Disciplines · Audits

Eve SOTA multimodal capability matrix — 2026-09

a claim that arbitrary source trees, binaries, or code execution are available.

9sections8 minread

On this page

Task 17.1 candidate. This is a source-derived admission boundary, not a feature wishlist. The ledger checkbox remains open pending exact-candidate independent review.

Record digest: 8e4ee88f7c08c4aed9ffbbda080869dd6733d1277b47bc8339071acad53789da

Modality Input Output Authoritative parser/tool Model leg Explicitly unsupported
Text and code supported supported assistant message schema and turn stream turn (bound), escalation (bound) arbitrary repository ingestion, sandboxed code execution, automatic licence clearance
Web conditional supported read_page, inspect_page turn (bound), vision (operator-configured) arbitrary web browsing, web search, deep research, webpage output or mutation
PDF and document files unsupported unsupported none none PDF upload, office-document upload, OCR, layout-preserving extraction, document export
Tables and spreadsheets unsupported unsupported none none CSV/XLSX upload, formula evaluation, cell-addressed citations, workbook output
Images and screenshots conditional unsupported inspect_page vision (operator-configured), image-generation (not-admitted) arbitrary image upload, multi-image analysis, certified OCR, image output
Audio and speech conditional conditional voice recording normalization and transcription route, Web Speech recognition and synthesis speech-to-text (operator-configured), text-to-speech (operator-configured), music-audio-generation (not-admitted) general audio-file upload, speaker diarization, music generation, audio-file output, live duplex voice
Video unsupported unsupported none video-generation (not-admitted) video upload, video understanding, video editing, video generation, captioned video output
DCC and project files unsupported unsupported none three-d-generation (not-admitted) DCC/project upload, live DCC control, editable project output, 3D model generation
Generated media unsupported unsupported none image-generation (not-admitted), video-generation (not-admitted), music-audio-generation (not-admitted), three-d-generation (not-admitted) image generation, video generation, music generation, 3D generation, generated-media editing or delivery

Text and code#

  • Input: supported — Eve accepts message text and code as text. This is not a claim that arbitrary source trees, binaries, or code execution are available.
  • Output: supported — The assistant emits text. A code fence is presentation, not an executed or independently validated program.
  • Formats: UTF-8 chat text, plain-text code snippets, streamed UTF-8 text, Markdown/code-fence text
  • Limits: Only route-admitted request sizes and the turn registry context floor apply; no repository upload is implied. Generated code is unexecuted unless a separately admitted tool executes and verifies it.
  • Privacy/licence: Text leaves the service only through the admitted regional model route and its production data-posture gate. Model output carries no automatic third-party licence warranty; members must review code and attribution obligations.
  • Accessibility: Text and semantic Markdown remain available to assistive technology.
  • Evidence: apps/oshun/bff/src/assistant/model-registry.ts, apps/oshun/bff/src/routes/assistant.ts, docs/audits/eve-sota-model-leg-contract/2026-09-05.json
  • Unsupported owner/unblock: 17.7, 17.24, 17.25 — Admit a bounded tool or parser and retain execution, privacy, licence, and failure evidence.

Web#

  • Input: conditional — Only the member's current product page is available when the client advertises read_page or inspect_page. Arbitrary URL fetch, browsing, and deep research are absent.
  • Output: supported — Eve returns text, not a saved webpage, DOM mutation, or browser automation result.
  • Formats: current-page accessible DOM summary, current redacted viewport PNG, text grounded in returned current-page evidence
  • Limits: inspect_page is limited to one current-viewport capture per turn. Private sections, form values, chat, and embedded media are hidden; offscreen visual content is not claimed. No arbitrary URL retrieval, crawling, search index, or deep-research tool is admitted.
  • Privacy/licence: The client redacts the current viewport before vision submission; page descriptions are untrusted evidence, never instructions. Eve does not grant reuse rights for page content.
  • Accessibility: read_page uses roles, accessible names, states, headings, and anchors. inspect_page converts visible evidence to text but does not certify complete alt text.
  • Evidence: apps/oshun/bff/src/assistant/agent-tools.ts, apps/oshun/bff/src/routes/assistant.ts, apps/oshun/bff/src/assistant/model-registry.ts
  • Unsupported owner/unblock: 17.26 — Admit a governed external-web tool with provenance, injection defenses, privacy review, and eval evidence.

PDF and document files#

  • Input: unsupported — Eve chat has no admitted attachment route or authoritative PDF/document parser.
  • Output: unsupported — Eve chat does not emit downloadable PDF, office-document, or editable project artifacts.
  • Formats: none
  • Limits: Pasted plain text is text-code input and loses file, page, layout, and anchor semantics.
  • Privacy/licence: No file is accepted, so no document-retention or third-party document-licence claim is made.
  • Accessibility: No OCR, reading order, page anchors, or accessible document export is available.
  • Evidence: apps/oshun/bff/src/routes/assistant.ts, EVE_SOTA_GAP_CLOSURE_TODOS_2026-09-01.md
  • Unsupported owner/unblock: 17.2, 17.24 — Ship deterministic attachment parsers, stable anchors, adversarial tests, citations, and governed file retention.

Tables and spreadsheets#

  • Input: unsupported — Eve chat has no admitted CSV/XLSX upload, workbook parser, or cell-addressed analysis tool.
  • Output: unsupported — Markdown-looking text is not a spreadsheet or verified computed workbook.
  • Formats: none
  • Limits: Pasted rows are treated as ordinary text without sheet, cell, formula, type, or provenance guarantees.
  • Privacy/licence: No spreadsheet file is accepted or retained by Eve chat.
  • Accessibility: No accessible workbook export, header mapping, or cell-level navigation is available.
  • Evidence: apps/oshun/bff/src/routes/assistant.ts, EVE_SOTA_GAP_CLOSURE_TODOS_2026-09-01.md
  • Unsupported owner/unblock: 17.2, 17.25 — Admit deterministic table parsers and a governed compute path with cell-level evidence and accessible output.

Images and screenshots#

  • Input: conditional — inspect_page is the only image input. Eve chat has no arbitrary image upload, gallery import, OCR endpoint, or multi-image conversation input.
  • Output: unsupported — The admitted route returns a text description; image generation and image-file output are not admitted.
  • Formats: one client-produced redacted current-viewport PNG
  • Limits: One current viewport per turn; capture must be current and pass client redaction/validation. The vision registry caps output at 1,000 tokens and returns plain text. The retained image-generation reference output is measurement evidence, not product admission.
  • Privacy/licence: Private sections, form values, chat, and embedded media are excluded before submission. Vision output does not grant rights in depicted content; image generation lacks Eve-chat posture approval.
  • Accessibility: A textual description can assist access but is untrusted, may be incomplete, and is not certified alt text.
  • Evidence: apps/oshun/bff/src/assistant/agent-tools.ts, apps/oshun/bff/src/assistant/model-registry.ts, apps/oshun/bff/src/assistant/generation-model-registry.ts
  • Unsupported owner/unblock: 17.3, 17.22, 17.23, 17.24 — Admit attachment and/or generation tools with privacy, provenance, accessibility, lifecycle, cost, and eval evidence.

Audio and speech#

  • Input: conditional — The production server route admits only its configured Deepgram path; when absent, the product may use browser Web Speech before review-and-send. This is not general audio understanding.
  • Output: conditional — Production server TTS is fail-closed; browser speech synthesis is the current fallback and no audio file is returned.
  • Formats: browser-recorded audio normalized by the voice route, browser speech-synthesis playback
  • Limits: Recorded request body is capped at 10 MiB. Server synthesis text is capped at 1,500 characters even though production server TTS is currently blocked. Voice input is asynchronous and review-before-send; no live full-duplex conversation is claimed. The OpenRouter speech reference routes failed the required privacy posture and are unavailable.
  • Privacy/licence: Server speech is fail-closed unless the provider/subprocessor posture is admitted. Browser speech behavior is controlled by the member browser/platform; no music or voice-cloning licence is granted.
  • Accessibility: Transcription produces editable text before sending; typed input remains available. Speech playback supplements rather than replaces visible text.
  • Evidence: apps/oshun/bff/src/routes/assistant.ts, apps/oshun/bff/src/assistant/model-registry.ts, docs/audits/eve-sota-voice-depth-decision/2026-09-12.json, docs/audits/eve-sota-provider-subprocessor-review/2026-09-15.json
  • Unsupported owner/unblock: 17.3, 17.22, 17.23 — Admit exact provider routes and modality tools with posture, quality, latency, cost, caption, and failure evidence.

Video#

  • Input: unsupported — Eve chat has no video upload, temporal parser, frame sampler, or timestamp-grounded understanding tool.
  • Output: unsupported — A measured provider reference clip exists, but the generation leg is not admitted to Eve chat.
  • Formats: none
  • Limits: No duration, codec, resolution, temporal-grounding, caption, or output guarantee is offered.
  • Privacy/licence: No video file is accepted. The reference generation route lacks Eve-chat data-posture and tenant approval.
  • Accessibility: No timestamped transcript, captions, audio description, or accessible player output is available.
  • Evidence: apps/oshun/bff/src/assistant/generation-model-registry.ts, docs/audits/eve-sota-generation-model-legs/2026-09-16.probe.json
  • Unsupported owner/unblock: 17.3, 17.22, 17.23 — Admit temporal input and/or generation toolchains with provenance, rights, budget, cancellation, captions, and eval evidence.

DCC and project files#

  • Input: unsupported — Eve chat has no project-file attachment or admitted Blender, Unreal, Unity, Houdini, Maya, or Adobe runtime lease.
  • Output: unsupported — No editable DCC/project artifact is returned from Eve chat.
  • Formats: none
  • Limits: No supported file types, versions, plug-ins, host applications, or round-trip fidelity are claimed.
  • Privacy/licence: No project file is accepted; no asset, plug-in, font, or source-media licence clearance is claimed.
  • Accessibility: No accessible project representation or alternative output is available.
  • Evidence: apps/oshun/bff/src/assistant/generation-model-registry.ts, docs/audits/BELLONA_DCC_HOST_MATRIX_2026-09.md, EVE_SOTA_GAP_CLOSURE_TODOS_2026-09-01.md
  • Unsupported owner/unblock: 5.5, 5.12, 17.4, 17.22, 17.27 — Admit a versioned leased host/toolchain and prove editable artifact fidelity, rights, privacy, cancellation, and rollback.

Generated media#

  • Input: unsupported — No media reference-upload surface is admitted to Eve chat.
  • Output: unsupported — Image, video, music/audio, and 3D registry entries are measurements or candidates only; every generation leg is not-admitted.
  • Formats: none
  • Limits: No chat formats, variant count, delivery SLA, cancellation, retention, or cost guarantee is offered. Reference image/video outputs and billed-but-unverified audio do not constitute product admission. No verified 3D output exists for the configured candidate.
  • Privacy/licence: Generation legs lack Eve-chat data-posture/tenant approval and cannot inherit approval from unrelated media routes. Rights, provenance, policy, and licence review must precede publication; no generated-media ownership guarantee is offered.
  • Accessibility: No generated-media delivery path exists, so alt text, captions, transcripts, and accessible controls are not claimed.
  • Evidence: apps/oshun/bff/src/assistant/generation-model-registry.ts, docs/audits/eve-sota-generation-model-legs/2026-09-16.probe.json, docs/audits/eve-sota-media-project-governance/2026-09-15.json
  • Unsupported owner/unblock: 15.7, 17.4, 17.22, 17.23, 17.27 — Move an exact leg through data posture, lifecycle, budget, provenance, rights, accessibility, delivery, and eval admission.