Darkfield THINK
— a model that designs the pipeline.
THINK is a multimodal language model trained to read industrial documentation, watch camera streams, author detection schemas, and supervise SEE. It runs when reasoning is needed — not on every frame.
What THINK does.
Ingests business context before detection begins.
THINK reads PDFs, SOPs, and KPIs. It asks clarifying questions in plain language, negotiates the output schema, and proposes notification rules — before a single frame is processed. The onboarding conversation is the specification.
Watches the stream; decides if the camera is sufficient.
THINK pulls a frame from the stream and evaluates it against the declared task. If the angle is wrong for a plate read, or the lighting is insufficient for small-object detection, it says so — and tells you where to put a different camera.
Composes the detection graph itself.
THINK chains detect / track / OCR / count / zone steps into a runnable pipeline. It writes the output schema. No CV engineer authored the graph — the model did.
Points at objects; resolves language to coordinates.
THINK can take a camera frame and locate a named zone, object class, or spatial relationship in pixel space. This grounds abstract language — "the staging area near dock B" — into the exact polygon SEE uses for zone detection.
Evaluates SEE; decides when and how to retrain.
THINK pulls a rolling sample of recent events and scores SEE's output against ground truth it authored. Below tolerance, it diagnoses the failure mode — labelling, coverage, or model-fit — and launches a per-camera finetune accordingly. It can also swap to a different base model when one fits the problem better.
Reasoning over
an operations onboarding.
A lightly edited trace from an actual onboarding session. Partner details anonymised. The full trace runs to approximately twelve turns; this is an excerpt.
CAM-11: angle insufficient — barcodes face away from camera at current position. Recommend repositioning to south wall, 2.4m height.
Meanwhile, here's a draft schema for your approval:
table pallet_positions · columns: camera, timestamp, pallet_id, zone, last_moved, dwell_minutes, alert_flag
pipeline · detect: pallet · track: on · OCR: barcode · zones: bay_a, bay_b, bay_c, yard · dwell threshold: 120 min alert · tier: P2 · channel: SMS · condition: dwell ≥ 120 min · recipient: shift_manager
Architecture notes.
THINK is a multimodal language model with a vision encoder, designed to orchestrate SEE end-to-end. The architecture is a transformer-class model trained on a combination of real industrial CCTV footage, annotated SOP corpora, synthetic operational scenarios, and reasoning traces generated from partner onboarding sessions.
The context window is sufficient to hold a full SOP document, multiple camera frames, and the running plan simultaneously. Latency target is seconds-to-minutes per planning turn, and under five seconds for a single-frame stream-quality check.
Known failure modes.
THINK performs poorly on novel verticals not represented in training — highly bespoke manufacturing processes with no analogue in the SOP corpus are the current outer limit. Mitigation: the onboarding window keeps a human in the loop until the schema has been verified against at least 500 real events.
All autonomous actions are logged to an immutable audit trail. Every per-camera finetune records the evaluation that triggered it and the validation that cleared it. The audit log is visible to the partner — the model itself is not user-tunable, by design.
Model card.
| model | Darkfield THINK |
|---|---|
| role | Planner · pipeline author · finetune controller · alert dispatcher |
| system | System 2 — runs on reasoning events, not on every frame |
| modality | Multimodal — text, image, video frames, structured documents (PDF, DOCX) |
| context window | Sufficient to hold a full SOP, multiple frames, and the running plan |
| latency | Seconds–minutes per planning turn · <5s per single-frame stream-quality check |
| params | undisclosed in private beta |
| training data | Real industrial CCTV · SOP corpora · synthetic operational scenarios · partner onboarding traces |
| availability | Private beta · partner access only |
// citations and linked papers → research.html#papers
See THINK plan
against your operation.
We're onboarding a small number of partners in private beta.