The register, in the open

Run our numbers against us.

This page is the product's spine made public: every defect we know about, every claim we make with its verified status, every mistake we've made with its root cause — and the exact scope of what our guarantees bind. Figures frozen 2026-08-16 13:55 UTC for this launch snapshot; production will render them live from the register.

The register right now

120
findings recorded · 101 open
19
corrections — published in full below
21
marketing claims under continuous verification
45
actions in the authorisation ledger

Status across all 120: 101 OPEN · 8 CLOSED · 6 INFO (positives) · 3 WITHDRAWN · 1 PARTIAL · 1 QUALIFIED. Withdrawn findings stay visible — a register you can silently prune is a register you can't trust.

What the severity tiers mean

The register's working definitions — so "serious" can't quietly mean something mild.

TIER 0 8

Foundational / framing — findings about how the estate itself is read and governed.

TIER 1 29

Serious. Wrong money, wrong records, unearned claims, or security exposure. Fix before launch.

TIER 2 31

Material. Real defects with bounded blast radius or a known workaround.

TIER 3 33

Minor. Inconsistencies and gaps that mislead but don't yet damage.

TIER 4 / INFO

Informational, including recorded positives — things that work and must be preserved.

The claims table

Every sentence our marketing makes is a row here, checked against production. A claim that can't survive its query doesn't ship. Current: 7 FALSE · 3 PARTIAL · 2 UNDERSTATED · 1 TRUE-NARROW · 8 TRUE — launch is defined as FALSE = 0, and yes, we're publishing the sevens.

IDClaimStatus
C1Dashboard: "6 MONITORED"FALSE
C2Projects page: "0 active of 0 total"FALSE
C4Audit page: "a read-only record of what the system has actually done"FALSE
C6Sign out ends your sessionFALSE
C11Enable build alerts — real-time notificationsFALSE
C17Marketing: "Live deliberation engine — 16 open-reasoning rounds on record"FALSE
C18FAQ: "Structured deliberation is live today"FALSE
C10A non-superadmin cannot read the audit logPARTIAL
C13Serving-check detects a stale-serving domainPARTIAL — crash-test scheduled
C20Homepage: "Bring your own keys. Every call audited."PARTIAL
C15Marketing: "39/39 tables RLS-enforced"UNDERSTATED (40/40)
C16Marketing receipts: 612 findings / 124 scans / 5,878 audit entries / 16,919 pingsUNDERSTATED — all stale low
C9Merge queue enforces diff-read-requiredTRUE-NARROW
C3"RLS 100%" / "RLS Coverage 25/25"TRUE
C5"Only emails listed here can complete sign-in"TRUE
C7Coverage measures whether each active project has completed a health checkTRUE
C8Hiding routine activity never conceals warningsTRUE
C12Security Score 74/100 grade CTRUE
C14Marketing: "7 isolated projects, own Supabase each"TRUE
C19FAQ: "Deeper scenario tooling is on the roadmap, not yet shipped. We would rather say so."TRUE
C21The product can route to a model and record the resultTRUE

The corrections ledger — all 19, verbatim

Every time our own analysis was wrong, what we claimed, and why it was wrong — root cause included, never deleted. Three of these correct the same finding three times; the sequence is preserved because it teaches more than the conclusion.

C314 Aug
Called RLS coverage OVERSTATED on the strength of 20 advisories
40/40 tables are RLS-enabled; the score measures enablement and enablement is total.
C514 Aug
Called the allowlist claim CONTRADICTED
Read last_sign_in_at as admission; it records only the identity-provider exchange.
D714 Aug
Allowlist description contradicts behaviour
Falls with C5.
D2914 Aug
Claimed no alarm has ever fired
Health alarms have fired 11 times; the sharper finding is that nobody reads them.
D3614 Aug
Claimed the whole governance half has never run
agent_tasks holds 7 real tasks with real successes and two real 401 failures.
D3814 Aug
Claimed the health alarms had no delivery channel
/api/notifications/stream and /dashboard/notifications both exist.
D4514 Aug
Called a 404 on /dashboard/gh-queue a defect
The real route is /dashboard/github-queue; a mistyped path correctly refused.
D5314 Aug
Implied handle_new_user() had been cleared when only handle_new_user_role() had
Read it in full; it is a trigger function with an immutable search_path and is sound.
POSITIONING14 Aug
Proposed risk-tiered authorisation as a new idea
It already exists — issue #161, task a6086c37, 4 June.
MASTER-REF14 Aug
Treated the recovery file's VERIFIED CURRENT STATE block as current
11 of its statements were stale; an append-only log stays accurate while its summary rots.
D63-draft14 Aug
Nearly filed: no provider secret exists in the vault at all, including Anthropic
vault.secrets returns 0 rows to this connection — the join was uninformative, not negative. Anthropic renders Connected, so the app can read what I cannot. Withdrawn before it was stated.
D32/FORK-Q214 Aug
Framed five months of zero deliberation runs as revealed preference — the market telling you the feature is not wanted
Wrong inference. Multi-model deliberation is a feature for when you would otherwise ask a colleague. A solo operator has no colleague and does not deliberate with himself. Zero usage by a one-person team is evidence about team size, not about demand. The owner made this point, and it is correct.
D36/D6614 Aug
Implied the governance half was built worse than the monitoring half
The deliberation integrity model — a database CHECK enforcing that no verdict is exposed before every blind verdict is committed — is the most careful engineering in the estate. The difference is not build quality. Crons run without a human present; governance requires one.
E214 Aug
Proposed adding one team_members row as a 10-minute test of multi-user access control
Invalid as designed. projects.team_id is NULL on all six active projects, so the team branch of the policy cannot match. Caught by checking the test before proposing it.
D3414 Aug · correction 1 of 3 on one item
Claimed spend control is off and unmeasured, and that the "$0.07 of $5.00" figure had nothing behind it
Wrong on both counts. cost_limits holds SIX enabled limits across THREE scopes with spent_today_usd tracked per limit. The $0.07 figure is a real row. The design is more granular than a single cap.
D7414 Aug · correction 2 of 3
Claimed the cost meter froze on 19 June and has not run for 56 days
Wrong — the meter works. The limits were decremented within two seconds of the last real model call. Rollover is lazy: no spend means no rollover. Nothing is broken; there has simply been no activity.
D77-framing14 Aug · correction 3 of 3
Two cost figures disagree by roughly 9×
Invalid comparison — a daily bucket against an all-time sum. Same-day comparison gives 13.2×, and exposes the mechanism the 9× framing concealed: $0.0700 is exactly 7 calls × $0.0100. The original figure understated the defect and hid its cause.
D8714 Aug · withdrawn in 20 minutes
A project failed 4,083 of 4,083 health checks and monitoring then stopped silently; the operator was never told
Both halves wrong. The project had been deliberately soft-retired by owner decision, recorded in the audit log with reason, reversibility and preservation flags — surfaced correctly and in full. Root cause: concluded from a probe without running the control.
D10415 Aug
Called a passing reconciliation "vacuous — the check it monitors is broken"
Inference presented as fact, and wrong. One query dated the pattern: the failure had been fixed 26 days earlier by a token rotation; the pass was genuine. Root cause: pattern-matched without dating the pattern.

What our guarantees bind — and what they don't yet

The tamper-evidence scope, stated before you ask

What binds today: database constraints make deliberation records write-once against every application path, every API route, and every non-superuser role. No panellist can see a peer before committing. Nothing is "verified" without a named verifier and a timestamp. An application bug and a hurried admin hit the same wall.

What does not bind yet: us, as the database operator. A superuser could disable the trigger or restore an altered backup. Independent anchoring — a published hash chain and customer-side verification, so the evidence checks out without trusting ZASIS — is on the roadmap, and until it ships you will not find the words "immutable" or "not even us" anywhere on this site.

Register refs: D114 (the constraint chain), D117 (the scope), D115 (first public record — in verification).

How to read our evidence labels

VERIFIEDQueried or fetched directly, this session, from the authoritative source. The query ships with the number.
USER_PROVIDEDSupplied by the operator — a screenshot, a statement. Trusted as testimony, marked as such.
AGENT_REPORTEDAnother system or model said it. Never presented as our own finding until independently verified.
INFERREDFollows from evidence but wasn't directly observed — always named with the check that would settle it.
UNKNOWNWe can't establish it. Written as "unable to determine" — an empty finding is honest; a filled-in guess is not.