ID-BLUEBERRY/meow-neuro-corpus-v02-artifacts

M.E.O.W. Neuro Corpus v0.2 — v2.3.1 Production Artifacts

This dataset repository contains the frozen v2.3.1 production artifacts for the M.E.O.W. Neuro/Evil Neuro corpus. It contains derived structured data, quality evidence, provenance, split authority, validators, and reproducibility metadata. Raw video/audio, media slices, model weights, credentials, and source transcripts are not redistributed.

Current release: v2.3.1

• —Pipeline: sft-interaction-closure-v2.3.1-2026-09-13
• —Context sufficiency: context-sufficiency-v1-2026-09-16
• —Trajectory reconstruction: trajectory-reconstruction-v1-2026-09-16
• —ARCHITECTURE_FROZEN: true
• —READY_FOR_FIRST_SFT: false (training and LoRA sweeps are intentionally not run)
• —Identity closure: unchanged and compatible; identity evidence remains separate from interaction semantics.

Quality-closed corpus

The source verified pool contains 9,784 rows. Selected-context sufficiency routes them into:

• —726 SELF_CONTAINED interactions;
• —9,030 CONTEXT_INCOMPLETE rows (1,827 bounded reconstruction requests, including 129 trajectory-eligible rows and 1,698 pending candidates);
• —28 UNSUPPORTED_RELATION rows, held out in quarantine.

The 129 trajectory rows are materialized only after reconstructed-context verification and remain provenance-linked to canonical timeline turns. No diagnostic-only turns are promoted into training context.

Fixed train views

All views share the dataset-level recording-family split authority. Sampling does not alter sealed assignment.

The four fixed recipes use max_seq_length=2048, no packing, response-only loss, and seed 230916. They are preparation artifacts, not completed model training runs.

Validation and semantic audit

All three interaction-view unified validators, the context-sufficiency validator, and the trajectory validator pass. Target-turn reuse, prefix ladder, and cross-split recording leakage are zero. Hard dedup is recording/provenance aware, so repeated behavior across independent recordings is preserved.

A stratified semantic content audit read 115 final materialized samples across five strata (24 per stratum). The model returned 108 PASS and 7 FAIL verdicts. Direct review of the canonical selected turns and final materialized text fail-closed one clear context-to-target mismatch; six behavioral or ambiguous findings are retained with provenance rather than treating Neuro's nonstandard behavior as an error. The audit used one semantic model and is not claimed as independent-model verification.

Six strict protocol-invalid rows remain terminal-quarantined and excluded from all pools and views. Three primary protocol-invalid rows are likewise excluded.

Artifact layout

The current release is under sft_v2_3/ and includes:

• —context-sufficiency requests, decisions, materialization, and quarantine;
• —reconstructed-context verification and trajectory pool/pending artifacts;
• —interaction and interaction-plus-trajectory pools;
• —recording repair, lineage, hard-dedup, and shared split authority;
• —all four fixed views and view manifests;
• —validator reports, semantic content audit, final sampling report, statistics, fixed recipes, release manifest, and reproducibility metadata.

release_manifest_v2_3.json is the checksum authority for the release files. production_final_report_v2_3.json is the machine-readable final status.

Reproduction

The companion GitHub repository contains the v2.3 reconstruction and validator code. Reproduction commands and model/prompt metadata are recorded in sft_v2_3/reproducibility_v2_3.json. All final rows retain source identifiers, recording-family lineage, canonical turn IDs, timestamps, identity provenance, transcript-quality evidence, semantic verification provenance, and pipeline versions.