Proposed protocol foundation for Astra v2 capability conquest. Last updated: 2026-05-27
This matrix turns the v2 conquest plan into a decision-complete tracking sheet. It stays protocol-first: a capability is not considered conquered until it is measurable in deterministic bench, live, holdout, and proof lanes.
Month 6 release note: this file is a capability-progress view, not a release-claim override. If a row is stronger than the current evidence bundle or support matrix, the release-facing wording must be downgraded to match the proof. For reader/file claims, “Strong” means controlled proof lanes for PDF, EPUB, and .srt/.vtt subtitle-file flows; it does not turn parser conveniences such as ASS/Markdown/TXT/HTML, OCR, Docx, comic, or image translation into public support claims.
Owned reading + subtitle chain (not a separate row here): As of 2026-05-27 the repo has a unified owned-reading storage model and vocabulary Reading queue across article / PDF / EPUB / subtitle-file shapes, with honest limits on universal reopen (month-3-closeout-inputs-2026-04-14.md, owned-reading-schema-v1-2026-04-14.md). Required live evidence for that line is now pnpm bench:live:lane:document-proof, covering bench-live/document-intake-basic (PDF + EPUB + VTT intake plus unsupported DOCX boundary), bench-live/document-intake-local-file-handoff, bench-live/pdf-reader-basic, bench-live/epub-reader-basic, and bench-live/subtitle-file-basic; bench-live/learning-loop-revisit-smoke remains the Reading queue revisit smoke. The subtitle-reader → explain/save → vocab/review path for files is written in subtitle-reader-learning-chain-2026-04-14.md; in-page caption adapters are scoped separately (video-subtitle-adapter-inventory-2026-04-15.md, support-matrix-video-addendum-2026-04-15.md) and YouTube proof is now guarded by pnpm bench:live:lane:youtube-proof.
平台支持等级与对外 claim 边界不由本文件定义，统一以
docs/investigations/support-matrix-2026-q2.md为 canonical 来源。
This file tracks the user-facing capability rows defined in bench-opt/capabilities.ts. Cross-cutting platform work such as managed-service routing, AI quality-control, and rules-system UX matters a lot, but it is tracked as a supporting execution stream rather than a separate capability row here.
- Strong — production-quality base already exists in-repo
- Partial — some runtime or harness support exists, but parity is incomplete
- Gap — capability is planned but not yet implemented or not yet represented in the harness
- green — lane already exists and is trusted
- partial — lane exists but is not yet parity-grade
- missing — lane does not exist yet
coverage— feature exists but is under-testedruntime— product/runtime work is still missingux— product polish or interaction quality gapprotocol— harness/proof/reporting gapnew-subsystem— new reader/ingest/OCR/class-specific subsystem required
| Capability | Astra current implementation status | Current harness coverage (bench, live, holdout, proof) |
Read Frog reference value | Immersive benchmark value | Gap classification | Conquest wave | Exit criteria |
|---|---|---|---|---|---|---|---|
| 网页翻译 | Strong — deterministic bench + live page translation already exist; article extraction and dynamic content are already in the bench surface inventory, and dedicated churn/layout-noise holdouts are green | bench=green, live=green, holdout=green, proof=green |
Bilingual / translation-only reading UX, context-aware reading flow, article-mode orientation | Core product surface with polished bilingual readability and dynamic page handling | coverage, ux |
B | Article/feed/docs/dynamic page variants all green; dedicated page-translation holdout stress green; proof-suite content-reading prompts show no web-translation regressions |
| PDF 文件翻译 | Strong — PDF reader first cut now has deterministic bench, live reader flow, holdout layout-noise coverage, and capability-proof contribution | bench=green, live=green, holdout=green, proof=green |
Reading-oriented UX patterns are useful, but PDF is not the primary inspected Read Frog differentiator | Explicit product surface for layout-preserving PDF translation | new-subsystem, runtime, protocol |
B | Bilingual + translation-only rendering stable; PDF extraction/order/alignment deterministic lanes green; PDF live reader lane green; multi-column and delayed-page holdouts green; proof-suite confirms document/PDF contribution |
| 视频双语字幕（YouTube 双语字幕） | Strong — the benchmarked YouTube bilingual subtitle path and learning-workspace proof lane are green end-to-end; runtime platform adapters now extend beyond YouTube, but broader platform coverage is not yet a separately conquered claim | bench=green, live=green, holdout=green, proof=green |
Read Frog explicitly highlights subtitle translation in the player flow | Explicit core video subtitle experience | runtime, coverage, ux |
B | bench:live:lane:youtube-proof covers player button, bilingual subtitles, seek recovery, track switch, transcript panel, transcript search/jump, save sentence Review handoff, and video-note create; broader video platform support remains scoped by the support matrix |
| Epub 电子书双语翻译 | Strong — EPUB reader first cut now has deterministic bench, live reader navigation, long-chapter holdout coverage, and capability-proof contribution | bench=green, live=green, holdout=green, proof=green |
Reading-oriented, bilingual study workflows are directly relevant | Public document/reader translation surface | new-subsystem, runtime, protocol |
C | Chapter extraction/order deterministic lanes green; live chapter navigation and reading-mode switching green; malformed/long-chapter holdouts green; proof-suite confirms reader-state contribution |
| 图片翻译 (Beta) | Gap | bench=missing, live=missing, holdout=missing, proof=missing |
No primary image-translation reference surfaced in inspected Read Frog materials | Official beta image translation surface | new-subsystem, runtime |
D | OCR/overlay beta benchmark lane exists; live static-image fixtures green; noisy-background and mixed-script holdouts meet beta threshold; failure categories are explicit |
| 漫画翻译 (Beta) | Gap | bench=missing, live=missing, holdout=missing, proof=missing |
No primary comic translation reference surfaced in inspected Read Frog materials | Official beta comic translation surface | new-subsystem, runtime, ux |
D | Panel/balloon fixture lanes exist; live comic page baseline green; overlapping-balloon and stylized-font holdouts reach beta-grade pass; protocol documents beta-only thresholds |
| 隐私模式 (Beta → core policy) | Partial — privacy-mode request-context sanitization exists for page translation and subtitle translation, popup study context suppresses richer fields in privacy mode, sensitive-input suppression exists for input translation, and the background transport boundary is authoritative for translation request context. Targeted live/holdout scenario paths exist for page-translation privacy, but current release evidence is not yet system-wide enough to market this as a strong privacy guarantee | bench=partial, live=partial, holdout=partial, proof=partial |
Context-aware AI reading UX implies strong privacy expectations | Product positioning strongly implies privacy guarantees | protocol, coverage, runtime |
D (standalone conquest) | Deterministic privacy assertions cover all major request surfaces; fresh live privacy replay artifacts are attached when a release wants to strengthen privacy wording; release docs can state the exact privacy boundary without over-claiming local-only or end-to-end secrecy |
| 鼠标悬停翻译 | Strong — deterministic hover support, live hover translation, moving-target holdouts, and proof contribution are green | bench=green, live=green, holdout=green, proof=green |
Inline / selection-adjacent UX is a strong inspiration source | Explicit hover-translation surface | coverage, ux |
C | Live hover lane green; moving-target and overlay-interference holdouts green; no conflicts with selection/input/float-ball flows |
| 字幕文件翻译 | Strong — subtitle-file ingest/preview/export now has deterministic, live, holdout, and proof coverage across .srt/.vtt flows; ASS/Markdown/TXT/HTML parsing remains opportunistic convenience support |
bench=green, live=green, holdout=green, proof=green |
Adjacent to subtitle concepts, but file ingest/export is a separate workflow | Public document/subtitle-file translation surface | new-subsystem, runtime, protocol |
C | .srt/.vtt parser/serializer deterministic lanes green; upload/preview live flow green; malformed timing and overlapping-cue holdouts green; parse/translate/export failures are operator-visible; public claims name SRT/VTT unless separate proof is added |
| 输入框翻译 | Strong — deterministic bench + live input translation are stable, and the field-matrix holdout verifies contenteditable, delayed hydration, repeated edits, password suppression, and cursor preservation | bench=green, live=green, holdout=green, proof=green |
Inline authoring assist and selection-adjacent UX offer strong interaction inspiration | Explicit input-box translation surface | coverage, ux |
B | Textarea/contenteditable/cursor-preservation deterministic lanes green; multi-field live flow green; delayed hydration and repeated-edit holdouts green; reporting distinguishes trigger/translate/writeback failures |
These tracks are real product work, but they are not separate capability rows in bench-opt/capabilities.ts yet. They support multiple rows at once and should be treated as execution streams that can block or de-risk conquest.
Status on 2026-05-27: first managed-service hardening slice validated.
What is already true:
- ordinary product copy should present Astra-managed service modes rather than provider/model controls
- service fallback/retry metadata is implemented for operator diagnostics
- service-mode telemetry aggregates latency/failure/fallback by user-facing mode
- deterministic + targeted live evidence cover active-session service/site restart behavior
- fallback / restart / transport behavior are benchmark-visible in scoped slices rather than purely anecdotal
What is still missing:
- broader backend roster behind managed service modes
- per-surface preferred service policy beyond the first scheduler
- broader multi-event operator history / failure history beyond the current popup-local support path
Still open.
Main remaining gaps:
- glossary / terminology contract is now canonical at request-time, but still lacks stronger proof/guarantee language beyond the current runtime/tests
- stronger context shaping controls
- bad-output detection beyond current placeholder protections
- benchmark-visible rejection of obviously bad model output
- broader user-facing disclosure beyond the current popup-local support path when transport falls back from direct to relay
Runtime support is materially better than before, but UX remains behind a productized rules system.
Remaining gaps:
- clearer advanced-rules affordances
- rule validation feedback with examples
- diagnostics for matched / excluded / untranslated regions
- Freeze this matrix.
- Keep
bench-opt/capabilities.tsand this document aligned. - Ensure every row has target behavior, harness-lane gap statement, and explicit exit criteria.
- 网页翻译
- PDF 文件翻译
- YouTube 双语字幕
- 输入框翻译
- EPUB 双语翻译
- 字幕文件翻译
- 鼠标悬停翻译
- 图片翻译 (Beta)
- 漫画翻译 (Beta)
- 隐私模式 (Beta → core policy)
说明：这里的 Wave D 指的是把 privacy 作为单独 capability row 完整征服的时点；privacy gating 本身需要更早并行推进，并且可以阻断 Wave B/C 的扩面 claim。
A capability is only marked conquered when all four are true:
- deterministic bench green
- live standard lane green
- hidden/holdout lane green
- proof-suite family/tasks show no regressions caused by the capability
The capability must also be represented in:
- status / reporting / operator output
- benchmark pack protocol
- a visible rubric and a hidden robustness lane
Cross-cutting platform tracks do not need to become separate rows before shipping progress, but if they regress badly enough to invalidate a capability claim, the capability is not credibly conquered either.