Explicit quotation-budget abstention

The numbered-unit development trial at d548c6b70 discarded an entire model response when one selected quotation exceeded 256 embedding tokens. Preserved diagnostics showed four valid Chinese selections alongside two oversized ones, and five valid Japanese selections alongside one oversized one. This change tests a specific reliability correction; it does not establish a better ranking algorithm.

Local contract

curateActivityUnitSelection and the provider adapter now share the same strict curation core. A selected quotation exceeding 8,192 UTF-8 bytes or 256 actual cl100k tokens abstains whole. Its original selection index, byte count, token count (unknown for byte overflow) and quote_budget reason remain explicit in a versioned curation record. Accepted indices map the compact proposal array back to the original model selection. The original complete selection is also retained.

Wrong hashes, unknown sources, invalid ranges, duplicate selections, non-unique occurrences, blank quotations and invalid structural boundaries still reject the whole response. These checks also run for oversized quotes, and processing continues through later references so an early budget abstention cannot hide another malformed selection. Quotations are never rewritten, clipped or shortened. The existing canonical locator, independent reviewer and artifact contracts remain unchanged.

The proposer prompt and wire format are unchanged in this experiment. Independent review receives the original complete packet and only the strictly validated quotations, without the proposer selection or curation rationale. The offline pure function makes the same behavior reproducible without another proposer request. Provider usage remains journaled even when the resulting proposal is empty. No serving caller, provider, subscription, database write or production flag is added.

Verification and development replay

The mixed-selection regression first failed with descriptor_validation under the prior behavior (13 tests passed, one failed), then passed after the change. Twelve additional pure-curator tests cover exact byte/token limits, compact-to-original index mapping, all-abstained results, defensive snapshots, and malformed/ambiguous/duplicate evidence before and after an oversized range. The integrated personalization, baseline and billing suite passes 522 tests with no failures or skips. Full workspace typecheck passes eight tasks and build five. Independent code review found no actionable issue and passed 34 focused tests plus 15 synthetic adversarial checks. Separate helper review found no blocker before the paid replay. Final-head CI is checked separately.

A frozen replay protocol reconstructs both diagnostic selections from only the saved bounded references into known public sources. It verifies the immutable twelve-world development snapshot, previous diagnostic inputs/results, original packet and presentation hashes, and current implementation/helper hashes. This reconstructs known selections; it does not recover or reinterpret raw provider responses. No new proposer call, private reader data or held-out cohort is used.

The local replay retains four Chinese and five Japanese quotations, with three explicit budget rejections. A separate protocol permits at most two independent-review requests, no retries, five minutes, and USD 1 of observed cost before stopping subsequent requests. Unknown cost stops later requests; these rules cannot cap in-flight or unknown charges. Every outcome is retained with a fresh private durable journal. Protocol SHA-256: 9a4d1bf59aeeb0020d1604eceded31b83809423eefbe6b3f4292289598a71605.

Real-model outcome and semantic limitation

Both independent reviews completed. Chinese received one supported combat/creative-attack/reward passage and three unsupported NPC-background or NPC-only interaction passages. Japanese received one supported group-chat-format passage and four unsupported narrator/style directives. Each resulting artifact contains one bundle. Two review-only requests cost USD 0.217671, with no unknown costs and complete journals. Result SHA-256: 06c6ca2570d71f4411aab52f509bcf4d02d0dc0b1edee169c1ced9fd47c68d6d. These are reconstructed diagnostic selections, not a successful rerun of the original failed smoke.

Inspection raises a specific concern about the Japanese supported judgment: sharing a scene with a cast and specifying distinct voices may establish a group-chat format without establishing a concrete reader activity or enacted relationship. The model called it supported; the development assessment treats that conclusion as disputed. It is not human-verified semantic truth and must not be silently promoted into a positive training example. The Chinese supported quote explicitly describes player attacks, doubled damage for creative attacks, and a reward after defeating the boss.

The earlier Spanish smoke returned six valid quotations that were all independently rejected as NPC/faction descriptions or configuration. This budget change addresses measured mechanical loss but does not resolve semantic selection. The next development experiment must distinguish reader-involving concrete interaction from NPC-only activity and generic format/narration promises, while retaining valid fictional mechanics written as instructions. Freeze contrast cases and their intended decisions before model calls; preserve all disagreements. A fresh preregistered ranking comparison and measured reader outcomes remain necessary before experimental promotion.

Both Discovery CI and OSS export CI subsequently completed successfully on exact curation commit 360e32239b8816a74458ea8b181ea07da4479e06. These checks do not establish semantic or reader uplift.