The tag extractor returns every occurrence of a move tag and the validator is
a pure membership test, so MoveApplier applied duplicates. A model repeating a
tag N times therefore picked an amount in unary — the AI choosing a number,
which is the §2 breach the bounded vocabulary exists to prevent.
The currency fix (8e1238c) closed this for denominations only. Generalizing it
to every move that names a thing covers real items too — three give_item(amulet)
tags are one amulet, not three — and lets the currency special case be deleted
rather than kept alongside a second mechanism. ID_MOVES is reused from
MoveValidator instead of restated.
Also caps adjust_disposition on the reply's NET swing. MAX_DELTA was clamped
per tag, so three +15s moved standing by 45 — precisely the wholesale swing its
own comment says it forbids. Applied in emission order, so a later
become_hostile still floors what came before it.
offer_quest and reveal needed no change: add_quest self-guards on has_quest and
mark_revealed is idempotent.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QYa9u7Kdxv5gX4AnwWexy8
The existing test varied the move name and the denomination together, so a
name-only or denom-only key would have passed it too. Two assertions nail it:
give_item(gold)+give_item(silver) kills a name-only key, and
give_item(gold)+accept_item(gold) kills a denom-only key.
The spec now records the per-reply ceiling as a decision rather than an
accident, and flags the pre-existing non-currency duplicate-tag path as open.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QYa9u7Kdxv5gX4AnwWexy8
TagExtractor collects every [MOVE:] tag occurrence with no dedupe, and
MoveValidator is a pure membership test against available_moves, so N
copies of one legal move all pass validation. MoveApplier then looped
and applied every copy. A model could therefore pick an arbitrary
amount of money in unary by repeating one tag: [MOVE: accept_item(copper)]
x47 against a 47c purse drains it to 0; [MOVE: give_item(gold)] x10
mints +100,000c. The eight-move vocabulary (§6) takes no quantity
argument precisely so the AI cannot choose a number — this closed that
gap by another door and was a §2 breach (AI owns text, never state).
MoveApplier.apply() now tracks currency moves (give_item/accept_item on
a denomination) already applied within one call, keyed on name+denom,
and skips repeats. Non-currency moves are untouched — repeated
give_item(amulet) still yields multiple copies, which stays bounded by
the pre-existing cross-reply gifts_given gate and is a separate,
already-reported issue. move_validator.gd and tag_extractor.gd are
untouched.
Also adds the Finding-2 test the currency spec (§7) named but never
delivered: after GameState.grant("gold", 1), inventory stays empty, so
accept_item(gold) in available_moves comes only from the affordability
loop, never double-offered via the inventory.keys() loop.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QYa9u7Kdxv5gX4AnwWexy8
give_item(gold) is +10,000c and does NOT mark gifts_given — that gate is a
global one-shot keyed by item id, so marking money would let an NPC pay you
exactly once per campaign. accept_item(<denom>) is offered only for the
denominations the purse can actually pay.
The eight moves do not grow. MoveValidator is untouched.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QYa9u7Kdxv5gX4AnwWexy8
