brain-navigator-r2

<!-- szl:artifact-identity-reconciled -->

Artifact identity (noted 2026-09-30). Besides the LoRA adapter, this repository's root carries a merged full-precision checkpoint (model.safetensors + config.json, loadable with transformers) produced by the receipted CPU merge of this adapter into its declared base (merge_receipt.json). The adapter remains the artifact of record for every figure on this card; the merged bytes carry no separate held-out receipt and add no claim. Metadata-only note.

SEPARATE RESEARCH ADAPTER · MEASURED NAMED-N RECORD · LATER GATE FAIL · NOT PROMOTABLE · NOT AUTONOMOUS

BrainNavigator-R2

What this is

A separate research LoRA adapter on Qwen/Qwen3.5-0.8B. The training_receipt.json records bf16 LoRA with r=16 and alpha=32, a synthetic NAVIGATE/ABSTAIN curriculum over 575 public handles, and 0 raw private-graph nodes admitted to gradients. It is not a model of the private 9,464-node graph and confers no private-graph access or content resolution in weights.

Evidence and artifact identity

Reviewed Hub revision: ad510c4b05429acb862b216718f098ace6d05af3. The eval_report.json records maturity=MEASURED_RESEARCH_ONLY and publication_eligible=false.

These values are from the generate section of eval_report.json, computed on 2026-08-29 on a local LoRA. Its separate software section concerns lexical retrieval over the public projection. Neither section is a public leaderboard or an independent third-party benchmark. The generate receipt does not identify an exact public adapter-file SHA-256 or a fully specified runtime build.

The separate post-publication gate receipt names Hub revision ad510c4b05429acb862b216718f098ace6d05af3 and reports 2/5 on the canonical held-out suite in transformers-cpu mode, with overall FAIL, heldout_passed=false, productionAuthorization=false, and publicationAuthorization=false. This is a different suite from the earlier named-N record; its results cannot be exchanged or combined with those earlier scores. It also records refusalPassed=false and a refusal regression: rate 0.2 against baseline 0.4. The linked receipt retains the execution warnings; earlier execution failures remain in the canonical gate archive.

The merge_receipt.json separately records a 2026-09-14 repair after the prior merged checkpoint was an empty base copy: 96/96 adapter modules applied, maximum weight delta 0.0025102924555540085. That merge record establishes the reported repair operation, not a new behavioral evaluation. The earlier named-N result does not automatically transfer to the repaired merged checkpoint.

What the evidence does and does not establish

Published files establish that scoped measured records and a later failed gate are available for inspection. Train loss is not evaluation; file integrity is not quality. The training receipt retains evals=none-this-run for its training step. That historical field does not erase the later generate record. Neither the named-N record nor the merge receipt qualifies this artifact for release. The provenance.json also retains blocked artifact-lineage, consent, privacy-review, deployment, and served-revision boundaries.

Release, autonomy, and deployment boundary

publication_eligible=false and autonomy_eligible=false remain in force. Proposal only: no qualification, promotion, deployment authorization, autonomy, or house-lab pin follows from these records or from public file availability. Lambda uniqueness remains Conjecture 1: open and advisory, never a theorem.

Relationship to other artifacts

This SKU does not overwrite SZL-Khipu-1.5B-BrainNavigator, SZL-Khipu-1.5B, or SZL-Khipu-1.5B-GGUF. The second-brain Space is a separate software retrieval surface, not proof of model or private-graph access.

Source of truth

The immutable Hub evidence linked above and the canonical post-publication receipt govern these claims. The historical source card is retained for provenance; its wording alone is not qualification evidence.