The Guardian
of the Archive

HeyGen's Avatar Shots composer reloads with whichever avatar was used last, not the one pinned to this project. It reverted to a different avatar, Tomás, at least three separate times across this production — the very first Shot 1 submission, the start of the five-shot batch after the credits upgrade, and again at the start of the full six-shot reshoot. Each time it was caught the same way: a screenshot check of the avatar chip before typing the prompt, never after.

Cinematic mode has no persistent "previous generation" reference, so the only thing holding the Guardian of the Archive's face, wardrobe and set consistent across seven separate generations is manual discipline — reselecting "Oracle" via the avatar chip, "My Avatars," "Choose look," repeated before every single submission — and a near-verbatim environment paragraph doing the rest (§05).

The Creator plan's Cinematic/Seedance quota was exhausted after a single shot, mid-session, with five more generations still ahead. The fix — an upgrade to a Pro plan, Monthly billing rather than Yearly, sized for a defined production sprint — cost more than credits: it also lost the exact Shot 2 prompt in progress, since HeyGen keeps no recoverable draft or prompt history on a submission interrupted by a paywall. That prompt had to be rebuilt from what was already approved rather than guessed at.

How it was caught: the generic "credits left" counter shown elsewhere in the account didn't warn before the specific Cinematic quota gated the feature outright — the two counters aren't guaranteed to track the same pool. The checkout screen's "Payment Processing…" spinner then sat indefinitely even though the charge had already succeeded — confirmed only by reloading the account panel and checking the plan name and balance directly, not by waiting on the spinner.

Standing note: the upgrade inherited the old plan's renewal date rather than starting a fresh cycle, so the next full charge landed one day later, not four weeks later. The plan is due to come back down to Creator once the current production batch is finished — not yet triggered.

"It needs to be rebuilt. It is too flat; the voice is flat. Increase the volume of the voice. There are some shadows on the face; the scene is too dark, which I do not like… The voice is creepy; it has to be changed… There is no catching audience, no traction of the video."

Four structured questions went to the team before any resubmission: the new title-card wording, how far to soften the lighting, the target voice tone, and whether to add sound design at all. The recommendation given — and approved: keep the noir setting but add a soft fill light so the face stays readable; replace "quiet, measured, without warmth" with "warm, urgent and grave, protecting the listener, never menacing"; add sound design, since a fully silent track was part of why the first cut felt inert; and give the episode its own title card, "ERUDITA — The Guardian of the Archive," rather than reusing Episode 2's "The Price of Silence."

Fix: all six shots regenerated with the revised environment paragraph and voice direction (§03, §05); a contact sheet confirmed the face legible in every frame while the room itself stayed dark.

After the reshoot was approved, the team flagged it directly: "there is a trim audio problem in the transition from the first to the second video, and eaten words in the first video." Whisper's word-level timestamps on every clip showed the real cause: Shot 1 has no dialogue, but HeyGen's render wasn't true digital silence — it carried faint page-turning and breath sound at roughly −45 LUFS. Loudness-normalizing that "silent" track to the same −14 LUFS target as the dialogue clips boosted it by roughly 27 dB, and a uniform 0.5-second crossfade at every join then blended that now-audible noise straight across the first word of Shot 2.

The same uniform-crossfade approach turned out to be wrong at every other join too, for the opposite reason: Shots 3 through 6 all begin speaking at literal frame zero, with no lead-in silence at all — so any crossfade at the start of those clips fades in the first word itself rather than blending silence into silence.

Fix: Shot 1's audio was replaced with true digital silence rather than a loudness-normalized "quiet" track. Every other join was recomputed per-shot from each clip's actual measured trailing silence (last-word-end-time to clip duration, via whisper), and rebuilt as trim-plus-concat with only a ~15-millisecond click-guard fade positioned inside the outgoing clip's real silence — never touching an incoming clip's first word. The video-side crossfade durations were updated to match each new audio join exactly: in a sequential crossfade chain, video and audio overlap durations at a join must be numerically identical, or lip-sync drifts cumulatively across the sequence.

The reshoot added a synthesized ambient tension bed under the dialogue, per the team's own approved recommendation that a fully silent track was part of why the first cut felt inert. Once the trim fix above landed, the team reported the first ten seconds — under the silent Shot 1 — had no audible sound at all. Measurement showed why: a volume=0.10 multiplier stacked on an already-quiet filtered-noise source, plus a slow three-second fade-in, put the bed at roughly −61 to −65 dB — technically present in the file, well under audible threshold. Raising it to a measured, deliberately audible −32 dB (about 15 dB below the dialogue) and shortening the fade-in fixed that specific complaint.

What followed was more direct feedback that overturned the whole approach, not just its level: "There is a lot of noise in the video now. Cut it; it is annoying. We are wasting too much time. If you want to add background sound, use the erudita audio at a very low volume; the voice is important." The synthesized bed and sting were removed entirely and replaced with a track the team supplied directly, looped with short crossfades to cover the full runtime and mixed roughly 23 dB below the dialogue — audible as texture, never competing with the voice.

Confirmed by the team, in three words: "It lands."

The initial direction set black-and-white as non-negotiable, named 8½ as the visual reference, and asked for Marcello Mastroianni's presence in the framing and performance. A clarifying question established this meant style homage only — never a literal likeness of the real actor. Tone was set as a Hitchcock-style announcement that must make the player feel they are entering "an emotional danger zone."

A first script was written and produced as far as Shot 1 before HeyGen's credits ceiling interrupted Shot 2 — the incident and its cost are recorded in full in §07 Bug 1, including the lost prompt that had to be rebuilt from scratch.

Given three source documents — Values, Characters and Cultural Heritage, The Architecture of Suspense, and The Drama of the Game — and direct feedback: "the game is not only about a family problem or drama. It is about the Foundation using the lace makers; it is about the Foundation cooperating with the Estado Novo; it is about discovering the truth; it is about informing UNESCO, or not, about the truth found." The script was rebuilt around that correction, expanding from four shots (~55–60s) to six (~80–90s) to give the institutional stakes their own beats rather than compress them into an aside.

A specific wording question was raised: should the prologue name the object as "a lace," or stay vaguer with "handcraft heritage"? The recommendation was to keep the concrete noun — "handcraft heritage" reads as nomination-file language, not spoken dialogue, and naming the material doesn't reveal the crime's mechanism, it just makes the mystery tangible. A further recommendation followed from the source material's own opening line — that ERUDITA is itself the name of the bobbin lace at the center of the crime: describe the lace, but withhold the proper noun connecting it to the game's own title, so that discovery stays a later payoff.

Asked directly for an honest assessment rather than confirmation: the Guardian of the Archive's warning-heavy dialogue sat closer to a moral narrator than the "guardian of the archive … does not decide what the players should believe" the character notes describe; a line gesturing at the UNESCO decision ("not far from here") didn't hold up geographically; and the camera direction asked for one continuous push-in across three separate HeyGen generations — the exact failure mode Episode 2's own production diagnosed and had to fix after the fact with a cross-dissolve. Two of those were corrected before production: the geography line, and the camera continuity (changed to three independently-framed static shots). The Guardian of the Archive's tone and Shot 3's density were left as open notes rather than decided unilaterally.

The assembled first cut came back rejected on every axis — voice, lighting, shared title, and simply not holding attention. The verbatim rejection and the four questions asked before touching HeyGen again are recorded in full in §07 Bug 2.

All six shots were resubmitted with the revised environment paragraph and voice direction, avatar reselected before submission for the third documented time in this production (§02). The reshoot was approved.

"You made a point here. I think the video 2 small problems: first, there is a trim audio problem in the transition from the first to the second video, and eaten words in the first video. Secondly, and very important, the player of the game is a tourist with no connection to the families involved in the story until he enters the game. So the last part needs a different call to adventure."

The sequencing that followed was explicit: fix the video and the ending first, inspect every trim rigorously, and only then build this didactic unit — which is why this unit was produced last in the record, not first.

The closing line was rewritten to move the stakes from a family the player doesn't yet have any claim to, toward the three objects the player is actually about to be handed — the lace, the museum, the foundation — and reframed the choice as what the truth will demand of the player, not what it will cost a family. Shot 6 was regenerated against the new line and verified word-for-word by whisper against the intended text (§04).

Per the instruction to "inspect each trimming voice video," every one of the five joins was checked individually — word-level timestamps per clip, not a single pass over the finished mix — before any fix was applied. The root cause (a not-quite-silent Shot 1, and word onsets sitting at frame zero on four of five later clips) and the fix are recorded in full in §07 Bug 3.

A direct, single-line report, resolved by measurement rather than by ear — see §07 Bug 4 for the level reading that explained it and the fix that followed.

Direct feedback, with an audio file attached, overturned the synthesized-ambient approach itself, not just its level, and specified the replacement: the team's own supplied track, at very low volume, dialogue prioritized. Recorded in full in §07 Bug 4.

Three words closed the audio work. The instruction that followed authorized this document: build the didactic unit, using the whole conversation, with the title "The Guardian of the Archive" and a focus on the problems encountered producing the avatar scene, script and content — which is the brief this document has followed throughout.