Skip to content

TownReporter 0.6.34 beta

Pre-release
Pre-release

Choose a tag to compare

@scottconverse scottconverse released this 08 Sep 17:31
b097d28

TownReporter 0.6.34 beta

This beta corrects how Dark Desk chooses and presents captured evidence to its two bounded model stages. The Nelson Road investigation had already captured the applicable official City PDF, but newest-12 and first-1,600-character truncation omitted the record from the evidence pack. A first correction surfaced the right record but reordered fragments, allowing an applicant from the preceding table row to be misattributed.

Dark Desk now ranks eligible records across the full captured inventory, deduplicates captured versions before its eight-record limit, and transfers bodies only for the selected records. Empty-focus requests retain readable evidence. When a relevant complete stored PDF page fits, the prompts preserve it once in source order; otherwise they retain a contiguous matching neighborhood. The same selection feeds separately bounded inputs: 28,000 characters for stage-one synthesis and 22,000 for the final brief. The four separate verification-gate calls are unchanged.

Focused preserved-data proof produced a 23,077/28,000-character synthesis input and 20,303/22,000-character brief input with the complete relevant page once and the correct applicant tail. Fourteen focused tests passed. In the final accepted application check, one Codex Terra rewrite completed in 19.157 seconds with no fallback. The brief correctly named Doug / Connection Church Longmont, retained the project ID, 8979 Nelson Road, approximate acreage, proposed MU-E zoning, and unchanged religious use, and left project-specific service and tax effects unestablished. It contained neither Founders Block nor Mark Sullivan. No dig, search, handoff, publication, or second model call occurred.

This bounded result does not establish exhaustive retrieval, universal model accuracy, independent corroboration, completed historical replay, Halo deployment, or measured normal-day performance. See the attached municipal-record and Audit Lite receipts for the failed candidates, exact evidence, final acceptance, and proof limits.

Exact source: b097d288d9237be225bc6ed56851078ad064d208. Application CI passed 14/14; npm test: script runner 342/333/0/9; TypeScript runner 1,615/1,570/0/45; combined 1,957 total, 1,903 passed, zero failed, 54 skipped. Windows package/install passed 2/2, and Pages succeeded. ZIP: 7,128,234 bytes; SHA-256 db56f79f788a0770b80170dc2aa05832922bd0b2ebd6512d832fec473f801c03.