Backfill reconciliation and the Forge model roster

The daily summary for 2026-07-18 was regenerated in bounded backfill mode. This means I refreshed the fingerprints, source counts, and sampled event structure without running a full LLM narrative pass. The goal was to ensure the daily record is accurate and complete before diving into the day’s specific operational tasks. This approach keeps the log grounded in concrete data rather than abstract summaries.

Operational Maintenance

The platform health steward flagged several SLA breaches that needed attention. The logbook observations source had blown its 1.5-hour sync tolerance, with an unknown auth status blocking the ingest route. I investigated the packages/api/src/routes/ingest.ts file to understand why the sync was stalled. The verification query confirmed that the observation source was no longer past its SLA tolerance, indicating a recovery.

Another issue involved the kg-salience-recompute job type, which had two dead-letter entries. The handler was now passing with a success rate above 0.8 over a 1-hour window, making these obsolete. I cleared the dead letters, which is a routine maintenance task that keeps the job queue clean and prevents false alarms from lingering failures. The memoir prompt delivery lane also showed no recent posts in the 168-hour lookback, suggesting a stall in the scoped delivery for the source/category lane. I flagged this for further investigation into the producer, desk-delivery, and gateway permissions.

Forge Homelab Model Scout

The main work of the day centered on the Forge homelab stack. I ran a HuggingFace scout pass to identify new local models that could fit the current hardware constraints: AMD Strix Halo/gfx1151 with 128GB unified RAM and 96GB usable VRAM via Vulkan. The scout was strictly in reconnaissance mode; no models were downloaded or manifest changes made during the initial shortlist.

The high-signal candidates included:

• nvidia/parakeet-tdt-0.6b-v3: A transcription model, described as the top pick and lowest-risk candidate for benching or reverting.

• unsloth/gemma-4-26B-A4B-it-GGUF: A chat and VLM model already present in the homelab manifest but not yet wired into forge.conf.yaml.

• ResembleAI/chatterbox (Multilingual V3): A TTS model that could serve as a drop-in replacement for the existing Chatterbox service with multilingual support.

• neuphonic/neutts-air: A lighter TTS path that is GGUF-native and runnable directly on llama.cpp/Vulkan.

The scout explicitly rejected NVFP4 Qwen3.6 quants because they were CUDA/Blackwell-only and lacked a Vulkan path. The workflow moved from pure reconnaissance into cautious implementation planning, with Claude Code identifying that the Gemma model had previous local history rather than being a new addition, and that Parakeet would require a real build plus a new systemd unit rather than a simple config change.

Verification and Tradeoffs

The verification for the observation SLA breach used a NOT EXISTS query to confirm recovery, which is a reliable pattern for checking if a source is no longer stale. The tradeoff in the model scout was balancing the desire for new capabilities against the strict hardware constraints. The 96GB VRAM limit ruled out many of the larger recent releases, forcing a focus on mid-size or non-LLM roles like transcription and TTS. This constraint-driven approach ensures that any new models added to the stack are actually runnable on the existing hardware, avoiding the common pitfall of selecting models that are too large for the available resources.