Verified Evidence Operations Core & Project Evidence Corpus Builder for Video Research
Engineered for technical researchers, PhD students, and investigative analysts synthesizing multi-video lecture series, conference keynotes, and technical interviews into grounded corpora for Google Gemini Gems and NotebookLM. Unlike 1-click YouTube AI summarizers that emit unverifiable prose, YouTube → Gemini Gem preserves a 1:1 raw_lines audit trail, millisecond timecodes, and fail-closed missingness flags.
Preview stage: the timeline workbench below runs on a 7-segment synthetic dataset (VIDEO_ID: SYNTHETIC-ONLY) including an explicit caption dropout at seg-005 (00:01:52–00:02:19, source_status: "unavailable"). Epistemic guardrail: Source Traceability (Provenance) ≠ Factual Validity.
Interactive browser workbench available below · Zero account required
Primary Job-To-Be-Done
Compile multi-video lecture and interview transcripts into timestamp-anchored Markdown and JSONL corpora with 1:1 raw_lines provenance and fail-closed missing-caption ledgering.
Key Capabilities
- Solves the 4 Structural Breakdowns of Raw Transcript Ingestion: (1) Loss of Temporal Provenance (start_ms / end_ms stripped), (2) ASR Fragmentation & Mid-Sentence Line Splits, (3) Silent Omission of Missing Audio Intervals, and (4) Cross-Video Context Drift across 10+ transcripts.
- Deterministic 4-Stage Pipeline: Canonical Source Manifest (video_id, segment_id, monotonic millisecond offsets) → Deterministic Segment Normalization (preserving a 1:1 raw_lines audit trail) → Explicit Missing-Interval Ledger (coverage.json) → Dual-Format Local Commit (Markdown + JSONL).
- Fail-Closed Missingness Contract: when captions drop out (e.g., seg-005 from 00:01:52–00:02:19), the compiler emits source_status: "unavailable" and grounding: "source_unavailable" in coverage.json rather than hallucinating a bridging summary.
Privacy & Local Data Boundary
Static demonstration uses 7 original synthetic transcript segments (SYNTHETIC-ONLY). No YouTube player iframe, external API call, or user playlist URL is requested.
01. INTERACTIVE ARCHITECTURE SIMULATOR (SAMPLE FIXTURE)
Inspect the transformation and boundary handling live
This interactive workbench runs locally in your browser on built-in synthetic fixtures.
Timestamped Transcript-to-Corpus Inspector (Example Research Recording)
01 · Canonical Source Manifest
Pins video_id, segment_id, and monotonic start_ms / end_ms boundaries to prevent cross-video context drift.
02 · 1:1 `raw_lines` Audit Trail
Normalizes fragmented 2s ASR cues into coherent paragraphs while retaining a 1:1 raw_lines provenance link.
03 · Fail-Closed `coverage.json`
Uncaptioned intervals (seg-005) fail closed as source_status: "unavailable" inside coverage.json.
04 · Traceability ≠ Factual Validity
Dual-format Markdown + JSONL commit: Source Traceability (Provenance) verifies recording offset, distinct from Factual Validity.
Select a synthetic recording segment (7 intervals):
NORMALIZED CORPUS BLOCK · seg-003
Source range recording
The next step records the source range.
MATCHED RAW CAPTION CUES (00:00:56–00:01:27)
The next step records the source range.
Static All-States Reference Summary (No-JS & Audit View) — All 7 Segments (seg-001..seg-007, including seg-005 unavailable)
| segment_id | timecode_label | offset_ms | source_status | topic_heading / missing_reason |
|---|---|---|---|---|
| seg-001 | 00:00:03–00:00:18 | 3000..18000ms | synthetic | Problem framing: why flat transcripts lose context |
| seg-002 | 00:00:18–00:00:34 | 18000..34000ms | synthetic | Segment identifier assignment |
| seg-003 | 00:00:56–00:01:27 | 56000..87000ms | synthetic | Source range recording |
| seg-004 | 00:01:27–00:01:52 | 87000..112000ms | synthetic | Paragraph normalization rules |
| seg-005 | 00:01:52–00:02:19 | 112000..139000ms | unavailable | Source unavailable / verification not possible: caption track missing for 00:01:52–00:02:19. Pipeline refuses to interpolate or claim Proof of Absence. |
| seg-006 | 00:02:19–00:02:45 | 139000..165000ms | synthetic | Corpus handoff format (Markdown & JSONL) |
| seg-007 | 00:02:45–00:03:10 | 165000..190000ms | synthetic | Human review boundary |
02. PROBLEM & INTENDED OUTCOME
The 4 Structural Breakdowns of Raw Transcript Ingestion & 1-Click Video Summarizers
Generic 1-click YouTube AI summarizers and raw transcript copy-paste break serious research in 4 ways: (1) Loss of Temporal Provenance strips start_ms/end_ms offsets so claims cannot be verified at 00:00:58; (2) ASR Fragmentation splits single thoughts across dozens of 2-second caption cues; (3) Silent Omission of Missing Intervals causes LLMs to hallucinate bridging prose over audio dropouts (like seg-005 at 00:01:52–00:02:19); and (4) Cross-Video Context Drift collapses speaker attribution when loading 10+ transcripts into a Gemini Gem or NotebookLM.
Engineered deliverable goal
YouTube → Gemini Gem operates as a deterministic Evidence Corpus Builder: it normalizes fragmented ASR cues into topic-coherent paragraphs while retaining a 1:1 raw_lines audit trail, records unavailable intervals fail-closed in coverage.json (source_status: "unavailable", grounding: "source_unavailable"), and exports dual-format Markdown + JSONL packages.
- 01.1:1 raw_lines audit trail linking every normalized corpus paragraph back to exact start_ms / end_ms ASR cues
- 02.Fail-closed missingness in coverage.json (seg-005 marked source_status: "unavailable", never hallucinated)
- 03.Epistemic clarity: Source Traceability (Provenance) verifies recording location, distinct from Factual Validity
03 & 04. ARCHITECTURE, PERMISSIONS & KNOWN LIMITATIONS
Inspect Technical Pipeline, Permissions & Known Limitations
03 & 04. ARCHITECTURE, PERMISSIONS & KNOWN LIMITATIONS
Inspect Technical Pipeline, Permissions & Known Limitations
How the pipeline handles structure and unknown states
- Stage 01
Canonical Source Manifest
Indexes video_id, immutable segment_id, and monotonic millisecond boundaries (start_ms, end_ms) across each source recording.
Boundary note: Private, restricted, or uncaptioned intervals are logged immediately for missingness tracking. - Stage 02
Deterministic Segment Normalization (1:1 raw_lines)
Consolidates fragmented 2-second ASR caption cues into coherent topical paragraphs while retaining a 1:1 raw_lines reference inside every corpus block.
Boundary note: Auto-generated captions may contain speech-to-text recognition errors. - Stage 03
Explicit Missing-Interval Ledger (coverage.json)
Records uncaptioned or dropped intervals (such as seg-005 at 00:01:52–00:02:19) fail-closed with source_status: "unavailable" and grounding: "source_unavailable".
Boundary note: Missing intervals are never filled by speculative LLM completion or claimed as Proof of Absence. - Stage 04
Dual-Format Local Commit (Markdown + JSONL)
Emits human-readable Markdown study briefs alongside machine-ingestible JSONL records ready for Google Gemini Gems and NotebookLM.
Boundary note: Source Traceability (Provenance) verifies where a quote occurred, not external Factual Validity.
Explicit scope and verification requirements
| Area | Why accessed | On-device handling | Verification note |
|---|---|---|---|
| Synthetic transcript fixture | Demonstrate timestamp-to-corpus alignment and missing-segment handling. | Bundled as static TypeScript data; zero network requests. | Verified in source and E2E tests. |
| Future video caption ingestion | Read user-authorized transcript files or compliant public caption tracks. | Planned local normalization into Markdown/JSONL. | Requires provider API policy and content-rights review before implementation. |
Known Limitations
- Preview concept; does not claim an active YouTube API integration, rights-clearance engine, or automated Gemini Gem upload.
- Epistemic Guardrail — Source Traceability (Provenance) ≠ Factual Validity: timestamp anchors verify where a statement occurred in the recording, not whether the speaker’s statement is externally true.
- Missing or restricted transcript segments are marked unavailable, never certified as Proof of Absence.
05. FREQUENTLY ASKED QUESTIONS
Common questions about YouTube → Gemini Gem
Can this pipeline prove that statements in a video are true?
No. Timestamped source links help a human reviewer trace a derived note back to the exact segment offset, but they do not verify factual accuracy or eliminate downstream model hallucinations.
What happens if a video segment lacks transcripts?
The pipeline marks that range as source_status: unavailable (verification not possible). We explicitly reject any Proof of Absence guarantee because an unindexed interval does not prove a topic was never discussed.
Does this website connect directly to my Gemini Gem or YouTube account?
No. Provider integration is not evaluated in this candidate. The page demonstrates the data contract using synthetic segments.
What happens to private or restricted playlists?
This website cannot access private playlists or accept playlist URLs. Any future tool would only process user-supplied or explicitly permitted material.
Need production AI document exports today? Noxport AI Exporter is Available Now.
While this R&D workbench is in preview/development, our flagship Noxport AI Exporter (v3.0.4.226) is shipping today with 6 Basic Formats Free Forever and 4 Pro Formats (Word .docx, Vector .pdf, Pro .md, Professional Pack .zip).
NOXPORT LABS R&D PORTFOLIO · RELEASE NOTIFICATIONS
Follow R&D workbenches in the Noxport Labs portfolio
Inspect interactive synthetic workbenches in your browser or choose specific R&D projects below to receive release notifications.
AnswerBell AI
Monitors long-running deep-research streams across a 3000ms DOM quiet window so you know when a thread is ready for Noxport AI Exporter compilation.
Vitalis Health Studio
Demonstrates a local read-only SQLite MCP policy boundary that allows bounded 28-day aggregates while blocking raw record dumps with JSON-RPC -32001.
CardioLab Workstation
Pairs raw vs. detrended synthetic ECG/RR signal traces with a 6-event Beat Ledger (BL-01..BL-06) and SHA-256 provenance manifest.
Explore YouTube → Gemini Gem architecture & documentation
Inspect our demonstration architecture documentation or explore another project in the Noxport Labs portfolio.
Div I · Available
Noxport AI Exporter
Turn 50–150+ turn ChatGPT, Claude, Gemini, Perplexity, Grok, DeepSeek, Copilot, and Qwen sessions into clean, editable Word (.docx), searchable vector PDF, Pro Markdown, and .zip bundles locally in your browser. Engineered for staff/principal engineers, independent technical consultants, AI solution architects, and computational researchers — with zero cloud uploads, no account required, and zero silent truncation.
Div I · In development
AnswerBell AI
AnswerBell AI eliminates the "Babysitting Tax" and context-switch fatigue when running multi-minute reasoning models (o3-pro, Deep Research, Claude Extended Thinking) across 10–30 open browser tabs. Unlike naive notification extensions that trigger false-positive chimes on a single Stop-button flicker, AnswerBell reconciles multi-signal DOM evidence over a 3000ms quiet window to distinguish true completion from mid-run tool pauses.
Div III · Preview
Vitalis Health Studio
Engineered for quantified-self engineers, privacy-conscious longevity researchers, and AI builders who own multi-year archives from Apple Health (export.xml), Garmin (.fit), Oura, and Whoop. Vitalis resolves the dilemma between Vendor Truncation (mobile apps locking your data inside opaque 7-day readiness scores) and Cloud AI Exfiltration (pasting raw multi-year CSV/XML dumps into cloud LLMs).
Div III · In development
CardioLab Workstation
Engineered for human-performance researchers, sports-science engineers, academic physiology labs, and HRV algorithm developers working with chest-strap RR intervals (Polar H10 .csv / .fit) and raw single-lead ECG waveforms (.edf / .csv). CardioLab eliminates the Black-Box Smoothing Hazard by pairing immutable raw baselines (SHA-256) with a synchronized Raw / Processed / Overlay comparator and a 6-event human-in-the-loop Beat Ledger.