[Product Slice] Make Deep Research complete one primary-source-backed, claim-verified report end to end

[helsome]

Problem

Deep Research has accumulated several strong subsystems — retrieval ranking, checkpoint/recovery, Gold Cases, Langfuse tracing, prompt-injection defenses, claim/evidence work, claim verifier work, and primary-source connectors — but the repository still lacks one explicit product acceptance target that proves these pieces work together as a complete research workflow.

Without that vertical slice, individual modules can be correct while the final report still loses source authority, evidence spans, verification status, or reproducibility between retrieval and synthesis.

Related: #13 #15 #20 #21 #30 #32 #39 #100

Current state

Folio currently has or is actively developing:

• Deep Research production workflow;
• source ranking/deduplication;
• primary-source document adapters (for example SEC / issuer sources);
• structured claim/evidence relationships;
• claim verifier core;
• inline citation/source inspection;
• Gold Case datasets and Langfuse trace integration;
• checkpoint and crash recovery;
• prompt-injection defenses for untrusted research content.

The missing piece is one canonical report flow that demonstrates the whole chain rather than another isolated subsystem.

Expected direction

Choose one real public company and one bounded research question, for example:

“What drove the company’s latest-quarter revenue/margin change, and what are the major risks to the current thesis?”

The production flow should be:

Research question
  ↓
Research plan
  ↓
Primary-source retrieval (filing / exchange / IR)
  +
Secondary-source retrieval (news/web when useful)
  ↓
Dedup / source ranking
  ↓
Evidence extraction with stable source + location
  ↓
Atomic claims
  ↓
Claim-level verification
  ↓
Synthesis
  ↓
Persisted report
  ↓
Citation / Source Inspector
  ↓
Eval / trace metadata

Primary sources should be preferred for factual financial claims; secondary sources may add interpretation or market context. Unsupported conclusions must remain visibly unsupported rather than being promoted into confident report prose.

Non-goals

• No requirement to cover every market or filing provider.
• No requirement to implement a new research planner if the current planner is sufficient.
• No requirement for a large multi-company benchmark.
• No broad UI redesign.
• Do not duplicate [Deep Research] Add claim–evidence verification and bidirectional citation traceability #13/[Finance Research] Add first-class filings, announcements, and research-report source pipelines #32/[Research Retrieval] Deduplicate search results and rank sources by authority, freshness, and diversity #39 implementations; this Issue is the integration and product-acceptance target.

Validation / acceptance criteria

• Run one real Deep Research task through the production research path using a real model.
• Retrieve at least one real primary source (regulatory filing / exchange disclosure / issuer IR document) and at least one additional independent source when relevant.
• Retrieval output demonstrates dedup/source-ranking behavior; syndicated copies do not count as independent supporting evidence.
• The final report contains stable claim IDs for factual/causal conclusions and evidence references that resolve to specific source spans/structured facts, not only source homepages.
• Each material factual/causal claim is machine-readably classified as supported / contradicted / insufficient evidence (or equivalent).
• Unsupported/contradicted claims are not silently rendered as verified conclusions.
• Report persistence + reload preserves claim/evidence/citation identities.
• At least one inline citation can be opened from the report into the corresponding evidence/source detail.
• Run metadata records model/provider, retrieval policy/version, relevant tools, and trace/eval identity where available.
• The run is represented in the current Gold Case / evaluation path, or produces an equivalent reproducible eval artifact with groundedness/evidence checks.
• Provide a reproducible test report with exact commands, Bun/OS, model/provider, query, source set, verifier summary and focused/full test results.
• Attach screenshots of the final report and at least one claim → citation/evidence inspection path.

Maintainer note

The goal is not architectural completeness. A narrow, credible, end-to-end research report is more valuable than adding another broad abstraction. Prefer wiring existing components together and fixing contract gaps discovered along the path.

[RXQ6]

/claim

我来接手这个 Issue。

按 CONTRIBUTING 里「大 Issue 可以拆成多个可独立合入的小 PR」的约定，我打算分阶段推进，不一次性交付全部 acceptance criteria：

第 1 阶段（可离线验证）：把「检索 → 去重/排序 → 证据抽取（稳定 source + 位置）→ 原子 claim → claim 级验证 → 合成 → 持久化 → reload → 引用解析」这条链在现有的 fixture / Gold Case 设施上先跑通并固化成测试，重点断言两件事：claim / evidence / citation 的身份在 persist + reload 之后保持稳定；未被验证支持的结论不会以已确认的口吻呈现。

第 2 阶段（真实运行）：补真实模型与至少一个真实 primary source（filing / exchange / IR）的运行，产出可复现的 test report 与截图。

实现上会尽量复用 #104 / #105 / #106 正在推进的统一证据契约与 report evidence bundle，不另起一套证据模型。

如果维护者希望调整切入顺序或 scope 边界，请直接指出，我按结论调整。

[github-actions]

@RXQ6 已认领该 Issue，并已自动设为 Assignee。提交 PR 时请在描述中写 Closes #102。

[RXQ6]

我在认领时说明过会分阶段推进，这里根据最新在飞的 PR 做一次 scope 收敛。

核查了当前所有在飞的 PR 后，我调整了切入方式，避免重复实现：

不新建证据 / claim 契约模型。 按 #67 上的评审结论，等 #104 的统一 EvidenceBundle 落地后在其上做接线 —— 我不会另立 ResearchClaim 或第二套 claim id 语义。

直接复用现有组件，不重复实现：

本 Issue 需要的子能力	复用谁
检索去重 + 来源排序	#39 / PR #58
claim 三态判定（supported / contradicted / insufficient）	#13 / PR #90
带稳定位置的文档证据	#32 / PR #95
报告证据 bundle 持久化与重载	#100 / PR #105
报告内行内引用 → 证据详情通路	#30（Part 1 已由 #88 合入）

本 Issue 我负责的 scope 收敛为：

把上述组件在生产路径上串成一次完整的、可复现的 Deep Research 运行 —— 真实模型 + 至少一个真实 primary source，产出带稳定 claim / evidence / citation 身份的持久化报告，并交付本 Issue 要求的可复现 test report 与截图。不做新的抽象层，只做接线和补齐契约缺口。

前置未就绪期间不阻塞等待： 先准备可复现的验收脚本（scripts/ 下，纯新增），等前置 PR 合并后填入真实运行结果。

如果维护者希望调整切入顺序，或认为某一部分已有归属应当让出，请直接指出，我按结论调整。