Methodology

Process

Each lecture in the Predictive History corpus is analyzed through a systematic, repeatable pipeline:

1. Transcript extraction — Auto-generated English subtitles downloaded via yt-dlp in SRT format with timestamps.
2. AI analysis — The full transcript is provided to Claude Opus 4.6 with the analysis schema and scoring rubric. The model reads the entire transcript and produces a structured JSON assessment.
3. Screencap extraction — For every timestamped item (quotes, rhetoric, predictions), a video frame is captured automatically via ffmpeg.
4. Publication — JSON is rendered to static HTML. No editorial intervention beyond the rubric design.

Prediction Tracking & Calibration

Every falsifiable prediction and claim extracted from the corpus carries a status — confirmed, partially_confirmed, disconfirmed, untested, or unfalsifiable — re‑evaluated against real‑world events. The reference document that drives scoring is briefing-data.json, accompanied by a human‑readable calibration reference and a long‑form geopolitical briefing. None of these files are hand‑edited; they are machine‑maintained by Claude Opus 4.6 under two scheduled runs.

Update cadence

How subject matter enters the briefing

The daily script's prompt hard‑codes a tracking list: Iran / Strait of Hormuz, dedicated ground‑invasion indicators (troop movements, Kharg Island, amphibious assets, draft signals, logistics staging), Gulf states, the Russia–Ukraine war, Latin America (Venezuela / Cuba / Colombia), US–China and Taiwan, the Korean peninsula, and countries exposed to a Hormuz blockade (Japan, South Korea, India, Pakistan, Bangladesh, Europe). For each theatre Claude runs web searches for the day's developments and decides what's material.

Subject matter broadens three ways:

1. A new daily entry. Each record carries its own date / tags / summary / details / prediction_impact / source — additive, no schema change.
2. A new theatre. The prompt explicitly allows Claude to add new theatre sections when a story grows large enough to warrant separate tracking.
3. A new metric. The metrics array is free‑form; entries are added or retired as the situation evolves (e.g. Ships Trapped in Gulf, Qatar Gas Production were added after the Gulf crisis escalated).

The prediction_impact field on each daily entry is the bridge to the predictions tables — it's the claim‑to‑evidence linkage the monthly run consumes when re‑scoring statuses.

Sources used to monitor falsifiable claims

Sources are cited per entry in the source field of each daily record, not a fixed global list. Recurring outlets across recent entries:

• Wires & general news: Reuters, AP, Bloomberg, CNN, CBS, NBC, ABC, Fox News, NPR, PBS, Washington Post, The Hill, Axios, Time, Newsweek, Fortune, CNBC, BBC, Al Jazeera, RFE/RL, Euronews, CBC, Globe and Mail, Irish Times.
• Regional / language‑of‑record: Haaretz, Times of Israel, Ynetnews, JPost, Israel Hayom, i24, Middle East Eye, Al Arabiya, Al Bawaba, The National, Gulf Times, Naharnet, L'Orient Today, PressTV, Tasnim, TASS, South Front, Turkiye Today, Arab Times, Pakistan Today, Express Tribune, Tribune India, ANI, WION, The Week India, BSS, Manila Times, Taipei Times, TVBS.
• Defense & maritime specialist: USNI News, The War Zone (TWZ), Naval News, Stars and Stripes, Navy Times, SOFREP, SOF News, The Aviationist, gCaptain, SimpleFlying, FlightGlobal.
• Official / primary: CENTCOM releases, U.S. State Department, UN News, Human Rights Watch, Amnesty International, IMF, HRANA, Iran International.
• Markets & commodities: TradingEconomics, Bloomberg, Investing.com, CoinDesk.
• Geolocation / OSINT: LiveUAMap, Mezha, Mappr.

The baseline anchor set used for the calibration reference is narrower — Al Jazeera, BBC, Reuters, NPR, CNN, Washington Post, CNBC, plus institutional research from CRS, CSIS, Atlantic Council, Chatham House, Carnegie Endowment, and the Arms Control Association — but the tooling does not enforce an allow‑list.

What the tooling does not do

• No source‑quality filter, no automated deduplication, no independent fact‑checker in the loop — source weighting is delegated to Claude's judgment on each run.
• No RSS or API ingestion. Every source is pulled ad hoc via web search during the run; the list of outlets above is descriptive of recent behavior, not prescriptive.
• If the daily run produces invalid JSON, the script reverts briefing-data.json from git. Bad runs are self‑healing, but nothing is automatically retried.
• Broadening what's monitored means editing the tracking list in the daily script's prompt. The current list reflects the theatres the project judged most consequential for the tracked predictions; it is not a claim of global coverage.

Scoring Rubric

Each lecture is evaluated on seven axes, scored 1–5. Higher is better. Click any score on a lecture page to read the full justification.

Epistemic Asymmetries We Try to Name

An audit project has defaults. Ours are these: we weight documented events over circumstantial ones, mainstream reporting over adversarial reporting, and confirmed facts over alleged ones. These are reasonable defaults. They are also defaults with known failure modes, and an honest audit should name them instead of pretending they don't exist.

The October Surprise trap

In 1980, allegations that Reagan campaign figures negotiated with the Islamic Republic to delay the US embassy hostage release until after the November election were dismissed as conspiracy theory by most of the press. The allegations were partially corroborated across the 1990s — first by Gary Sick's 1991 book, then by a 1992 Senate investigation that found “credible evidence” of back‑channel contact, and again in March 2023 when Ben Barnes told the New York Times on the record that he had accompanied former Texas Governor John Connally on a 1980 Middle East trip delivering exactly that message. The category of the allegation shifted from conspiracy theory to plausible, partially supported, incompletely documented.

The lesson for an audit project: a posture that reflexively codes unverified → wrong will systematically undercount exactly the kind of event that takes decades to settle. For a project auditing a speaker whose recurring theme is cover‑ups and hidden strategic logic, that asymmetry is not neutral — it silently biases the audit against the speaker's category of claim while taking the official rebuttal at face value. A reader who senses that tilt will lose trust in the audit, not in the speaker.

What we do about it

We do not validate rumors. We also do not silently dismiss them. The schema and site carry four specific affordances to keep the middle ground honest:

• contested_unresolved prediction status. Reserved for claims that (a) cannot be confirmed from open sources, (b) cannot be ruled out from open sources, and (c) are being publicly asserted by non‑fringe actors. The status_note names the documented facts the claim rests on and the evidence that would resolve it. Used sparingly to avoid either endorsement or reflexive dismissal.
• Open Questions page. A standing register of questions where the documented record and a publicly‑circulating counter‑reading both remain in play. Each entry lists its documented facts, the alternate reading, who asserts it publicly, what would resolve it, and our honest current assessment. Entries leave the page when new evidence promotes them to confirmed or disconfirmed.
• alternate_readings on daily briefing entries. Where a documented event is being read differently in public discourse, the briefing records both readings side by side — with evidentiary status, who is asserting each, and why the question is (or isn't) resolvable. Noted, not endorsed.
• steelman on the verdict of each lecture. Where a lecture's specific claims and reasoning fail but the underlying concern is legitimate or shared by mainstream scholarship, we say so. A viewer who heard the lecture and felt something real beneath the argument deserves a rigorous frame for that intuition, not a dismissal that pretends the intuition wasn't there.

The limits of the limits

These affordances still privilege open‑source evidence. We cannot, from public reporting, prove or disprove a genuine cover‑up; the best we can do is refuse to pretend an unresolvable question is resolved. If decades from now a declassification moves a current Open Question into confirmed or disconfirmed, the audit trail should show that the question was named honestly while it was open. That is the standard we hold ourselves to.

Limitations & Caveats

• AI analysis is not infallible. LLMs can misidentify rhetoric, misjudge accuracy on niche topics, or miss cultural context. Every justification is published so readers can evaluate the reasoning.
• Auto-generated transcripts contain errors. YouTube's speech-to-text occasionally garbles names, technical terms, and non-English words. Where a quote seems off, check the video.
• Scores reflect the rubric, not a verdict on the speaker. A low score on one lecture does not characterize the entire channel. Patterns emerge from the corpus, not from individual data points.
• This is content analysis, not mind-reading. We analyze what was said, how it was framed, and what was omitted. We do not speculate about the speaker's intentions or affiliations.
• The schema may evolve. If the rubric is refined, earlier analyses will be re-run and the version history published.

Reproducibility

Every claim in every report is traceable to a timestamp in a public YouTube video. The analysis schema is published — anyone can feed the same transcripts and rubric to an LLM and compare results.

Fair Use Notice