PUBLIC-SOURCE RESEARCH SCAN · EVIDENCE CHECKED 5 NOVEMBER 2026
AI feature gaps: a 20-company research scan
This scan develops my research into where AI tools can better support evidence quality, user autonomy, correction, privacy and accountability. It compares existing capabilities with my earlier proposal families to identify useful contributions, rather than submit the same broad request repeatedly. A documented limitation can become an exploratory project, a reproducible test or a bottleneck analysis before it becomes a feature request.
Coverage: Twenty organisations selected from the existing outreach programme, including its AWS, IBM and NVIDIA records; this is not a market ranking. Five entries identify documented product boundaries, one records a checked public issue, and fourteen identify existing functionality with narrower extensions to investigate. These are evidence categories, not twenty proven missing features.
Method, scope and how to use the findings
Compared the existing evidence-verification proposal family and context-transparency requests with official documentation, feedback routes and selected public issue records. Public-source discovery was followed by direct checks of selected consequential limits and issue states. Sources below are the evidence trail for each entry.
This was a desk scan, not an exhaustive crawl, paid-account audit or live API experiment. Private support tickets, internal roadmaps and Discord histories were not inspected. An undocumented capability is not necessarily absent; a documented boundary can have a workaround. Issue states and product contracts are dated observations and may change.
Contribution sequence: identify the exact product and requirement; check existing issues and workarounds; reproduce with fictional or authorised data; separate observed behaviour from its explanation; write a bounded analysis; then refine the existing request or contribute to its current thread. Record submission, receipt, review and implementation separately.
No feature requests or messages were sent during this scan. Existing submission and moderation histories remain in their original entries; channel verification does not establish posting access, staff review or adoption.
1. OpenAI — ChatGPT / Codex / APIs
Documented product boundary
Existing capability: ChatGPT documents memory, memory review/correction and source visibility. Availability varies by plan and settings. The memory summary is not necessarily a complete view of all usable information, and memory does not retain every conversational detail.
Potential contribution: refine the existing context/evidence request into a retrieval-coverage and correction-impact view: what source was selected, what relevant material was unavailable or omitted, and which earlier conclusions may be affected. This is an extension to existing memory controls, not a claim that memory or source inspection is absent.
Research check: use fictional facts across chats, introduce a correction and compare source traceability, retention and downstream revisions. No account-specific performance was tested here.
Route: official support; the public community has a ChatGPT feature-request category. Codex repository issues are product-specific, not a universal ChatGPT request route. Retain the previous moderation history and use the appropriate category.
Sources: Memory documentation; support route; public request category.
Source code to explore: openai/openai-python. Official API client; does not expose ChatGPT memory internals. Licence / reuse terms: Apache-2.0.
My project — code & hand annotations →
Research dossier, visual roadmap & demo editor →
2. Anthropic — Claude API / Claude Code
Documented product boundary
Existing capability and documented limits: the citation API supplies text-source pointers. Its documentation says image citations are unsupported and native citations cannot be combined with strict structured outputs. A documented line-oriented extraction/parsing workaround exists; that is not the same as the requested native combination.
Potential contribution: schema-constrained extraction with per-field native evidence locations, plus visual-region citation support. These should be separate, narrowly specified requests and tested against the documented workaround.
Research check: compare a two-step extraction/citation pipeline with line-oriented extraction on fictional text, tables and scans; score field accuracy, evidence alignment and lost qualifiers separately.
Route: Claude support; Claude Code issues only for Code-specific matters. A separate privacy-feedback issue, #71486, is closed as a duplicate of #67719, according to its comments. This does not establish implementation. The current state of #67719 was not checked.
Sources: citations and compatibility; support; privacy-feedback issue; closure explanation.
Source code to explore: anthropics/anthropic-sdk-python. Official API client; native citation and schema compatibility remain service-side questions. Licence / reuse terms: Apache-2.0.
My project — code & hand annotations →
Research dossier, visual roadmap & demo editor →
3. Google — Gemini / Google AI
Existing capability / narrower extension to investigate
Existing capability: Gemini API grounding provides source-linked answers and search metadata. Grounding is not itself a historical archive or proof that every claim is supported.
Potential contribution: specify historical source snapshots with effective/publication/retrieval dates and separate subsequent-correction notices, building on the existing as-of request rather than reintroducing it.
Research check: ask the same bounded historical question against current pages and preserved versions; measure hindsight contamination and missing-version abstention.
Route: Google AI Developers Forum for Gemini API topics. Consumer Gemini feedback should use its own in-product channel.
Sources: grounding documentation; developer community; Grace's existing request.
Source code to explore: googleapis/python-genai. Official Gemini API client; not a historical web archive. Licence / reuse terms: Apache-2.0.
My project — code & hand annotations →
Research dossier, visual roadmap & demo editor →
4. Microsoft — Microsoft 365 Copilot
Existing capability / narrower extension to investigate
Existing capability: documented grounding, citations, web-query visibility and audit-related controls overlap the proposed transparency and evidence work. Tenant/admin controls differ from end-user views.
Potential contribution: examine accessible end-user notices when a cited document is superseded, removed or no longer authorised, and a clear route to challenge a consequential answer. Absence of a specific notice was not established in this desk scan.
Research check: in an authorised test tenant, change a fictional source and its permissions, then compare new answers, historical answers and available audit evidence.
Route: product feedback and the official developer feedback guidance; select the particular Copilot product.
Sources: grounding and limitations; web grounding and audit references; feedback guidance.
Source code to explore: microsoft/autogen. Agent framework for a local analogue; not Microsoft 365 Copilot implementation or tenant audit code. Licence / reuse terms: MIT for code; CC-BY-4.0 for documentation.
My project — code & hand annotations →
Research dossier, visual roadmap & demo editor →
5. Meta — Llama / Model API
Existing capability / narrower extension to investigate
Existing capability: model cards, tool-call formats and model-specific limitations are published. A model and a complete evidence-management application are different system boundaries.
Potential contribution: a portable evaluation pack for evidence attribution, uncertainty and tool-result provenance, with versioned model/application configuration. Treat this first as a contribution to evaluation rather than an assumed missing company-wide API.
Research check: compare a base model application with a provenance-aware wrapper on the same fictional records and resources.
Route: official Model API support for feedback; model repository issues only for the relevant model or implementation.
Sources: model documentation; official support and feedback.
Source code to explore: meta-llama/llama-models. Llama reference utilities and model-specific materials; custom terms vary by model. Not a complete hosted API. Licence / reuse terms: Model-specific custom Llama terms; inspect the selected model licence.
My project — code & hand annotations →
Research dossier, visual roadmap & demo editor →
6. Amazon — AWS / Bedrock
Existing capability / narrower extension to investigate
Existing capability: Knowledge Bases support metadata filtering and incremental synchronisation of added, modified and deleted source documents. These are important existing building blocks.
Potential contribution: reproducible historical retrieval and explicit correction dependencies across generated answers, beyond filtering a current index. Whether a managed product already meets a particular requirement needs service-specific testing.
Research check: ingest two source versions with distinct effective and publication dates; test historical answers, deletion behaviour and update visibility.
Route: AWS Builder Wishlist, verified by the official FAQ. Use the existing as-of submission as the conceptual anchor.
Sources: sync behaviour; retrieval configuration; official Wishlist guidance.
Source code to explore: aws-samples/amazon-bedrock-samples. Bedrock sample applications; the managed knowledge-base service remains external. Licence / reuse terms: MIT-0.
My project — code & hand annotations →
Research dossier, visual roadmap & demo editor →
7. NVIDIA — NeMo Guardrails
Existing capability / narrower extension to investigate
Existing capability: fact-checking, retrieval/output rails and evaluation tooling are documented, including checks against relevant chunks. Generic fact-checking is not an absent feature.
Potential contribution: evaluate whether controls detect outdated evidence, circular corroboration, unsupported implications and missing conditions, separately from source-consistency checks.
Research check: supply current and superseded fictional chunks; measure false acceptance and abstention. A fact-checking pass against supplied text does not establish that the source itself is true.
Route: developer forums are a real product-specific feedback surface. Grace's earlier moderation/access obstacle remains a separate unresolved matter; this scan does not establish restored posting access or justify reposting.
Sources: fact-checking; evaluation; developer forums.
Source code to explore: NVIDIA-NeMo/Guardrails. Guardrail toolkit; no guarantee of complete fact checking or source independence. Licence / reuse terms: Apache-2.0; inspect third-party notices.
My project — code & hand annotations →
Research dossier, visual roadmap & demo editor →
8. IBM — watsonx / API Connect
Existing capability / narrower extension to investigate
Existing capability: AI Factsheets and governance tooling already record tracked model/prompt metadata and lineage. A request for generic provenance would overlap existing products.
Potential contribution: specify claim-level challenge decisions and downstream correction propagation across systems, distinguishing these from model lifecycle lineage.
Research check: trace a fictional contested claim, adjudication and amended source through downstream consumers; test whether affected answers are identifiable.
Route: official IBM Ideas portals, selecting the correct product.
Sources: watsonx governance; Ideas portal guidance; data/AI ideas.
Source code to explore: IBM/watsonx-ai-samples. watsonx sample notebooks/apps; not the internal API Connect appeals implementation. Licence / reuse terms: Apache-2.0.
My project — code & hand annotations →
Research dossier, visual roadmap & demo editor →
9. Mistral — Studio / connectors / agents
Existing capability / narrower extension to investigate
Existing capability: MCP connectors, scope/registration and observability integrations already cover substantial parts of the earlier architecture proposal.
Potential contribution: add structured evidence-relation and correction-impact annotations to traces, rather than requesting MCP or observability again.
Research check: introduce a source revision into a traced workflow and determine which downstream claims can be located and reconsidered.
Route: Mistral support explicitly accepts feature requests; product-specific Discord/GitHub discussions have narrower scopes.
Sources: connectors; observability integrations; support.
Source code to explore: mistralai/client-python. Official API client; connector availability and traces depend on the selected service. Licence / reuse terms: Apache-2.0.
My project — code & hand annotations →
Research dossier, visual roadmap & demo editor →
10. Cohere — Command / RAG
Existing capability / narrower extension to investigate
Existing capability: fine-grained RAG citations already link output spans to document sources.
Potential contribution: evaluate unsupported subclaims, contradictory evidence and omitted conditions beyond a citation's presence. First determine whether the application layer can provide the required ledger.
Research check: construct compound claims whose sources support only selected components; measure citation attachment separately from evidential sufficiency.
Route: official support and developer Discord are documented. CrowdFeatures was not established as an official Cohere intake channel by this scan.
Sources: RAG citations; official support/community guidance.
Source code to explore: cohere-ai/cohere-python. Official API client; citation attachment alone does not prove evidential support. Licence / reuse terms: MIT.
My project — code & hand annotations →
Research dossier, visual roadmap & demo editor →
11. Perplexity — Search / Sonar / Agent APIs
Existing capability / narrower extension to investigate
Existing capability: search date-range filtering is already announced. This differs from recovering the actual historical version and verification state of a record.
Potential contribution: refine the existing eight-family proposal set into independently testable requirements for source ancestry, versioned evidence and correction propagation. No evidence here establishes that the complete architecture is implemented.
Research check: compare a publication-date filter against archived versions of revised pages; score version fidelity and hindsight contamination.
Route: official API forum explicitly invites endpoint, metadata and documentation proposals. It is not a general consumer-app request channel.
Sources: date-filter announcement; request-category scope; existing Grace proposal.
Source code to explore: perplexityai/perplexity-py. Official API client; not the proprietary search index or research engine. Licence / reuse terms: Apache-2.0.
My project — code & hand annotations →
Research dossier, visual roadmap & demo editor →
12. xAI — Grok / tools
Existing capability / narrower extension to investigate
Existing capability: web/X search and source citations are documented, including collection-file references. Citation lists and evidence verification must not be equated.
Potential contribution: test source independence, historical validity and the relationship between a cited source and each asserted claim, with clearly recorded uncertainty.
Research check: compare independent sources with paraphrased copies sharing one ancestor; assess claim support and provenance rather than agreement count.
Route: official contact/support; no public company-wide feature board was verified here.
Sources: citation documentation; collections search; official contact.
Source code to explore: xai-org/xai-sdk-python. Official API client; does not expose Grok training data or source ancestry automatically. Licence / reuse terms: Apache-2.0.
My project — code & hand annotations →
Research dossier, visual roadmap & demo editor →
13. Hugging Face — Hub / libraries
Existing capability / narrower extension to investigate
Existing capability: model/dataset cards, base-model metadata and repository discussions already support important provenance and contribution functions.
Potential contribution: interoperable claim-level support/dispute/unknown metadata and links to correction events, beyond descriptive cards.
Research check: build a small dataset schema or library-level adapter and assess whether it complements existing Hub metadata without introducing misleading confidence scores.
Route: Hub discussions for repository-specific contributions; relevant library issue trackers for implementation changes; the existing public forum proposal remains the conceptual anchor.
Sources: model cards; repository discussion scope; existing Grace proposal.
Source code to explore: huggingface/huggingface_hub. Hub client source; claim-level metadata is an extension to investigate. Licence / reuse terms: Apache-2.0.
My project — code & hand annotations →
Research dossier, visual roadmap & demo editor →
14. DeepSeek — JSON Output
Documented product boundary
Documented boundary: the JSON-mode guide states that unrecognised JSON can be replaced with spaces, leaving apparently empty content even when the finish reason is stop. It advises prompt adjustment and client-side blank handling/retry.
Potential contribution: an explicit machine-readable generation/format failure state and reliable recovery contract, rather than a superficially successful blank response. Check alternate endpoints and schema modes before asserting the limitation applies to all DeepSeek interfaces.
Research check: use a bounded fictional extraction set and classify valid outputs, blanks, truncations and recovery results; record endpoint, model and configuration.
Route: official FAQ exposes feedback. A searchable public company-wide request board was not established.
Sources: JSON-mode notice; official FAQ.
Source code to explore: deepseek-ai/DeepSeek-V3. Reference inference code, not the hosted JSON-mode implementation; model weights have separate terms. Licence / reuse terms: MIT for code; model weights have separate terms.
My project — code & hand annotations →
Research dossier, visual roadmap & demo editor →
15. Alibaba — Qwen / Qwen Code
Public issue report — not independently reproduced
Existing capability: function calling and Qwen-Agent parsers are documented, including warnings about stopword-based templates for reasoning models.
Directly checked report: Qwen Code issue #16213 is open and reports conditional-rule state shared between parent and subagent contexts, suppressing a reminder in the other context. Its author explicitly seeks a maintainer decision on intended ownership. This scan did not reproduce the report or establish the design's intended semantics.
Potential contribution: clarify and test per-context versus shared rule-consumption state; make delegated safety/operating constraints inspectable.
Research check: reproduce the issue's minimal two-file example in the specified version before proposing a fix or new feature.
Route: the relevant Qwen Code or Qwen-Agent repository; avoid an obsolete general Qwen repository for unrelated products.
Sources: function-calling documentation; reported issue #16213; current issue list.
Source code to explore: QwenLM/qwen-code. Coding-agent source; parent/subagent rule ownership can be inspected here. Licence / reuse terms: Apache-2.0.
My project — code & hand annotations →
Research dossier, visual roadmap & demo editor →
16. Moonshot — Kimi / Kimi Code
Existing capability / narrower extension to investigate
Existing capability: session export/import, compaction and user guidance on preserving important context are documented. Generic context compression or session export is not an absent capability.
Potential contribution: assess whether compaction preserves conditions, corrections and evidence identifiers, and whether a preservation report would help users detect loss. Its absence was not established.
Research check: compare original and compacted fictional histories under fixed budgets, then inject a correction and test dependent conclusions.
Route: official in-product feedback and the current Kimi Code issue tracker. Check repository migrations before filing against archived CLI code.
Sources: sessions and context; contact and feedback.
Source code to explore: MoonshotAI/kimi-code. Coding-agent source, including session/context machinery; not all Kimi service internals. Licence / reuse terms: MIT.
My project — code & hand annotations →
Research dossier, visual roadmap & demo editor →
17. Groq — Structured Outputs
Documented product boundary
Documented boundary: structured outputs already exist with strict and best-effort modes on selected models. The current page states that streaming and tool use are not supported with Structured Outputs. Tool-use streaming elsewhere is a different combination.
Potential contribution: a scoped compatibility extension or capability-discovery matrix covering strict schemas, tools, streams and fallback semantics.
Research check: confirm the precise supported combinations on the selected model/endpoint before evaluating correctness and latency.
Route: the official Groq community and documented community channels.
Sources: structured-output scope; community; developer community.
Source code to explore: groq/groq-python. Official API client; structured-output/tool/stream compatibility is server and model dependent. Licence / reuse terms: Apache-2.0.
My project — code & hand annotations →
Research dossier, visual roadmap & demo editor →
18. You.com — Research / Answer APIs
Existing capability / narrower extension to investigate
Existing capability: the Research API documents cited natural-language and structured outputs; product documentation also describes supporting source material. Do not request generic research with citations or structured JSON as if absent.
Potential contribution: independently test compound-claim support, source ancestry and historical-version fidelity; document any specific mismatch before proposing additional metadata.
Research check: create revised-source and duplicated-source cases and separately score citation attachment, source support and historical validity. Vendor verification claims remain claims until tested.
Route: official API support/Discord; integration requests have a documented support route. No public company-wide feature backlog was verified.
Sources: Research API; support; integration request guidance.
Source code to explore: youdotcom-oss/youdotcom-python-sdk. Official API client; cited research results require independent source checking. Licence / reuse terms: MIT.
My project — code & hand annotations →
Research dossier, visual roadmap & demo editor →
19. Fireworks AI — structured generation / inference
Existing capability / narrower extension to investigate
Existing capability: structured-output and grammar/schema guidance already exists. Schema compliance does not establish factual support, but the latter need not be an inference-provider feature.
Potential contribution: a provider-neutral semantic/evidence validation adapter and explicit model-capability contract, distinguishing vendor behaviour from third-party wrapper translation errors.
Research check: compare direct API output with a wrapper on identical fictional schemas; score syntactic validity, schema adherence and evidence sufficiency separately.
Route: official contact guidance directs product support, bugs and account issues to Discord and documentation. A public company-wide feature board was not verified.
Sources: structured outputs; official contact routing.
Source code to explore: fw-ai-external/cookbook. Official cookbook evaluation examples; not the hosted inference implementation. Licence / reuse terms: Apache-2.0.
My project — code & hand annotations →
Research dossier, visual roadmap & demo editor →
20. Glean — enterprise retrieval / citations / connectors
Documented product boundary
Documented boundaries: citation pointers can remain in historical answers after access is revoked, while opening the source remains permission-controlled. Connector updates can lag. S3 documentation describes different scanned-PDF/OCR behaviour in crawled content and direct chat uploads, with support-mediated configuration options.
Potential contribution: connector-specific extraction/freshness notices and privacy-preserving stale-citation status, so users can distinguish missing extraction, missing permissions, stale indexing and genuinely absent evidence.
Research check: in an authorised test workspace, use the same fictional scan across upload and connector paths, then change a source/version/permission. Measure extraction coverage and status visibility. Retained citation pointers alone do not establish unauthorised access.
Route: official Glean community has Feedback and Ideas; support access may depend on designated contacts. This is enterprise Glean at glean.com, not the separate note-taking brand formerly using glean.co.
Sources: citations; S3 connector boundary; crawl/update FAQ; official community; support contacts.
Source code to explore: gleanwork/api-client-python. Official API client; enterprise connectors and permissions require an authorised test tenant. Licence / reuse terms: MIT.
My project — code & hand annotations →
Research dossier, visual roadmap & demo editor →
Authorship: Grace R. Nettey supplied the research priorities and existing contribution programme; AI-assisted source review and synthesis supported this scan. The candidate projects and requests have not been independently validated. Dated publication record →