October30 current correction supersedes the unsplit30s experiment recorded below. New WAN source edits cap15s and request the source-based whole-second duration; fractional delivery normalization is a Plotthirst-funded repair cost without another customer fee. Optional source-edit prompt planning is a separate approved OpenAI text operation, in bounded groups of ten with one repair maximum. Its frozen quote covers token bounds plus one compute/request/buffer/monthly allowance (site-wide divisor>=1000); known usage settles once, unused reservation is released, and video generation still requires separate review.100 reviewed prompts/25min and500000 input characters are supported; large text storage does not add a per-clip planner charge. Unknown usage and unpriced normalization/storage/transfer overages remain platform reconciliation exceptions. These editor changes do not price the120min Seedance scenario here or establish full margin coverage. All six existing stale source fingerprints remain deliberately unreset; no paid validation.

Scenario: 120 minutes, 720 ten-second clips, one take each, standard Seedance 2.0 reference-to-video, 720p 16:9, image references (no reference video), native audio. Regular customer multiplier 1.8, 100 credits per dollar, no discounts or retakes. This is credit consumption value, not necessarily cash paid after credit purchase bonuses.

This is NOT a verified all-in checkout quote. Missing from this subtotal: video prompt planning, reference analysis, any new references beyond these 18, full delivery packaging/original-media storage and download/stream transfer allowances, and the monthly shared operating allowance. Separate music, narration, re-generations and extra takes also change the scope.

The story and final-review rows are estimates. Image rows use the current local image helper and assume one 1536x1024 production image per sheet/storyboard and one 1024x1024 portrait per sheet. No image/video provider requests were made. Public sample viewing costs the visitor nothing.

Operating allocation gap

The original standalone estimate import returned no populated billing section. Subsequent Admin browser verification on September 9 shows a loaded sandbox allowance of $140.0154826191 and 1,000 estimated actions. Neither observation verifies production settings; $250/month below remains an illustrative scenario. With $250 / max(1000, estimated actions), the initial shared cost is $0.25 per eligible billable action. At 1.8x it adds $0.45 before any AWS estimate buffer. Do not multiply that across arbitrary model calls without checking which actions the actual ledger bills.

The preview helper explicitly applies this share plus a 20% AWS estimate buffer. In that $250/1000 scenario, the 720 preview quotes alone become $403.20 instead of $14.40 (+$388.80). The video/image quote helpers used above do not include that same share. This inconsistency must be reconciled before claiming one final customer total; do not double debit at parent and child.

Evidence

- Local validation: local-runtime/sample-price-validation-2026-09-09.json

- Pricing helpers: server.py; plotthirst_clip_preview_runtime.py

- Standard Seedance route: provider-routing-catalog.json; $0.0112 / 1,000 output video tokens, 216,000 tokens per ten seconds at 720p / 24fps = $2.4192 provider, 436 customer credits after rounding.

- Official model page: https://www.atlascloud.ai/th/models/bytedance/seedance-2.0/reference-to-video

- Fresh screenplay primary measured cost $6.9527046, excluding historical recoveries; final-review allowance $0.80–$1.80 provider. Customer estimate uses approximately $7.75–$8.75 x 1.8.

Release status

Corrected sample is packaged for sandbox review; not deployed. The bounded final-review engine exists, but mandatory automatic invocation and customer quote/debit wiring are still pending. User policy is automatic final review and necessary bounded corrections, not a customer toggle. This sample uses the eight already approved editorial fixes, without another paid review.

2026-09-10 cost review: the only server change since the reviewed hash is a local-enabled, authenticated admin sample compiler preflight endpoint (plus line endings). It makes no provider calls, reservations or debits. Existing movie pricing/workflow is unchanged. Local admin inspection is an unbilled PC-only exception; paid planner/adaptation and media integration remain unimplemented. Smoke tests prohibit provider network calls.

2026-09-10 sample production handoff review: sample imports now validate the 21 published artifacts by existing integrity checks, rather than requiring unpublished intermediate writer reviews; completed artifact import is not a claim that those reviews ran. The saved 720 screenplay assignments and existing full reference sheets are reused without generation, model calls, reservations or debits. Per-segment readiness no longer requires unrelated project references. No price/model/markup changes. Existing paid segment generation helpers are unchanged; their end-to-end worker/billing acceptance remains pending, including the previously recorded clip-preview billing inventory gap. The sample preview settings do not transfer to the production copy; review saved project settings before submission.

2026-09-10 video-default persistence cost review: authenticated account-scoped sample video settings read/write and existing production-copy field updates are ordinary unbilled project configuration, with storage/request overhead under existing settings operations. No provider calls, quote, reservation or debit added. Video model/shape/resolution changes affect future work via existing fields; existing take state is retained. Pricing/catalog/markup are unchanged; generation still needs its normal current price validation.

2026-09-10 follow-up: video preferences use a small account-scoped storage record rather than the whole app-state document. Project-state writes now use unique temporary files and serialize final replacement, fixing the reproduced concurrent-save collision. These are storage reliability/configuration changes only; no provider, pricing, reservation or debit changes. The dedicated preference module is now fingerprinted for future cost-monitor review.

Sample merge cost note (2026-09-11): local explicit publication copies existing completed media; no provider regeneration or second generation debit. No movie-generation price change. Hosted publication/storage/transfer adapter and its cost coverage are not implemented by this local merge; not included as a newly priced operation.

### 2026-09-11: AI playlist planning

Implemented in the local build: player prompt, whole-segment duration control, full-screenplay selection, saved personal playlists, free quote and explicit maximum-credit confirmation. One GPT-4.1-mini call; no video generation. Settlement uses actual returned tokens plus one operating allowance, account multiplier/discount and configured buffer. Local own-key mode excludes provider wholesale from site credits. Duplicate submissions cannot rerun the provider; failed invalid output releases the hold. Uncertain settlement is retained for reconciliation, not automatically retried. Seven focused billing/selection tests pass; frontend typecheck/build passes. No paid provider validation or website deployment for this feature yet. Existing unrelated billing inventory/cost-monitor gaps remain open.

2026-09-11 playlist production (local): Player now offers a one-minute trailer preset, reviewable selected segment list and Generate this playlist with a separate production confirmation and inline job progress. Canonical sample segment IDs now map into player clips; generation maps exact whole ranges into the existing sample production project and rejects incompatible layouts. Existing planner/storyboard/video pipeline and saved project defaults are reused, completed media preserved, stopAfterClips enabled; no automatic stitch or extra planning debit. AI playlist fee is separate from existing video production charges. Existing production overhead/quote gaps remain: this change does not certify an all-in video quote. TypeScript and local bundle build pass. No paid trailer/clip run or deployment performed; end-to-end paid acceptance remains pending.

2026-09-11 Terra playlist default: user approved switching selector to gpt-5.6-terra. Trailer preset now invites slight reveals from later scenes that raise questions without explaining outcomes. Quotes include model identity; execution honors saved quote model. Terra uses Responses medium reasoning directly with no fallback duplicate call, a reasoning output allowance, normalized usage and reported cache-write/cache-read pricing. Existing playlists preserved. Local UI rebuilt and idle server restarted; paid run of revised preset not performed. Existing unrelated cost monitor and billing inventory gaps remain.

2026-09-11 playlist editorial directions v3: requires early/middle/late comparison without coverage quotas, explicit intended playback order, one grounded reason per selected ID and concise category-level exclusions (including reasons for excluding broad movie portions). New results validate reasons and retain response order; player displays reasons/exclusions. Old saved playlists remain readable. Original source context still drives generation. Eight focused tests pass. Additional response text uses the existing single quoted call, with typical output allowance increased; no new review/debit/stage and no paid run. Existing broader cost-audit gaps remain.

September 14 Seedance 2.5 option audit

Resolution-dependent estimates now replace the old flat 0.134 USD/sec approximation. See SEEDANCE_25_OPTIONS_2026-09-14.md and shared/seedance25-options.json. This does not re-audit the existing Seedance 2.0 full-movie baseline or resolve previous operating-cost gaps.

Seedance reference input preparation — local, September 14

- Shared Atlas Seedance 2.5 reference path inspects input videos before provider submission. Below 409,600 pixels it prepares a separate H.264 copy with a small safety margin and even dimensions. Content is fitted without cropping; tiny padding handles frame rounding. Audio is copied, frame rate is not overridden, duration/audio/dimensions are verified. Valid-size inputs are not re-encoded. Unsupported dimensions fail explicitly; no duration truncation or automatic paid retry.

- Verified supplied 854x364 MOV -> automatic 992x422 copy, 29.374s, original SHA unchanged. Immediate phone-download copy is 992x424, H.264/AAC, 24fps, 29.374s.

- 35 unit/contract tests pass plus real-file conversion. Runtime requires ffmpeg and ffprobe in the executing worker; confirm that in the eventual hosted release. Lambda package guard includes the helper. No deployment or paid generation this turn. Temporary download is local sandbox only and must not be added to git or production samples.

- Cost review: no new AI/provider operation or customer debit. Local conversion is owner-run. For future hosted enablement, input fetch, temporary storage, CPU and upload requests are currently an explicitly unpriced Plotthirst-absorbed compatibility cost, including own-key users. Repeated attempts can repeat preparation; persistent caching and measured cost allocation remain follow-up items. Existing operating-cost audit gaps remain open.

WAN 3.0/Prime resolution-specific rate review added locally 2026-09-14; see docs/WAN3_QUICK_SOLO_2026-09-14.md. Existing whole-movie baseline remains flagged pending wider review; these provider subtotals are not all-in customer prices.

2026-09-14 local framing correction

Production image canvases follow the project ratio within the old non-square pixel budget; custom OpenAI sizes use the existing non-square estimate bucket. No extra AI preparation call or automatic paid retry added. Actual custom-size billing and existing overhead gaps require review; no baseline hashes refreshed. See PRODUCTION_FRAMING_2026-09-14.md.

September 15 input repair cost review

Quick Create now passes Seedance 2.0 reference-video duration to the token-rate calculator for both quotes and settlement. Grok 1.5 1080p T2V/I2V uses the official $0.25/output-second tier; preset voices are free inputs, and reference mode stops above 720p. Existing full-movie baseline and operating allowances remain pending wider review; baseline hashes remain unchanged. No new planning or paid retry stage was added.

2026-09-17 Suno menu scope

Fenix and Auk are disabled by owner choice; the Fenix integration is archived for possible future testing. This menu/archive change introduces no provider calls or new charges. Historical archived prices are not current quotes. The broader source-cost review remains stale and its baseline hashes have not been refreshed.

September 22 selected-segment dependency review

Automatic selected-segment production can now create missing reference subjects before storyboards. This adds the existing reference child to this entry point, not a new rate or second parent debit. The target filter runs before sheet/portrait pricing, credit balance validation and reservation. A regression test quotes one selected missing sheet plus one portrait at 5 credits each as 10 credits; unrelated and already-ready references are excluded. Existing reference child settlement uses its saved reservation and target prices. Worker sequencing tests verify references finish before storyboards and clips. Existing media and source files are preserved.

The old 18-reference full-movie subtotal is not a bound for all screenplay subjects; page 6 has additional props. Do not present that subtotal as a checkout estimate. The same six stale source fingerprints are also reported in the existing release checkout before this change. They remain flagged: this focused review does not silently certify the old all-in estimate or change its baseline hashes. Broad operating allowance and own-key gaps remain as documented.

2026-09-24 local customer estimate display

The new generation review displays current child-helper credit forecasts before

screenplay queueing, and server estimates before Ultimate submission. This is a

visibility/consent change, not a price-sheet change or reconciliation of this

report's unresolved cost components. The earlier subtotal must still not be

presented as a verified all-in movie cost. Extend/Loop explicitly identifies its

initial reservation as a subtotal. Baseline hashes intentionally remain stale

until a full pricing review; no paid generation or deployed acceptance claimed.

### 2026-09-25 — Retire Sora video generation (local)

Removed Sora 2 / Sora 2 Pro from shared video and routing catalogs, legacy menus and TypeScript fallback options at owner request. Historical clip/spec data remains. Saved retired selections fail with a choose-another-model error before provider dispatch and new cost quotes. No replacement provider is silently selected and no new rates or billing operations were added. Live deployment remains pending after the prior ECS alarm rollback; full-cost audit gaps and existing stale baseline remain open.

October 3: WAN split editing is a separate optional Quick Solo operation, not part of this Seedance 120-minute movie subtotal. Its reviewed reservation, one prep/assembly fee, discounted output settlement and platform-funded overhead exception are documented in SPLIT_VIDEO_EDIT.md and BACKEND_COST_AUDIT_POLICY.md. Existing movie baseline source re-audit gaps remain; no stale fingerprints were refreshed.

October4: optional WAN split continuity adds local extraction of each previous

edited final frame and reuses it as the next input. The extra input count and

bounded frame preparation join the existing approved edit/service units, with

private media/storage/transfer coverage exceptions documented in

BACKEND_COST_AUDIT_POLICY.md. This does not change the120-minute Seedance movie

subtotal or certify its unpriced components. No baseline hashes refreshed.

### October 23 audio-reference input review

Quick Solo/Multi audio input coverage now follows shared model-specific contracts. This is existing video-call input, not another generation stage: no provider rate, markup, discount, reservation or debit rule changed. Multi quote/cache/hash uses each clip’s exact named audio subset; split retry identity includes audio and preserves completed children. Existing upload/storage/audio normalization is reused; global operating-cost gaps remain open rather than being declared covered. H3 input file count does not add wholesale generation cost per the primary Atlas reference guide. No customer prices/settings changed or paid jobs run. Static billing inventory remains unchanged. Six existing source fingerprint mismatches remain, without baseline reset. See VIDEO_AUDIO_REFERENCES_2026-10-23.md for scope, contract sources and acceptance limits.

October24 input-name repair: mixed References lists resolve to exact typed subsets before existing per-clip pricing. No new paid operation, audio transformation or storage write; existing upload receipts are reused. No rates/discounts/reservations/debits changed. Mock12-clip quote sums every10s clip once; no real provider or customer credit mutation. Billing inventory unchanged; existing six stale monitor fingerprints remain.

October 24 — Video reference scope (local; publication held)

Quick Multi now passes each clip's selected, owned video subset and measured input duration through its existing quote/cache/child settlement path. Seedance 2.0 input-video tier and input-plus-output rates therefore apply to that clip's actual inputs. No new parent debit, rate/discount change or paid retry was added; completed reuse remains. Web, saved setup, ZIP and MCP handoff share this path. Existing Solo own-key handling is unchanged. Upload verification, catalog/preview and compatibility-preparation costs reuse current Plotthirst-funded maintenance; their unresolved overhead gaps remain explicit. No paid generation, financial change or production enablement occurred. See QUICK_MULTI_VIDEO_REFERENCES_LOCAL_2026-10-24.md for manual review and 231-check/browser evidence. Existing six stale cost-monitor source fingerprints remain unreviewed; do not refresh baseline hashes to silence them.

October26 local Multi source-edit repair: reuses Solo split-editor per-portion15s reserve/measured settlement, selected input seconds, discounts and one combined preparation/assembly service debit. No ordinary Multi stitch charge or second parent generation debit. Existing own-key Solo unchanged; Multi source editing rejects own-key requests. Temporary H264/PCM MOV chunks avoid AAC duration padding but may increase bytes; existing service/storage/shared-overhead recovery remains an explicit unverified allowance. Upload duration/catalog repair adds no paid generation.266 Python checks/Chrome synthetic verification, no paid/customer changes; see QUICK_MULTI_SOURCE_EDIT_2026-10-26.md. Source publication pending, previous source5e74fde; six stale source fingerprints remain unresolved and unreset.

October26 multiple-source follow-up: initial editor release was b194226. New versioned plans select up to ten originals with independent ranges/references and source-specific contiguous continuity. Provider units retain maximum15s reserve/frozen measured settlement. ONE shared service allowance uses distinct USED original seconds plus selected output seconds and mapped frame work; no unused-source read/charge or extra parent/stitch debit. Originals cached, completed chunks reused. Website/mobile/API workers share the path; dedicated MCP source-edit schema is not newly exposed. Source-edit own-key still rejects; Solo/unversioned approvals remain compatible. No paid planner, rates/balances/discounts/settings changes. Existing storage/transfer/shared-overhead gaps remain; no profit claim.282 Python checks,11 frontend reports, typecheck/build and synthetic Chrome pass. See QUICK_MULTI_SOURCE_EDIT_2026-10-26.md. Same six stale source reviews unreset; source push and hosted acceptance remain distinct.

October26 settlement follow-up: unchanged frozen source-edit15s units/discounts/proration and one service debit. Missing approved units now stop before provider work. Only verified picture<=15s with container/audio tail<=0.12s may settle by picture duration; retain original measurement and never clamp longer picture. Plotthirst absorbs any minor supplier rounding/container basis difference; actual wholesale billing remains unverified. Fresh Auto uses its existing bounded saved-file download plus one local ffprobe; legacy padded checkpoints may need one additional bounded owned-file read. No duplicate provider/parent/stitch charge, changed rate/ceiling, own-key change or paid retry. Ordinary Auto measurement unchanged. Versioned source timing checks run before new quotes, without rewriting accepted plans. Explicit admin diagnostic uses one exact owned worker record/checkpoint under the existing platform-funded support exception. Static inventory unchanged. Broader storage/transfer/shared overhead and the six existing stale movie-cost fingerprints remain unresolved/unreset. Actual failed media and recovery require live verification.

October26 frozen source-edit continuation review: read-only preview checks exact owned job, closed prior hold, approved per-portion/service unit totals and an owned saved output. Confirmed reserve equals original accepted ceiling minus cumulative ledger spend, never a new current-price quote. Deterministic hold ID plus queue-generation CAS bounds repeated/concurrent requests; completed media and stable debit IDs are reused. Original one service allowance and per-unit discounts/rounding remain; no duplicate parent/stitch/title fee or new paid retry for the retained clip. Source own-key rejects. Support preview/checkpoint reads are Plotthirst-funded; no claimed fix for broader AWS/storage/shared overhead recovery. Unknown queue acceptance retains the hold pending reconciliation and never releases it under possibly running work. One reviewed static reserve path: plotthirst_batch_edit_resume.resume. Actual saved first clip is verified15.000s picture/15.023s container and settles360credits in pure calculation; no live debit/requeue yet. Six stale movie-cost fingerprints remain unreset.

October 28 — Atlas media models and runtime list pricing

The new Atlas media choices do not change this report’s selected Seedance sample scenario. Each newly selected model is priced by its own reviewed list rate or Atlas origin_price calculator; this document is not an estimate for a Kling/PixVerse movie. Existing six stale source fingerprints and all prior full-cost/margin acceptance gaps remain open. Do not reset the baseline hashes to clear them.

October30 WAN unsplit Auto: a new versioned30s reserve/frozen per-variation unit allows whole Solo sources totaling at most30s; final debit prorates measured duration from the approved unit, with stable debit event and existing unused-hold release. No split/service/parent fee is added. Standard/Prime output per-second list rates, multipliers, account/own-key rules and model set remain. Existing split15s and older approvals are preserved. Existing Auto saved-file read/probe is reused; verified30s picture with <=0.12s container tail uses picture basis, with the same platform-funded minor-tail exception as existing split editing. Static billing inventory unchanged after review; broader storage/transfer/shared overhead coverage remains unresolved.99 focused Python checks incl real-media split/continuity/padding, TS/build and static audit pass. Same six movie-monitor fingerprints remain stale/unreset. No customer/paid generation test performed; publication pending.