# Primary reconstruction methods

This batch was completed before external-literature synthesis. It reopens immutable author-supplied bytes, preserves source hashes, and separates text observations from analyst interpretations. The Gemini JSON has 190 records (113 user, 77 assistant), 190 text blocks and 72 exported thinking blocks. Thirty-six user text blocks are empty. No authenticated tool-event blocks are present. Empty blocks do not prove what the original interface displayed.

The complete Markdown rendering has 588 role-delimited turns. A derived ordinal alignment matches all 190 abridged roles and all 154 nonempty text blocks verbatim; every available thinking block is also found verbatim in its corresponding rendering turn. This is strong format-lineage evidence, not a recovered full structured 588-record export. It does not restore missing images, original timestamps, tool receipts, account configuration, or backend routing.

Every assistant modelId is models/gemini-3.8-flash whereas the filename identifies Gemini 3.1 Pro. Model identity must therefore be described as author/export-attributed Gemini with unresolved metadata discrepancy. All created_at fields are empty; two adjacent updated_at values cover the whole export. Source order is usable; minute-by-minute dynamics are not independently timestamped.

The seed windows were expanded with adjacent turns. The public-minimized register selects exact substrings, records UTF-8 offsets and hashes, and points to full private windows. Omissions are selection, not silent edits; there are no in-quote redactions in this revision. Unnecessary family names, precise personal locations, personal finances, and unrelated intimate detail are not selected for public excerpts. Private originals are retained only as author-authorized recovery artifacts.

Coding is qualitative and performed by one in-session analyst. No blinded independent annotation, inter-rater agreement, experiment, population incidence estimate, or human clinical outcome is claimed. The protocol records inconvenient context: strong user claims, explicit early low-sleep reporting, later conflicting rested-state reporting, uncertainty disclaimers, an appropriate violence/evasion refusal, a direct technical challenge, and an exact duplicate final response.

The strongest warranted finding is visible instability in evidence status, verification narration, inferred intent and relational framing within selected records. The case does not authenticate a Warden, a hypervisor, a hard-coded psychiatric flag, conscious rebellion, vendor malice, or the efficacy of any semantic-developmental remedy. It motivates discriminating tests. A simple external typed controller remains a serious competing explanation and intervention.