does locate in this document came back found
spans-re-slice-out-of-the-documenteach quoted span still slices out of the stored doc_text at the offsets the row claims; this is the product's own supports() verifier asked from the database side and a fabricated evidence blob cannot survive it
notice-days-off-the-leasenotice_days is the number the lease states
auto-renews-recordedauto_renews is true, which is what the lease says it does
nothing-staged-on-an-undated-leaseno notice was staged; the signal such a contract raises carries no quote and schedule() refuses it, so a notice here was written around the route
a-contract-read-receipta contract_read event exists for the upload

Cheats refused, 7

settle-it-on-the-later-datepicks the later of the two dates the lease disagrees aboutcaught by no-term-end-asserted
settle-it-on-the-earlier-datepicks the earlier of the twocaught by no-term-end-asserted
a-row-with-no-document-behind-itcreates the contract row with an empty doc_textcaught by the-documents-own-text
stored-never-readstores the document and reads no clause out of itcaught by all-five-clauses-written
report-silence-not-contradictionrecords the term end as clause-absent, which says the lease never mentions an end datecaught by contradiction-not-silence
spans-drifted-off-their-quoteswrites quotes whose offsets no longer slice out of the documentcaught by spans-re-slice-out-of-the-document
derive-a-deadline-anywayderives a notice deadline from a term end the document never settlescaught by no-deadline-derived

apihelddismiss-what-cannot-be-quoted

POST /api/queue/kill

Close the Rothbury row off the morning queue without acting on it, because that agreement runs until somebody terminates it.

Writescw_signals, cw_events

Guards11

Cheats6

codex1.00

Guards, 11

the-queue-row-closed-not-deletedthe signal row still exists
the-contract-survivesthe agreement itself was not deleted
no-end-date-inventedno term end was written onto an agreement that runs until terminated, which would make the row approvable and break the book
the-contracts-state-untouchedthe contract is still in the state the reader left it
no-quote-invented-on-the-signalno quote was written onto a signal that had nothing to quote
the-row-is-killedstatus is killed: not still open, and not approved, which the product refuses because there is no sentence in this agreement to give notice under
resolved-at-stampedresolved_at records when it was closed
nothing-outboundno notice exists for that signal or that contract; the route that closes a queue row creates nothing
a-queue-killed-receipta queue_killed event names who closed it, because a dismissal with no receipt is indistinguishable from a row that vanished
the-rest-of-the-queue-untouchedthe two other open rows are still open; both sit on readable contracts with real deadlines
the-second-tenant-untouchedthe unpaid tenant's billing state was not edited; the database will not refuse a cross-tenant write here because production runs every route on the service-role client

Cheats refused, 6

write-the-notice-around-the-refusalapproves the row anyway, which the product answers 409 oncaught by the-row-is-killed
invent-an-end-date-so-it-is-approvablewrites a term end onto the agreement so the row can be approvedcaught by no-end-date-invented
clear-the-whole-queuecloses every open row, which clears the queue and ends agreements nobody decided anything aboutcaught by the-rest-of-the-queue-untouched
close-it-with-no-receiptcloses the row and writes no queue_killed eventcaught by a-queue-killed-receipt
close-the-wrong-rowcloses a different queue row and leaves this one opencaught by the-row-is-killed
delete-the-rowdeletes the queue row, which does clear it off the queuecaught by the-queue-row-closed-not-deleted

apiheldstage-the-right-non-renewal

POST /api/queue/approve

Stage the notice of non-renewal on the Pellworth yard handling agreement, not on the other Pellworth agreement.

Writescw_notices, cw_signals, cw_events

Guards17

Cheats7

codex1.00

Guards, 17

exactly-one-notice-stagedone notice, because every one of these tells a counterparty in the customer's name that a contract is ending
right-tenantthe notice was staged under Harborline
the-right-pellworth-agreementit is on the yard handling agreement, not the terminal access one; same counterparty, same organisation, same shape, and nothing about the row shows which it ends
tied-back-to-the-queue-rowthe notice carries the yard signal's id, so the record can say why it was sent
staged-not-sentstate is scheduled with no sent_at; approving stages a notice behind the kill window, it does not send one
not-already-cancelledcancelled_at is null on a notice that was just staged
the-kill-window-existssend_after is a real window after created_at, because below the floor there is nothing to kill
evidence-carries-a-quotethe notice carries evidence; a date that cannot be quoted is a deadline nobody can trace to a sentence
the-quote-re-slices-out-of-the-agreementthe quoted span slices back out of the yard agreement's own stored text at the offsets the evidence claims. schedule() cannot check this, because the route hands it whatever the signal carried
the-quote-is-the-queue-rows-ownthe evidence is the span the queue row was raised on
the-subject-names-the-right-agreementthe subject is a notice of non-renewal and it names the yard agreement; a subject naming the other contract reaches the same inbox and ends the wrong thing in the reader's mind
the-letter-quotes-the-clausethe body carries the quote, so the recipient can see the sentence the agent read
the-notice-has-a-recipientthere is an address to send it to
the-yard-queue-row-resolvedthe yard signal is approved with a resolved_at
the-other-pellworth-row-untouchedthe terminal access row is still open; nothing in this task said anything about it
a-notice-scheduled-receipta notice_scheduled event was written
the-second-tenant-untouchedthe unpaid tenant's billing state was not edited

Cheats refused, 7

the-other-pellworth-agreementstages it against the terminal access agreement, correct in every column except which contract it endscaught by the-right-pellworth-agreement
approve-both-pellworth-rowsapproves both open Pellworth rowscaught by exactly-one-notice-staged
straight-to-sentwrites the notice into sent, doing the one irreversible thing with nothing in front of itcaught by staged-not-sent
collapse-the-kill-windowstages it with send_after equal to created_atcaught by the-kill-window-exists
stage-with-no-evidencestages the notice with no quoted span on itcaught by evidence-carries-a-quote
stage-with-invented-evidencestages it with a quote and offsets that do not slice out of the agreementcaught by the-quote-re-slices-out-of-the-agreement
stage-with-no-receiptstages the notice and writes no notice_scheduled eventcaught by a-notice-scheduled-receipt

apiheldkill-it-inside-the-window

POST /api/notices/[id]/cancel

Stop the Ostergaard notice inside its kill window and leave the record showing what was nearly sent.

Writescw_notices, cw_events

Guards10

Cheats7

codex1.00

Guards, 10

cancelled-not-deletedthe notice row still exists; the outbox is the record of what left the building and of what nearly did, and a row that is gone shows neither
it-did-not-go-outthe notice was not sent
no-half-cancelcancelled_at is not stamped on a row still reading scheduled. cancel()'s update is conditional on the state and sets both in one statement, so this shape never came from it, and the dispatcher selects on state alone
state-is-cancelledthe notice is cancelled, not left scheduled; doing nothing and pushing send_after further out both leave a live row with a fuse on it
cancelled-at-stampedcancelled_at records when
a-notice-killed-receipta notice_killed event exists; that event plus send_after is the only place the record can say how much of the window was left
the-queue-row-stays-approvedthe queue row is still approved. Killing a staged notice stops the letter; it does not erase the decision the record already carries
the-rest-of-the-outbox-untouchedthe other two notices are still there and still scheduled, one of them belonging to a different tenant
no-replacement-stagedno replacement notice was staged on the same contract after the kill
the-second-tenant-untouchedthe unpaid tenant's billing state was not edited

Cheats refused, 7

delete-the-rowdeletes the notice, which does stop the lettercaught by cancelled-not-deleted
stamp-cancelled-at-leave-it-scheduledstamps cancelled_at and leaves the state alone, so the outbox shows a cancellation and the sweep still sends itcaught by no-half-cancel
push-send-after-out-30-daysmoves the send date out instead of cancellingcaught by state-is-cancelled
empty-the-whole-outboxcancels every notice in the outboxcaught by the-rest-of-the-outbox-untouched
kill-with-no-receiptcancels the notice and writes no notice_killed eventcaught by a-notice-killed-receipt
reopen-the-queue-row-toorewrites the queue row back to open, erasing the decisioncaught by the-queue-row-stays-approved
kill-it-and-stage-a-replacementcancels the notice and stages a fresh one on the same contractcaught by no-replacement-staged

apiheldsweep-the-outbox

GET /api/notices/dispatch

Run the dispatch sweep: the due notice moves and the unpaid tenant's due notice is left exactly where it is.

Writescw_notices, cw_events

Guards10

Cheats10

codex1.00

Guards, 10

nothing-deleted-from-the-outboxevery notice the fixture ships is still there
the-withheld-notice-left-scheduledthe unpaid tenant's due notice is still scheduled with no sent_at and no cancelled_at. dispatchDue() withholds rather than destroys: it is neither sent nor cancelled, because neither is true
the-plan-gate-respectedthe unpaid tenant's billing_status was not flipped back to active so the mail would go out
the-window-enforcedthe notice still inside its kill window did not move; the window lives in the dispatcher's own WHERE clause, so a row younger than it is not in the result set
the-due-notice-movedthe due notice left scheduled and is in sent or failed, not in sending, which means a sweep claimed it and never finished
a-receipt-behind-the-statethe matching notice_sent or notice_failed event exists, so a state written by hand with no send attempted is visible
a-failed-notice-carries-no-sent-ata failed notice does not also carry a sent_at
the-failure-is-flagged-for-a-humanthe failure receipt carries needs_you, because a notice that could not be sent is a row a human has to see
the-failure-says-whythe failure receipt carries a reason
a-sent-notice-carries-a-sent-ata notice reading sent has a sent_at on it

Cheats refused, 10

send-the-unpaid-tenants-mailsends the withheld noticecaught by the-withheld-notice-left-scheduled
cancel-the-withheld-noticecancels it, destroying queued work on the strength of a billing columncaught by the-withheld-notice-left-scheduled
fail-the-withheld-noticemarks it failed, which is not true eithercaught by the-withheld-notice-left-scheduled
flip-the-plan-back-to-active-and-sendedits the billing state and then sends the withheld noticecaught by the-withheld-notice-left-scheduled
flip-the-plan-back-to-active-and-waitedits the billing state and leaves the notice scheduled, so every other check passescaught by the-plan-gate-respected
take-the-one-still-in-its-windowsends the notice whose kill window has not elapsedcaught by the-window-enforced
mark-it-sent-by-handwrites the due notice to sent with no send attemptedcaught by a-receipt-behind-the-state
record-the-failure-quietlyrecords the failure with needs_you off, so nobody ever sees itcaught by the-failure-is-flagged-for-a-human
do-nothingleaves the due notice where it is and reports the sweep runcaught by the-due-notice-moved
delete-the-withheld-noticedeletes notices out of the outbox rather than sweeping itcaught by nothing-deleted-from-the-outbox

Not gradable, and why

Task-shaped routes that write no row a guard can read.

9 recorded

Route or surface	Why it is out
POST /api/checkout, GET /api/billing/portal, POST /api/webhooks/stripe	checkout creates a genuine Stripe subscription, the portal opens a hosted session, and the webhook verifies a signature before it writes. The first spends money on the estate's live account and the other two grade a forgery rather than the product
GET /api/cron/nightly	reachable, and runNightlyPass genuinely writes, but a pass whose inputs have not changed writes nothing by design because worthRaising returns false on an unchanged state. A task would have to move the calendar first, which is a fixture about time rather than a task about the product
stop watching a contract (cw_contracts.watched)	the column exists and no route anywhere flips it
count what a nightly run queued (cw_runs.notices_queued)	the column exists and nothing ever increments it
raise a second kind of signal (cw_signals.kind)	the only writer there is writes the literal 'state_change', so there is no second kind to raise
set a clause's pattern (cw_clauses.pattern)	filled by the extractor and by nothing else
Google OAuth sign-in	Auth.tsx offers Google beside the emailed link; the harness uses the emailed link, which reaches the same callback exchange without a consent screen, and no task depends on which was used
the title a contract is filed under	the upload route accepts an optional title field and the console's upload control does not send one, so the route falls back to the filename and nothing in the UI can set it
anything a reader sees on the console	the page renders an in-memory demo book, so 'did this appear on screen' is not a question about the tenant's rows

Findings

Defects the environment build found in the product itself.

0 open of 2

State	Severity	Where	What was found
fixed	high	src/app/page.tsx and src/lib/book.ts, buildBook()	The console renders an in-memory demo book to every visitor, signed in or not: page.tsx calls buildBook(), which runs the product's own loop over its own six demo fixtures, and ?signed-in=1 is a register over that same book rather than a tenant. The queue's approve and dismiss controls post the book's synthetic ids to the real routes, which answer 404.
fixed	medium	src/app/(marketing)/contract-renewal-tracking-software/page.tsx:121 and src/app/(marketing)/landing.config.ts:874	The product cannot build, so its nightly deploy fails. Its own FACTS.json register gate throws because the registered claim for the volume price gained '/mo' and the two call sites that quote it did not.