tube-bridge-research

Tube Bridge Research

This is the single entry point for the MCP, research method, corpus doctrine, examples, and reusable templates. Treat tube-bridge as one research capability rather than a bag of unrelated tools.

Select the research mode

• Quick lookup: one bounded fact about a source; use explicit language and timestamp provenance without manufacturing hypotheses.
• Focused study: one source or tightly bounded question; use FRAME-LOCK and claim classification.
• Comparative investigation: disputed, causal, or multi-source question; require source lineage, competing hypotheses, and adversary review.
• Living corpus: repeated monitoring; additionally preserve state snapshots, update traces, review intervals, and resolution criteria.

Read references/methodology/00-research-method.md before a focused, comparative, or living investigation.

Non-negotiable rules

• Keep OBSERVATION, FACT, SOURCE-CLAIM, INFERENCE, and UNKNOWN distinct.
• Freeze a timestamped frame_id before evidence retrieval. Changed questions, standards, applicability, or materiality create a new prospective run; they never rewrite the old verdict.
• Treat a timestamped transcript as tool-observed evidence of what the selected subtitle track renders. Generated, translated, or dubbed text is not automatic proof of exact speech, speaker endorsement, or a world claim.
• Check available subtitle tracks and pass lang explicitly whenever the intended language can be established. With lang=None, v1.1.0 deterministically stays in the first ASR/default language family and never crosses into an unrelated manual dub; explicit selection still gives stronger provenance.
• Preserve video identity, requested and returned language, track type, timestamp spans, retrieval warnings, and upstream source lineage.
• Count independent observation paths, not videos. Several videos can repeat one origin.
• Verify corpus hits against source transcript text. Similarity scores rank candidates; missing hits do not prove absence.
• For causal or predictive questions, keep competing mechanisms and seek distinguishing evidence.
• Attack the leading explanation before synthesis. Core frame, provenance, confidence, and stopping checks are mandatory-material. A gate PASS means fit to present, not certainly true.
• Do not claim that Corpus v2 hierarchy, dual representation, lexical search, or logical relations exist in the released runtime. Current corpus behavior is flat overlapping-window vector retrieval.
• Do not fabricate unavailable transcripts, speaker identities, frames, provenance, certainty, or narrative glue.