Blackline Atlas LFM2.5-VL Adapter
This is the final public Blackline Atlas Liquid-track adapter for the hackathon
submission. It is a PEFT/LoRA adapter for LiquidAI/LFM2.5-VL-450M.
Intended Role
The adapter is a guarded visual site-brief component. It compares a selected
Sentinel/SimSat baseline image and current image, then writes a structured
civilian-scope brief:
- visible scene
- likely visual change
- limitations such as cloud, low resolution, or no-data tiles
- relationship between the source report and what is actually visible
- one of
discard, defer, or downlink_now
It is not autonomous alert authority. Blackline Atlas keeps final triage under
deterministic civilian guardrails and fails closed when model output is invalid.
Training Data
The corpus contains normalized, license-aware training shards for source-led
satellite triage. It is not a raw mirror of third-party datasets. It includes
planner/tool rows, paired-image visual brief rows, hard negatives, and
SimSat/Sentinel examples with provenance/audit notes.
Training split: 30,858 rows.
Eval split: 3,421 rows.
Objective
Supervised fine-tuning for structured visual brief generation over civilian
infrastructure disruption examples. The model is trained to separate source
facts from satellite-visible facts and to prefer low-confidence discard/defer
when imagery is cloudy, low resolution, or not directly evidentiary.
Evaluation Summary
- Eval loss improved from
3.0021 to 0.3273.
- Corpus-native 22-case SimSat gold eval:
22 / 22 valid JSON outputs19 / 22 analyst-schema valid reports9 / 22 action matches
These results support a guarded analyst-narration lane. They do not justify
autonomous alerting.
Runtime Use
Blackline Atlas uses this adapter after a source lead resolves to Sentinel
current/baseline imagery. Contact sheets may be supplied as orientation-only
context. Mapbox tiles are not evidence. SAM/SAM3 masks are not part of the
judge runtime path for low-resolution Sentinel pairs.
Expected JSON actions:
discarddeferdownlink_now
Malformed, tactical, source-only, or low-confidence output should be repaired
only when safe; otherwise it is withheld.
Safety Boundary
Allowed use: civilian resilience, humanitarian logistics transparency, public
accountability, and macro-scale disruption triage.
Disallowed use: tactical targeting, strike support, military asset ranking,
troop/convoy tracking, sabotage guidance, or claims of real-time surveillance
beyond the source data and satellite imagery.
This adapter is not autonomous alert authority and should not be used as a sole
decision-maker for emergency, military, law-enforcement, or targeting decisions.
Limitations
- Sentinel imagery may be cloudy, low resolution, stale, or missing.
- The model can overstate source facts if not guarded; runtime code strips or
withholds source-only casualty/impact claims.
- Action match is not strong enough for autonomous scoring.
- Satellite-visible evidence is macro-scale only. Tiny objects and tactical
interpretations are out of scope.
