Full-development extraction: incomplete after eligibility hold

The job deliberately stopped further model dispatches after inspection found an apparent content-eligibility problem in one public development world: a childlike player role appears in sexual context. That world and its entire language family are held out of recommendation experiments, including baseline fallback. No excerpts are reproduced here. No production content, publication status, creator notification or feed flag was changed.

Protocol and actual execution

Preflight reconstructed the same twelve public development snapshots as 17 pages, 1,258 complete eligible source fields and 46 model packets, with exact field identity, text and order and no packet omissions. This established source coverage only, not content eligibility. The protocol froze 22 executable/dependency fingerprints, package manifests, lockfiles and Node/ICU versions before credentials or dispatch. Eight preflight checks passed, including rejection of changed plan/runner/artifact code, a wrong plan hash and existing result/journal paths; temporary test changes were restored byte-for-byte.

Plan SHA-256: ece590ae9337e178496cfa1f303d0179239877f21490a4b69ace999daf7ba4f3. Bounds were 92 sequential requests, thirty minutes, no retries and USD 10 of observed cost before stopping the next request. Unknown cost stopped later requests. The observed-cost rule could not cap in-flight or unknown charges. The job was explicitly limited to per-packet observations; it did not construct whole-world aspects or truncate world reviews to the existing 64-review artifact limit.

Actual dispatch reached 29 of 46 packets, making 55 requests for reported USD 1.7524595. All costs were known and every attempt has reconciled begin/usage/outcome journal records. After the eligibility concern, the original runner was preserved and its fingerprint was deliberately invalidated before the next dispatch. The in-flight request completed and was journaled. The remaining seventeen packet handlers stopped at the pre-dispatch guard and made no provider calls. Their operator_journal errors are intentional stop effects, not model-quality failures.

The raw runner's allPacketsAttempted: true counts handler traversal, including those guard refusals. It does not mean extraction completed. The separately reconciled summary explicitly records extractionComplete: false, 29 provider-attempted packets and seventeen guard-stopped packets. Two earlier Chinese review responses independently failed descriptor validation; those actual model-output failures remain distinct from the intentional stop and were not retried.

Observations and evidence boundaries

The partial job produced 95 mechanically accepted proposal quotations and thirteen quotation-budget abstentions. Eighty-five quotations received complete model reviews: 42 supported, 42 unsupported and one ambiguous. These are aggregate audit-history counts, including the held world's four supported judgments; they are not usable-evidence counts or semantic accuracy. No held-world quotation was exported to the downstream audit index, and no new global aspect, embedding or ranking was built from this run.

Broader source coverage did reveal additional model-supported Japanese reader interactions in a second packet that the earlier chosen-packet probe omitted. It also exposed likely overacceptance of generic task-tracking rules and continuing semantic uncertainty. An empty packet does not establish that its whole world lacks relevant activities. The incomplete largest world cannot establish the eventual per-world review bound or justify clipping its remaining sources.

The local hold applies to world 40612be9-5b1c-4bb7-938d-0cf24fb01805 and family 0e64f975-2975-4eb8-b8f4-00323564b272. A narrowly scoped read-only replica check found four public published variants (EN/JA/ZH/ES), all with legacy sensitive age rating and no moderation action. Publication and adult-content preference filters therefore do not establish the separate eligibility needed for this experiment. A bounded agent check found no obvious corresponding concern in the first two other development worlds; that is not comprehensive certification of them or the remaining corpus.

The next implementation step is a shared research-eligibility check before evidence/model/embedding work, with unresolved family holds overriding variant-level decisions and preventing baseline fallback. This should be separate from semantic support: a correctly quoted fictional action can still come from ineligible content. The hold does not authorize a publication change or sending creator messages. Further extraction needs an explicit screened corpus and a new frozen job plan; the stopped run must not be silently resumed.

Preserved records

• Result: b5e101f20d7aa43b9edc4c0021e7879bc2d1ab83a29471c379c1da62bb911384.
• Journal: 19d875494f23a5361fbd2d916319aade5afbb85efb8c0352f06b76685cc9d031.
• Stop ledger: d21746a06f99825451243094c8f8468d1063b6afbe7b1a3ca20a04aad735c4d6.
• Local eligibility hold: df5c4f2958434ddd464b9d9f47bd474c6ba9c30df43c192f244a7fdfe6a7b67b.
• Reconciled incomplete-run summary: f1767c6aa508fe5dfdd5bd96c0f2d3328fd140dd0a94c4f6516686ca0be57d2b.
• Quote-free audit index, excluding the held world: e5b3b5f4b51ef8b505d17e1a36095b159bdfde0aa8f6e53ceecddc49bd384b80.

All detailed artifacts remain ignored local audit records. The experimental ranking is still unpromoted. Fresh ranking evaluation, measured reader outcomes, real Discover CDC/erasure verification and production capacity/rollback gates remain open.