
FACT-CHECK RECORD
When a Shot Gets Blocked: The Doll Island V2 Workarounds, Documented
Our own fact-check gate stopped two shots. Here is what replaced them.
ASSET LEDGER

A 98-file canon library, a highest-version rule, and byte-for-byte archive verification.
Here’s how it started: Season 1 was already delivered, and while we were putting together the results, we presented a retired older cut as if it were the current one. Caught on the spot, of course. The file had been sitting in the Drive root folder with a perfectly respectable name — and the wrong chapters inside. Nobody had been careless. Nobody had ever defined which file counts.
When Night Huntress Season 1 — eight chapters plus a compilation — wrapped, its assets lived in three kinds of places: local working folders, generation-platform history, and leftover exports from earlier deliveries. The same thing existed in several look-alike versions, and anyone picking up the work, including yourself a week later, could grab the wrong one. So we stopped and did the boring thing: an audit. Every asset collected into one canon library, a written rule for versions, and the masters verified into the archive. This article is the full record. It pairs with Article 1, the face-reference workflow: that one covers citation discipline while generating; this one covers governance of the files themselves.
The first page of the audit record is plain: 94 files in the incoming library, all flat, no subfolders, in four groups — 36 storyboard assets for Chapter 3 (first/last frames for nearly all 16 shots), 33 character turnaround sheets (Li Ye’er, Bai She, Mr. Crow and the supporting cast), 22 location plates, and 3 props. That evening the owner uploaded 4 more character sheets — two versions of Chen Du, Qianmian, and White Glove — and I downloaded and eyeballed every one. Total: 98.

Spread out like that, the problem showed itself. Four sets of files shared a name across versions: the Costume A turnaround existed as an original, an r2, and an r3; Lin Xiao, the Crow-49 sheet, and the Yuyan crossbow each had two. Which one is “the face, current version”? Before the rule, the honest answer was: whoever remembers. And in the Drive root sat the mislabeled compilation — 575.81 MB, old chapters inside — which is what ambiguity costs when it grows up. It hadn’t even been saved wrong, exactly. Nobody had ever said it didn’t count.
One more thing the audit surfaced: at that point the library held no finished Season 1 films, no chapter videos, no posters, no logo. That’s by design, and worth stating plainly — the canon library holds source material for generation and reference; finished masters travel a separate archive line (see the verification record below). An audit gets confusing fast if you don’t decide what the library is *for* first.
The decision of October 16, 2026 is short. In practice it runs like this:
| Rule | Content | Error it blocks |
|---|---|---|
| Single source | The Google Drive “Night Huntress / incoming library” (98 files) is the only canon; assets are cited from here and nowhere else | Grabbing files from stale local folders or platform history |
| Highest version wins | When a file name has multiple versions, the highest r value (revision number) is authoritative — no case-by-case judgement | Subjective “this one looks newer” picks |
| Legacy assets retired | All older local videos and assets outside the canon are retired: no longer cited, no longer presented | Deprecated episodes or images passed off as current work |
The middle rule deserves a word, because “no case-by-case judgement” sounds lazy and is anything but. Case-by-case judgement hands the decision to whoever is standing there — who may be tired, may not have been around for that revision, may be you in a week with half the context gone. A fixed rule doesn’t need the right person. It just needs to be followed. The face-reference close-up from Article 1 lives inside this system: the reference is a member of the canon library, so every generation cites the same file and the same face — not whichever copy was nearest that day.
One gate remained. The two Season 1 compilations — the submission cut and the v2-1080p cut, about 1.1 GB each — were archived on October 21, 2026. Because of their size they traveled through Dropbox: downloaded, merged, and uploaded by the owner into Drive’s “Night Huntress / 04_Masters & Deliverables.” Then we read back the Drive-side sizes and compared them with the local masters, byte for byte:
| Master | Bytes (identical on both ends) |
|---|---|
| Compilation, submission cut | 1,116,570,054 bytes |
| Compilation, v2-1080p | 1,095,364,996 bytes |
Not one byte off, either time. It sounds pedantic until you remember how large cross-platform transfers actually fail: a file truncated at 99% looks perfectly normal. A size check is the cheapest integrity evidence there is — no playback, no transcoding, just a number that either matches or doesn’t. Why masters must live in your own storage gets its platform-side telling in Article 8, The Sora Shutdown Playbook: platforms retire; a canon library doesn’t.
Only then did Season 1 close out on the asset side: canon in the library, legacy retired, masters verified in the archive. The work itself is public on the Night Huntress YouTube channel.
Per this site’s house rules, this section stays in:
…[truncated 3127 chars] on this site, The Sora Shutdown Playbook: the platform side of masters-in-your-own-hands — platforms retire; a canon library does not.
My recommendation: run a canon audit after the first season — or first project — of any AI pipeline. Fix a single source, fix a version rule, retire the legacy files, verify the masters into the archive: half a day of work. A wrong-version incident gives no warning; it happens the moment you ship an old cut believing it’s current. The later the rules are set, the more files have scattered, and the more the audit costs. The principles behind this are on the About page; how our articles are structured is on Start Here.
A canon library is the project’s single approved asset source: assets are cited from here only, and “newest” is decided by a written rule, not by anyone’s memory. An AI season produces hundreds of files in multiple same-name versions, so “just use the newest” in practice means “use the one someone believes is newest” — which is how wrong versions get shipped. The Night Huntress canon library (Google Drive, decided 2026-10-16) held 98 files at decision time.
No. It is this project’s version-naming convention: each revision of a same-name file increments its revision number (r value), and the audit rule says the highest r always wins. The point is not the naming scheme — dates or sequence numbers would work too — but that the rule is fixed in advance, with no case-by-case judgement.
No. Retired assets are marked out of circulation: no longer cited, no longer presented. The point is preventing misuse — a deprecated cut being passed off as current work — not destroying records. The files remain where they were, outside the delivery path.
It is not. Files of about 1.1 GB moving through a third-party relay (Dropbox, in this case) — download, merge, upload — can be truncated at any hop and still look normal. Both compilations (1,116,570,054 and 1,095,364,996 bytes) matched the local masters byte-for-byte on the Drive side; only then did we call the archive done.
Yes — and the earlier, the cheaper. The rule set is three lines (single source, highest version wins, retire the legacy), and needs no paid tooling. Setting the rules while files are few takes half a day; waiting until hundreds of files have accumulated turns it into an audit the size of this article.
Reviews, pricing, and hands-on notes for each tool — including what they can do but shouldn’t — are kept current at DC Affiliate.
From canon libraries and version rules to delivery verification, this production governance is available as a done-for-you or advisory service at David Chen Studio.