How this was assembled, and what it cannot establish
An evidence dossier is only as good as its account of its own weaknesses. This page states the method used for each finding, what that method can prove, and where it stops.
Method, check by check
| Method | What it establishes | Limit |
|---|---|---|
| Playlist metadata extraction yt-dlp --flat-playlist --dump-single-json |
Owner, last-modified date, per-video uploader, duration and view count. Established that the playlist is a personal bookmark collection last touched 2022-02-07 holding seven videos, none uploaded by its owner. | Playlist “last modified” is not an upload date. Single-video metadata calls are frequently blocked by YouTube bot checks from datacenter addresses; where a call failed, the gap is stated rather than filled. |
| Transcript retrieval and verbatim matching | That the fiction disclaimer occupies t = 61–93 s of the original, that it is also present in the 2026 WTQ re-cut, and that the captioned line circulating in clips appears verbatim at t = 856 s. | Auto-generated captions contain transcription error. Timestamps refer to the 2:30:48 original; re-cuts renumber time. A match proves shared source, not which copy someone watched. |
| Archive capture history Internet Archive CDX API |
That a URL existed at a given date. This is the strongest durable signal available on the open web, and it is what makes the chronological argument work. | Absence of a capture is not absence of publication — the Archive’s crawl coverage is uneven. It proves “no later than”, never “no earlier than”. |
| Description-signal analysis | Self-disclosed production method (“Script —”, “Narration —”), named stock-footage vendors, affiliate and course funnels, hashtag stuffing, suppression framing in titles. All quoted from the channels’ own descriptions. | Signals are evidence of a commercial content workflow, not proof of intent, and not proof that a specific frame was machine-generated. |
| Primary-source record checks | Court-martial and sentencing record, fugitive status, the non-existence of “SEAL Team 9” (U.S. Marshals Service); the real SAC Looking Glass alert period (National Park Service); programme existence via agency releases and FOIA documents. | Record checks resolve credentials and programme existence. They do not resolve motive, and they cannot rule out material that has never been released. |
| Tiering against documentary standard | Each of the 31 claims in the taxonomy is placed by one test: does a primary document independently corroborate it? | Tiering is a judgement about evidence quality, not about truth. A Tier 3 claim is undocumented, which is not the same as disproven — and Tier 1 shows how often the outlandish thing turned out to be real. |
What this dossier cannot do
Pixel-level authentication is impossible here
The artifact that prompted this is a phone recording of a screen showing a re-encode of an AI upscale of 2012 consumer video. At that level of degradation nobody can authenticate by artifact analysis — and artifact-hunting on such a file misleads in both directions: it invents synthesis where there is only compression, and it can smooth over synthesis that is really there.
Provenance beats artifact analysis
So the load-bearing argument here is chronological, not forensic: the footage has been continuously online since January 2012, and photorealistic generative video of a speaking human is a 2024-and-later capability. That argument survives any amount of re-encoding. It is also the argument you can reproduce yourself with the verifier.
Attribution to a state actor is not established
The evidence points to commercial content farming: affiliate funnels, self-disclosed stock-footage workflows, serialisation for watch time, hashtag optimisation, generalist channels. This dossier makes no claim of foreign-state involvement, because the record does not support one.
Downstream amplification cannot be excluded
Documented state operations characteristically amplify existing domestic conspiracy content rather than originate it — the pattern in the Russian Internet Research Agency record (real programs §6.1). Nothing observed here indicates state origination; amplification further downstream is outside what public metadata can settle.
Negative findings are bounded
“No record exists” means no record was found in the public and declassified sources checked. That is a real finding — MKULTRA’s records surfaced within three years of a destruction order, and real programmes leak, get subpoenaed and get FOIA’d — but it is not the same as proof of non-existence, and this dossier does not dress it up as one.
The simulator predicts nothing
The Looking-Glass Engine runs synthetic toy dynamical systems with no real-world data feed. Its only claim is about the shape of predictability itself: that the horizon moves outward logarithmically in measurement precision. Every forecast it emits carries an expiry date and a named falsifier for exactly that reason.
Rules this dossier holds itself to
- Every factual claim carries a link to a primary or authoritative source.
- Quotations are reproduced exactly, including hesitations and disfluencies. Nothing is paraphrased inside quotation marks.
- Contractor marketing figures are labelled as such — Lockheed’s “>80 percent accuracy” claim for ICEWS is a claim, not an audited result (Wired’s real-world test found 4 of 16).
- Failed predictions are reported as failures. The 2012-12-21 convergence did not happen.
- Where a check did not work, the failure is reported rather than filled with a guess. The verifier never fabricates a result or shows a confidence score it cannot compute.
- Uncomfortable facts on all sides are included — including that the community circulating this material knew about the conviction, and that the described capability is genuinely real but bounded.