Private research infrastructure

Turn a research brief into structured evidence.

Signal Room runs a four-stage video evidence pipeline: discover leads from YouTube, pull captions, extract contradictions with an LLM, and curate the strongest signals in a review queue. All of it stays off the open web.

Discover YouTube search across staged query batches, never re-running a used query
Caption YouTube captions fetched via InnerTube, parsed and stored timestamped to D1
Analyze LLM reads caption text and extracts claim pairs, actor stances, contradiction signals
Review Keyboard-driven curation queue — star, annotate, and score evidence items
Pipeline active D1 indexed Access gated
How it works
Discover

Query-level search with saturation tracking.

The orchestrator generates search batches from your brief using an LLM and dispatches them to YouTube.

  • No repetition: every query run is logged to D1 and filtered out of all future batches
  • Yield feedback: how many new unique leads each query produced informs the next generation prompt
  • Expanding coverage: each cycle explores angles the previous batch didn't touch
0×Duplicate queries
YT APISearch backend
Caption

YouTube auto-captions, fetched and timestamped.

For each discovered video the pipeline fetches the auto-generated caption track using InnerTube, YouTube's internal request layer.

  • No API quota: InnerTube bypasses the official Data API entirely, no rate limits or keys required
  • Full text stored: caption XML is parsed into timestamped segments and written to D1
  • No audio processing: YouTube's own captioning does the work; coverage follows wherever auto-captions exist
YTCaption source
±2sTimestamp precision
Analyze

LLM extracts contradictions across sources.

Each caption batch is sent through a structured LLM prompt that extracts claims, stances, and contradiction pairs.

  • Claim extraction: factual assertions are pulled with speaker attribution and timestamp
  • Contradiction flagging: direct conflicts with prior statements are stored as pairs with confidence scores
  • Async pipeline: analysis runs on the VPS via router.darra.ai and surfaces in the dashboard once D1 is written
N:NPair matching
0.1–1.0Confidence range
Intelligence

Entity aggregation across the entire corpus.

The intelligence layer operates above per-video analysis, aggregating extracted claims by named entity across the entire campaign corpus.

  • Entity timelines: every statement attributed to a person or organization is collected chronologically across all videos, not per-video
  • Stance drift detection: where an actor's stated position on a specific topic changed between two time-separated appearances, the delta is flagged
  • Cross-source contradictions: conflicts between what two different entities claim about the same event surface as corpus-level findings invisible at the per-video layer
  • Corroboration mapping: which actors are reinforcing or undermining each other's claims, scored by frequency and confidence
CrossSource matching
EntityAggregation
Saturation

Every query logged. No search runs twice.

The saturation tracker maintains a complete, per-query history of everything the system has ever dispatched.

  • Hard deduplication: before any new batch is staged, every candidate query is filtered against the full history log
  • Yield scoring: how many new non-duplicate leads each past query produced is stored and fed back into the next LLM generation prompt
  • Depth estimation: the Query Plan view in the dashboard shows remaining estimated search depth per topic cluster and flags when a cluster approaches saturation
0×Query reuse
YieldScore feedback
Review

One item at a time, keyboard-driven.

The review queue presents one evidence item at a time: video, captions, and extracted contradiction signals together.

  • Keyboard-driven: arrow keys to navigate, S to star, Enter to mark seen and advance
  • Annotation layer: labels, notes, and a 0–10 operator score write back to D1 per video
  • Starred export view: a separate queue collects all starred items for structured export
⌨Keyboard nav
D1Durable storage

Operator access only.

This workspace is private. If you have credentials, authenticate to open the pipeline console.

pipeline active · access gated