12. One way a goto ends#

Date: 2026-08-10

Status#

Accepted. Amends ADR 4 in one particular — the model no longer decides between moving and offering — and amends ADR 6 by collapsing the two refusals it describes into one. Both conclusions stand.

Context#

ADR 4 gave the model a choice: one plausible passage, move the book; two or more, call offer_positions and let a thumb answer. That is a judgement, made at 2am, on a mumbled sentence, by whichever model is configured — in practice Haiku, chosen for speed.

It goes wrong in a way that is invisible from a desk and obvious in a bed. The model would decline the judgement and hand it back as prose: that’s further on than you’ve got — do you want me to take you there anyway? Which is the question ADR 4 exists to have deleted, arriving through the model’s mouth instead of through the tool it removed. Or it would offer a list of one place they had already heard, get the tool’s refusal — “That is one place they have already heard. Move them there instead.” — and argue with it, or re-offer, or stop.

Both are the same failure. The decision is arithmetic on two facts the server has and the model does not reliably use: how many places are plausible, and whether the single one is behind them. The model is being asked for a calculation, and what it gives back is a conversation.

There is a second cost. move_to and offer_positions are two tools with a mutual exclusion between them, so the exclusion has to be written twice — once in each — and the copy in move_to has to fire before library.move_to writes, because the row reaches the page fifteen seconds later as the refusal of its next report. A rule stated in two places is a rule with a seam.

The alternatives we rejected#

Prompt wording: tell it harder not to ask. The prompt already said, in several places, not to ask which place and not to ask about spoilers. This codebase has made this call before — acted exists because “never offer and move in the same turn” was a prompt rule the tools now enforce. The wording was not the problem; being asked for the judgement at all was.

Keep both tools and make move_to refuse a place they have not heard. Fixes the spoiler half and leaves the move-or-offer judgement, the two-place exclusion, and the bounce on a single heard place all where they were.

Have find_passage draw the list itself, with no second call. The furthest version of this idea, and it takes away the one judgement the model is good at and should keep: reading the passages it was given and saying “none of these is really it”. A screen of four wrong places is worse than being told it could not find the moment.

Decision#

Every goto ends in offer_positions, whatever becomes of it. The model names the passages that could really be the moment, best first. The tool decides:

  • more than one place → the list goes on the screen;

  • one place at or past the position → the list goes on the screen, one row, covered — the press is the consent, which is what ADR 4 built it for;

  • one place behind the position → Library.move_to, there and then, returning a Moved — unless a list is already on the screen, and then it is the second case instead: one row, and a press.

move_to comes off the model’s tool list. Library.move_to stays and is what the third case calls.

The single-heard-place Refused becomes that move. It was always the same rule; it was just expressed as a bounce that cost a round trip and could be argued with.

The exclusions collapse into one place. A turn that called recall refuses the whole tool, so a question can no longer end in either a list or a jump, and that is now one branch rather than two in two files. A turn that has already taken somebody somewhere refuses a second go.

A turn that has already drawn a list is the one exclusion that is not a refusal. It may draw another — that is a change of mind, the last offer wins, and nothing has left the process yet — but it may not move, and the move cannot be caught after the fact, because Library.offer_positions makes it before it returns and a Turn is assembled long after that. So the fact that a list is up travels down with the call, as may_move, and the place behind them comes back as a row to press. Only the caller knows the screen is busy and only the library knows what the arithmetic came to; the argument is where those two meet. It is deliberately not a count of ids at the caller: the ids are deduplicated and the ones resolving to nothing are dropped before any of this is decided, so two ids are routinely one place.

Consequences#

The model stops deciding whether to move. It judges plausibility — which passages could really be the moment — and nothing else. That is the judgement it is good at, and the one that has a cheap failure: name too many and the cost is an extra press, where deciding to move wrongly costs the hour they were in.

“Are you sure?” has nowhere left to come from on a goto. There is no consent argument to solicit and no move-or-offer fork to hesitate at. The only pushback left in the system is recall saying something has not come up yet, which is an answer rather than a question.

An explicit “take me to the end” costs one press. The place is ahead, so it goes up as a covered row. Given the alternative was a sentence asking for consent and a second turn to give it, a thumb on a row is both faster and more deliberate.

A turn can no longer move the book twice. It used to be able to move, think better of it, and move again, with the last winning. Now the second call is refused. A page dragged twice inside one answer arrives somewhere the sentence does not describe.

The prompt loses about a third of itself, including all of the spoiler apparatus that ADR 11 made unnecessary and the paragraphs that explained when to move and when to offer. What is left is routing, plausibility and how to speak.

It cost by-position navigation, and this section did not say so — amended after the fact, which is the honest place to record it. A chunk id can only mean the passage that matched, so with ids as the only currency the book could be sent to places that can be described and to no others. “Chapter 8” is not in the book’s words and no search will ever find it; nor is “the end”, nor “back an hour”. The sentence above about “take me to the end” costing one press was written as though that still worked, and it did not: there was no id for it.

The gap survived review because offer_positions contains a move, so it reads as a replacement for move_to until you ask what is still addressable. The desk suite could not catch it either — every test names a passage a search returned, because that was the only thing the tools allowed.

offer_positions takes a chapter or a position_ms as well now (issue #115). The decision above is unchanged and this is not a second way for a goto to end: a named place becomes one more candidate and meets exactly the same arithmetic, so a chapter they have heard is a move made there and then, and one they have not is a covered row and a press. The recall and offer-then-move exclusions hold on it, which matters more here than for ids — a route reachable with a bare number would otherwise be the way around those rules rather than an exception to them.

The model can still name four bad places. The tool cannot tell a plausible passage from an implausible one — that is the judgement it deliberately keeps — so a poor search still produces a poor list. The listener cancels it, which is inert, and asks again.