LoRA adapter that tunes Gemma-4 12B IT to drive the orchestrator role of Spotlight, an open-source OSINT investigation harness, on its constrained 12B local tier. The orchestrator manages the phase pipeline (brief → methodology → research cycles → fact-check → report), delegates to investigator and fact-checker subagents, and pauses at human-approval gates.

What v5 fixes over v4

v4 proved the pipeline and passed three of four targeted capability probes, but its report-drafting probe exposed an execution failure: it correctly identified that report.html was still a template and that unsupported findings needed lower confidence, then repeatedly reached for unsuitable shell edits and failed to land the report changes.

v5 adds corrected continuations in the exact failure contexts:

populate report.html from the case artifacts instead of narrating over a template copy;
align report confidence with the deterministic fact-check verdicts;
use anchored edit/write operations for structured report files;
recover from a failed sed edit by reading the target and using a suitable file-edit operation;
report partial progress honestly and only present completion after the deterministic report validator passes;
re-delegate fact-check corrections when the evidence validator rejects a verdict.

Training

Data: 26 examples — 14 real gold trajectories, 6 v4-authored corrections, and 6 v5 report-execution corrections sliced into real trajectory contexts.
Method: Unsloth SFT with response-only loss masking (train_on_responses_only, <|turn>model spans), LoRA r16, 3 epochs, max sequence length 40,960. Final masked train loss: 0.7510.
Trained: 2026-07-10 on a RunPod A100-80GB (29 minutes of training; adapter tensor verification before upload).

Use

Serve the quantized artifact instead: tomvaillant/gemma4-12b-spotlight-orchestrator-v5-GGUF. Its card documents the exact llama-server flags required by the harness.

Validation status

Released 2026-07-10. The preserved v4 failure fixture now passes without another tune: Spotlight validates the structured evidence chain and deterministically renders all three report artifacts. v5 remains the final behavioral adapter; a fresh full 12B E2E is the post-release system validation.

Limitations

This is a narrow behavioral adapter for Spotlight's tool schema and orchestration prompt, not a general-purpose factuality model. Deterministic fact-check and report finalizer remain part of the deployed system. Unverified information must be treated as a lead, not as a publishable fact.

Lineage

v1 (unmasked SFT) → v2/v2-fixed (masking fix) → v3 (open-tier dataset) → v4 (gate, evidence, and repetition corrections) → v5 (report-gate execution).