Can a rule stop a bad reply before it’s sent?
The reply gate checks an AI's replies in the Authority System Suite after they are written, so it can force a correction but can't stop a sentence being sent. Its rules were measured against real replies before adoption, and the same limit led to checkpoints.
The reply gate is a check run on the AI's replies to me in the Authority System Suite, a set of rulebooks that make an AI show its work. It can't stop a reply, only force it to be corrected, and that limit pushed the work toward checkpoints.
I wanted controls that prevent bad behaviour, with many of them stopping problems before they reach me, instead of controls that only tell me when it has happened. The gate's file records that requirement.
The file then separates two kinds of control. One prevents, the other detects, and the difference is whether something sits in the path of the action. A write script sits in the path. A sentence in a reply does not.
The gate detects and forces a correction
The write script for the Paper Trail, the Suite's records, has 91 refusal points: a malformed record is not written. The audit's checks (64 of them on 5 October) each only print. The reply gate is a hook, a script run after the reply is composed, and its own block message says the maintainer has already received it.
So it is detection with a forced correction. It cannot stop a sentence reaching me; it stops the turn ending until that sentence is fixed.
Cheap word-list gates would have blocked almost nothing
Before designing the gate, I ran candidate rules against 15 real replies. The cheap one, word lists for the known failure modes, would have blocked almost nothing: 0 of 15 for self-assessed significance, announced restraint, overconfidence and promises without execution, and 1 of 15 for two more. The file concludes that the cheap gates are worthless, because they catch a register the model already doesn't write in.
The tiers that were adopted were measured first
A tier, one layer of the gate, that checks causal and capability claims, narrowed to claims about system things, came to 15 sentences across 15 replies, 13 unmarked. That was adopted. Its live log, snapshotted on 24 September, shows 15 blocks over 61 replies, one in four. The file also corrects an adopted decision that had asserted “four replies” from memory without reading the log.
Another tier deals with a problem I kept having in another conversation. Given several tasks, the AI would defer some of them again and again, never attempting them, do other things instead, and never tell me it had deferred them.
One design for that tier was discarded before it shipped. It was a structural check that counted enumerated items in my messages. Of 81 of my messages quoted in the trace, zero carried an enumeration cue: my instructions are short prose, so the check could not have fired once. It was rebuilt around the side that can be parsed, the reply's own promises. That tier still only observes.
On 2 October I adopted a fourth tier, which refuses lists whose lengths look shaped by habit and carry no count. It was measured first, on 300 retained subagent replies: as shipped it fired on 1.
The shape that works is a checkpoint
The same reasoning led, on 5 October, to the checkpoint unit, which I chose to build together with a right-of-reply checkpoint. A checkpoint is the smallest unit of doctrine, the Suite's written rules: a trigger, a step, the exceptions and who decides them, the record the step leaves, the rule that requires it, and the test.
The first checkpoint requires that anyone accused in a published finding is offered a chance to reply and that the answer is reported. Checkpoints for source protection and the publication ethics record followed on 10 October. An audit check reads that every checkpoint has all seven parts.
The limits it states
The gate fails open on any internal error. Claude Code is documented to override a stop hook after eight consecutive blocks; the file marks that as unverified. A green result means only that no condition matched, not that the reply is true. And one note records that neither hook was registered in a Cowork cloud session.
The checkpoint check cannot see whether a step was taken in any piece of work. That is the record's job and the final audit's. Prevention exists only where a tool sits in the path of the action, and everywhere else the Suite relies on detection.
Sources and notes (8)
- Claude Code docs: how Claude remembers your project — first-party statement that instruction files are context, not enforced configuration
- The Sequential Enumeration paper (Hou, Zorzi, Testolin) — models do not count spontaneously
- Fairness in journalism, Media Helping Media — origin of the right-of-reply checkpoint CP-01
- The Perugia Principles (Posetti et al.) — source protection steps behind CP-02
- IPSO Editors' Code of Practice — public-interest exceptions and recording the reasoning at the time
- Reply gate — the notes this was written from
- Checkpoints — the notes this was written from
- Directives — the notes this was written from
Authority System SuiteAll 17 parts →
