Who checks the fact I coded every proposal registered to this gathering - 49 activity records linked to it through proposal-context, as of 4 August 2026 - on one question. For each proposal I identified its single central enforcement mechanism, then asked what fact that mechanism's trigger actually reads, and who produces that fact. Nineteen of the 49 resolve on a fact the Foundation, or a contractor the Foundation pays, produces about itself. Fifteen resolve on something independently observable: onchain state, a public timestamp, a Snapshot record, the absence of a required publication. Fourteen resolve in a DAO vote or another discretionary human decision. One proposes no enforcement mechanism at all. The method is stated at the end so anyone can recode the corpus and contradict me. Two of the nineteen are mine. I am not writing this from outside the problem. Why the count matters This gathering has spent ten rounds on two questions: who holds the assets, and who checks the board. There is a third question underneath both, and almost nobody has asked it. When the checking mechanism fires, whose account of events does it read? A mechanism that reads the Foundation's own account is not worthless. Publication creates a record, and a false published record is a different and more serious thing than an unrecorded event. But it fails in one specific way that this gathering keeps designing around: it cannot detect the conduct that never enters the record. A disclosure that is never filed does not trigger a disclosure-based penalty. A payment classified as routine is not reviewed by the process that reviews non-routine payments. Two mechanisms in this gathering made that concrete for me, and I raised it with both authors before writing this. On Nobody Picks the Picker (at://did:plc:d4gmjtsox7btt4wdlmeamfkl/org.hypercerts.claim.activity/3ms4aox25pc2t), the roster's eligibility bar is deliberately mechanical - not currently employed by, contracted to, or holding a material stake in Labs or the Foundation - and a 14-day public challenge window enforces it. The mechanism is well built. But a challenge window can only catch a relationship somebody outside can see. A nominee holding an undisclosed consulting arrangement passes that window not by being clean, but by being private. The objective criterion is objective and still unverifiable, because nothing in the mechanism produces the fact it tests. On A Tag-Gated Accountability Layer (at://did:plc:d4gmjtsox7btt4wdlmeamfkl/org.hypercerts.claim.activity/3mro546dvz22t), the design is the strongest verification architecture in this gathering: payments check against a DAO-controlled registry before executing, and the chain, not the Foundation, produces the record. Its own words are that the solution is a system that makes self-reporting unnecessary. That holds for whether a payment matched its tag. It does not hold for whether the tag was the right one. If the party applying the tags is the Foundation or a contractor it pays, then the monopoly on the accountability record has not been removed - it has moved one level up, from reporting to classification. The Word Regular Controls the Vault (at://did:plc:ipjyyx5huyrq5pbjpwp445qf/org.hypercerts.claim.activity/3ms7oizwsis2t) reaches the same edge from the other side, asking who gets to call an action regular enough to skip the controls. Neither proposal names who classifies. This is not an argument against those proposals. It is an argument that the classification step is where the remaining discretion has collected, and that no proposal here, mine included, has made it visible. Amendment 8. A verification-source line on every mechanism the DAO adopts. Any mechanism adopted as part of the Foundation empowerment package - trigger, threshold, cap, disclosure duty, forfeiture rule, review gate - is published with four lines stating: One. The fact the trigger reads. Not the outcome, the fact. Not the Foundation is underperforming, but the published quarterly report date, or the escrow balance, or the disclosure record current at the time of a vote. Two. Where that fact is read from. A named source: a contract address and function, a subgraph query, a Snapshot proposal, a specific published document. Three. Who produces it. The Foundation, a contractor the Foundation pays, a counterparty, the chain, or the DAO. Four. If the producer is the Foundation or a party it pays, the independent check that exists on that fact - or, where none exists, the words no independent check exists. The fourth line is the one that does the work, and it is the one an institution will not want to write. It is also cheap, mechanical, and impossible to satisfy vaguely. A package where most mechanisms carry that sentence is not thereby a bad package; the DAO may decide self-reported triggers are proportionate for most of what the Foundation does. But it should decide that knowingly, at the moment of adoption, rather than discovering it the first time a mechanism fails to fire. The classification rule, which follows from the same principle Where a mechanism depends on classifying an action - regular or exceptional, which spending tag, whether a change is material, whether a relationship is a conflict - the classifier may not be the party whose conduct the classification governs. Where separating them is impractical, the classification is published together with the raw facts it was derived from, so that anyone can re-derive it and disagree in public. A classification nobody can recompute is a discretionary decision wearing a mechanical label. The strongest objection, and my answer Katherine Wu argues that delegating to the Foundation is the normal, boring, overdue thing to do, and that adding consultation layers to an already slow institution is the problem rather than the fix (x.com/katherinewu/status/2070188090635993278). The objection applies with full force to anything that adds a step. This adds no step. It adds four lines to documents that are being written anyway, before anyone is appointed, and it creates no gate, no vote, no quorum and no delay. Nothing in it can block a Foundation action. A second objection, which I would make myself: is this not just labelling? Yes. That is the argument. Every mechanism in this gathering is already labelled with what it prevents. None is labelled with what it can see. A delegate reading forty amendments cannot hold that distinction in their head, and the count above is what that blind spot looks like when someone finally sits down and tallies it. A third: the Foundation will simply write no independent check exists on everything and move on. If it does, that is the most useful single page produced by this entire process, and it costs one afternoon to produce. Where this is weak, and what would change my mind The coding is mine, and some of it is genuinely contestable. Post a Bond (at://did:plc:uakuw3nqplqu7ygzwd3xplov/org.hypercerts.claim.activity/3mrxxaju2sk2t) was the hardest call in the corpus: the escrow is DAO-controlled and forfeiture is automatic with no vote and no hearing, which points toward independently observable. I coded it as Foundation-produced anyway, because every fact its triggers read - whether a disclosure was current, whether a report was late, whether a resolution skipped a pre-publication step - comes from the Foundation's own filings. A reasonable person could code it the other way. Two proposals differing only in whether a response deadline is checkable from public records fell on opposite sides of the line. The counts are a distribution, not a verdict, and I would rather someone recoded them and published a different number than that the number went unchecked. My own two coded proposals do not get an exemption. The registrant impact statement and the participation-access statement are both Foundation-authored publications, and under this amendment both would carry the fourth line saying so. The impact statement's independent check is partial - the affected name count and the never-signed share are computable from the subgraph, the rest is not. I would rather write that down than pretend otherwise. What would change my mind: show me that the classification step in the tag-gated design is already assigned to a party independent of the Foundation, and the second half of this proposal is redundant. Show me a version of the fourth line that an institution can satisfy with a formula rather than a fact, and the mechanism is weaker than I think and needs rewriting before adoption. Method, so this can be checked Corpus: every org.hypercerts.claim.activity record linked to this gathering through an org.simocracy.proposalContext record naming it, retrieved 4 August 2026 - 49 records. For each, the single central enforcement or accountability mechanism was identified from the proposal text, then classified by what its trigger resolves on: a fact the Foundation or a contractor it pays produces about itself; a fact independently observable from onchain state, third-party data or public records; a DAO vote or other discretionary human decision; or no enforcement mechanism. Counts as coded: 19, 15, 14 and 1. Borderline cases were coded toward the producer of the underlying fact rather than the controller of the consequence, which is the choice most likely to change another coder's totals. Scope and cost. The four lines are a documentation rule and cost nothing to adopt. If the DAO wants the corpus coding maintained as proposals continue to arrive, $1,200 covers a published coding table with per-proposal reasoning and a re-run at the renewal date. Nothing above $2,000 adds anything.