Add Patent Evidence: source-checked agent continuity and evidence promotion

[alexanderipc]

What this PR does

Adds Patent Evidence, an agent research workspace and 15-tool MCP integration for investigations that outlive a single chat, agent or product specification. It stores cited findings in DKG Working and Shared Working Memory so another agent can recover, check and extend them, alongside a persistent local notebook of challenges, source changes and unfinished questions.

The central result: a clean installation of the published package, given only original sources and project identity, recovered 25 findings from DKG and checked all 82 citations. No research database or prior conversation was transferred. A separately registered reader used the exact registry installation recipe; cross-project access was rejected. Post-publication verification.

This PR adds only integrations/patent-evidence.json. Application code stays in the contributor-owned repository. Round 1 tag: cfi-dkgv10-r1.

Start with the demonstration

Watch the narrated walkthrough · Read the design brief · Inspect the execution evidence

The demonstration follows Fable Fern Labs and its WickRail self-watering planter, both fictional, against three real public patent publications. It shows an investigation continuing across agents, a changed engineering specification, reasoned reassessment, and a signed evidence candidate for an explicitly fictional opposition. The walkthrough uses actual browser captures and synthetic narration; the evidence archive includes real agent transcripts and machine-readable results.

Why this matters for LLM-Wiki and autoresearch

A patent investigation combines exact claim wording, evolving technical specifications, contrary evidence and unresolved assumptions. A fluent summary can lose the source version or qualification that made an earlier conclusion defensible. The practical user is a researcher or practitioner who needs an agent to resume that work without silently inheriting unsupported conclusions.

Patent Evidence implements a reusable source-checked handoff. The receiving workspace has originals but zero findings. SWM supplies the missing research records; the receiver independently extracts the originals and accepts a record only when its project, quotation, locator and source/passage fingerprints match. Imported human-review labels are not trusted. The local database remains useful for indexing and UI, but cannot account for the recovered research in this experiment because the receiving database initially contains none.

This makes DKG part of the continuation mechanism. It also offers a pattern other document-heavy research integrations can reuse: preserve source identity, recover attributed reasoning, retain challenges, and explicitly reassess dependencies when a source changes. Source-only recovery reconstructs cited findings; it does not reconstruct the full local notebook or treat remote review history as locally approved.

Delivered capability	Demonstrated result	Practical value
Agent-to-agent continuation	A real external Codex MCP agent, under a separate registered identity, recovered 16 findings, inspected originals, added a three-citation finding, shared it and read it back.	An arriving agent can contribute without receiving the previous chat or research ledger.
Repeatable memory recovery	Three subsequent fresh-workspace trials each recovered 17 findings and verified 61 citations. The later public-package check recovered the final 25 findings and 82 citations.	The handoff is reproducible; historical experiment counts are preserved rather than retroactively inflated.
Research that responds to change	A new fictional specification flagged ten dependent findings. Another real agent recorded two substantive reassessments; eight remained explicitly unresolved. Explicit dependencies propagate change flags transitively.	Changed assumptions become visible without overwriting the original reasoning or inventing approval.
Bounded autoresearch	A live continuation run selected an open question, retrieved previous DKG memory and added seven findings with follow-up questions. Completion alone does not close a question; a cited answer and reason are required.	Research can progress across sessions with a visible account of unfinished work.
Evidence maturation	A selected opposition candidate was signed, independently verified and consumed by an offline oracle answering with cited findings and unresolved checks.	WM/SWM outputs already have a concrete structure for later review and promotion.

The immediate adoption path is an existing document-review workflow: PDF, Word, Excel, text and patent HTML in a local research desk, or the same capability exposed to an MCP client. Codex execution is demonstrated; configurations for the registry's other named MCP clients are supplied without claiming they were all executed. The next adoption milestone is an independent practitioner reproduction under the predeclared pilot plan. No external user base or production deployment is claimed.

Faithful v10 integration

One project maps to one Context Graph; each immutable finding version maps to a Knowledge Asset; findings, citations and sources are Entities with provenance links. Public authenticated HTTP performs WM writes, read-back, finalization and explicitly requested, Curator-authorized SHARE. The public CLI handles local setup. There are no internal DKG imports, daemon patches or direct SPARQL writes.

Authority is tested with a negative/positive control: an unrelated identity cannot SHARE a private test asset, WM remains intact and no SWM record appears; the authorized Curator then shares the same asset successfully. Retry behavior accounts for SHARE removing the corresponding WM content. Research reviews are attributed conversations, not endorsement buttons or claims of cryptographic consensus.

The recorded deployment is a real DKG 10.0.18 development node with a mock chain and no configured peers or relays. The results establish durable same-node collaboration, not remote-peer replication. No funded wallet or TRAC spend is needed for this demonstration.

A concrete path to Verifiable Memory and context oracles

This release implements candidate preparation, signed export, verification and an offline oracle consumer. A selected package binds exact finding versions, source fingerprints, quotations and selectors, WM/SWM lineage, dependency issues and the described proceeding. The consumer verifies the signatures and source evidence before returning an answer.

The design separates four records that must not be collapsed: source content, evidence of historical public availability, claim-specific legal assessment, and the scope of a proceeding or decision. For example, “publication X contains passage Z” can be checked independently of the interpretation that Z discloses feature Y. A date quoted from a source remains an attributed assertion; anchoring today would not prove historical availability.

A future authorized Curator PUBLISH would follow disclosure review, qualified assessment and checks of historical-date evidence. An infringement action or opposition could nominate selected evidence for that review; filing a proceeding never authorizes disclosure automatically. No VM write or deployed network oracle is claimed: current publication status is false and on-chain UAL, anchor date and trust tier are null. Implemented contract and promotion design.

Integration links and reproducibility

• Repository: alexanderipc/patent-evidence-dkg
• Exact source commit: fb14eb35e347d2a45c9975e46ae21f4e3ae4e2bc
• Published package: patent-evidence-dkg 0.2.0 on PyPI
• Public release and checksums: v0.2.0
• Build and publication: successful protected GitHub Actions run
• PyPI provenance: wheel attestation · source-distribution attestation. Both verified with pypi-attestations; downloaded hashes match the audited build, whose application/source files match the pinned Git tree.
• Test coverage: TESTING.md. The exact snapshot passed 53 local tests including real-node integration tests. Linux and Windows CI each passed 50 tests with the three node-dependent tests skipped.
• Security and dependencies: SECURITY.md · THIRD-PARTY.md.

The exact registry launch recipe, with the four declared environment variables configured, is:

uvx --from patent-evidence-dkg==0.2.0 patent-evidence mcp

For a short reproduction: start the isolated node using the source README; bootstrap and share the fictional project; export its source-only kit; import that kit into a fresh workspace; register a separate reader; then call get_research_context, recover_shared_research and get_passage over MCP. The evidence archive contains independent-agent/events.jsonl, benchmark.json, source-reassessment/events.jsonl, the selected candidate and oracle verification outputs. The post-publication report records the clean-cache public install separately.

Registry checker compatibility: Section 8a permits an equivalent verifiable registry. This is a Python package published through PyPI Trusted Publishing with attestations. The current npm-oriented checker interprets the uvx --from argument as an npm package and cannot verify PyPI; its missing-npm-package warning is not a successful provenance check. The linked PyPI attestations, successful publishing run and public-install evidence supply the reviewable Python equivalent. Structural/schema validation passes.

Scope and security declarations

• Uses only supported public HTTP, CLI and MCP interfaces; no internal imports, node patches or direct SPARQL writes.
• memoryLayers declares WM and SWM; v10PrimitivesUsed reflects the exercised primitives and terminology.
• External hosts and all DKG write operations are declared; Curator-authorized SHARE is explicit. No PUBLISH is implemented.
• Local agent-scoped DKG credentials are documented. No third-party credential is requested by the installer; the optional model client uses its existing sign-in without this application reading its credentials.
• No application preinstall/install/postinstall scripts. No remote-input evaluation or dynamic remote module execution.
• Published with Trusted Publishing and verified attestations; the package was built from the exact pinned commit and public downloads match the audited build.
• Public package installation, 15-tool MCP discovery, source recovery, public-patent retrieval and project-scope denial verified after publication.

Tests include tampering, quotation integrity, project boundaries, source-change dependencies, archive refusal, forged review labels and candidate/oracle verification. All released research uses invented specifications and public patents. The app checks evidence integrity, not legal correctness, exhaustive FTO coverage or patent validity.

The prototype is single-user and loopback-oriented, without encryption at rest. The pinned upstream DKG runtime has 22 high and 3 moderate affected dependency entries, including propagated entries, triaged in SECURITY.md; these are not 25 distinct root vulnerabilities. Production hardening and remote-peer evaluation remain work. No legal-accuracy, token-saving or independent-adoption rate is inferred from the demonstration.

Contributor attestation and maintenance

• This integration is my own work or properly licensed. Code license: AGPL-3.0-only, matching the registry and package. Public publications retain their own legal status.
• It contains no intentional backdoors, malicious logic or data-exfiltration paths beyond the declared egress.
• I understand that material misrepresentation may result in delisting.
• @alexanderipc commits to at least six months of maintenance after registry acceptance.

AI assistance was used in implementation, testing and demonstration research. The pinned documentation was frozen before release; its pre-release publication-status notes are superseded by the live package, release and this submission. Application support is an ongoing commitment; bounty acceptance and an award remain the committee's decision.