Amir explicitly made the transcript product active ahead of Keep, delegated a stronger brand and asked for hosted ASR/LLM processing, Paddle and a Pageweft-like usage model. Sonaweft is the selected brand (bounded exact-name searches found no obvious matching software listing; this is not trademark clearance). Preserve instascript imports/command and existing remote URLs. Noteweld is the proposed next Keep brand; its implementation remains deferred.
Source decisions: CLI-001/002/003/005/009/010/013/022/023/033/037/041 and the current chat. Broad audiences and good defaults; English first; user-authorized media; Linux CLI/self-hosting plus hosted UI. Success: ingest media, hear/read timestamped ASR, keep original text, save separate corrections, optionally derive a faithful summary/research questions, export Markdown/JSON/SRT/VTT and a versioned bundle. AI review must not silently change the source or pretend to verify facts. A selected all-or-none hosted package is charged only after all its deliverables complete; retained private ASR checkpoints support explicit retry.
Release choices
- Host FastAPI and bounded workers on the existing owner Modal account, with separate Sonaweft volumes, secrets, ledger and Paddle product/prices. CPU/GPU scale to zero. Retained storage, scheduled cleanup, API calls and checks remain usage, not a guaranteed $0 bill.
- Speech: faster-whisper 1.2.1, pinned PyAV 15.1.0, SYSTRAN’s MIT large-v3 conversion on an on-demand GPU. Local CLI keeps CPU/int8 and can reuse existing small weights. Large weights are not installed on the laptop.
- Optional LLM: verified existing DeepSeek credential; official available IDs
deepseek-flashanddeepseek-v4-pro(GET models returned 200). Separate standard/advanced choice, explicit transcript-sharing consent, bounded input/output and no automatic paid retries. Provider credentials stay out of browser/storage exports. Groq remains a cheaper ASR candidate ($0.04/h turbo, $0.111/h full Whisper in current docs) but no Groq credential/account is present; no fictional integration is enabled. - Payments: one-time USD $5/$10 processing packs, no recurring access fee. Tariff: $0.15/job + $0.015 per started minute; optional standard review adds $0.10 + $0.005/minute, advanced adds $0.20 + $0.01/minute. This is a service price, not exact provider reimbursement or proven net profit. A 60-minute recording costs $1.05 transcription, $1.45 with standard review or $1.85 advanced before pack sales tax. Actual measured costs still need comparison.
- Bounded trial: first 30 seconds, transcription only, once per workspace/IP hash, at most three globally per UTC day and twenty per UTC month, additionally subject to admission/storage caps. Free use is never unlimited; no $0.50-per-IP guarantee.
- Uploaded source/result retention 48 hours; financial/recovery records persist separately. At most five retained uploads per workspace, 512 MiB source storage globally to start, two active jobs per workspace; initial file/duration limits are displayed by the deployed configuration. Owner estimated admission caps: $1/day and $2/month, including failed attempts. These do not cap the entire Modal/provider bill.
- Preserve immutable ASR; edits have revisions and cannot overwrite another session silently. The shared
instascript-export/v1schema carries stable segment IDs, timing, provenance, hashes and separately marked corrections/review. A failed import should later be atomic in Noteweld; its importer is not implemented by producing an export. - Browser Vim and Obsidian-style editing belong to Noteweld: CodeMirror Markdown/Vim, app navigation separately opt-in, wiki links/backlinks, tags/themes, bounded local graph, lexical search and portable vault export. No proprietary Obsidian runtime or arbitrary shell/Neovim process is embedded.
Evidence and limits
Baseline: 11 dependency-light tests passed. An actual local ASR probe found Nix library-path requirements and PyAV 19 incompatibility (metadata_errors removed); the pinned release corrects the latter. Subsequent evidence belongs in RELEASE.md, not in unsupported quality claims. No application can promise perfect speech recognition, universal platform URL access, verified truth or guaranteed income. Paid access remains gated by the real seller/domain approvals and isolated payment tests described in OPERATIONS.md.