The arguments that decide an agent ledger's fields, and why copying the fields skips them

The reason a copied schema rots is visible in what the tensions actually ask for. None of the three is a field rename. The first forbids a capability: the ledger may not block, veto or gate the action it records, and the post tags this explicitly as a constraint on the whole design rather than a key in the record ^[6][15]. The second forbids an operation: no overwrite, so "we decided A, then B instead" becomes two records with a relationship between them rather than one field whose value changed ^[9]. Only the third asks for something you could add to a YAML block, and even there the ask is metadata about where an existing value came from ^[12][2]. A team that lifts the field list inherits none of the three, because none of the three lives in the field list ^[3].

The provenance argument is the sharpest thing in the thread. `version: 7` tells a future reader what supposedly governed the decision; it says nothing about how the system established that version 7 was authoritative at the moment of the decision, and the post treats those as different trust claims ^[12]. Per the post, two commenters, Self-Correcting Systems and pm25coder, reached the same place from opposite directions: a version fetched fresh from its authority at 09:22, a version read from a five-minute cache, and a version inherited from session state all produce an identical field while supporting completely different claims about what the system could reasonably have known ^[13]. The supplied text breaks off at "The fix is to", so the remedy is not something I can report ^[14].

Count the baseline record's evidence block and the gap is concrete: two artifacts times artifact, authority and version is six leaf assertions, and not one of them records how the version was read ^[4][1].

pm25coder's objection is the one worth sitting with. A ledger that only narrates can quietly become fiction, and trust, on that reading, comes from being able to gate rather than merely describe ^[7]. The author accepts the diagnosis and draws the boundary a step earlier: enforcement lives at the policy and tool boundary, while the record carries a `policy_evaluated` result showing that a check ran and what it concluded, never the enforcement decision as its own authority ^[8]. That turns one component into two, and the seam between them is where the ledger's usefulness as evidence now sits.

The append-only rule has the same shape, a constraint sold on what it protects rather than what it stores. Rewrite the March record in August and you have destroyed the ability to answer whether the March decision was reasonable given what was known in March; the noisier history is the correct trade, and compaction can always produce a clean current-state view afterwards ^[10]. That is the same discipline the post attributes to Forensic Receipts ^[11]. For an existing decision log, the tensions convert into diagnostics that need no schema at all: whether it can refuse an action, whether a change of mind rewrites the old row, and whether a row distinguishes an authority version that was fetched from one that was remembered.