| name | source-evidence-dataset-development |
| description | Orchestrate evidence-driven TianGong TIDAS data creation or update from PDFs, Word files, web pages, URLs, APIs, databases, reports, or scientific literature. |
Source Evidence Dataset Development
Use this top-level skill when there is no import-ready LCA package and the task must develop or update TIDAS rows from evidence.
Boundaries
- This skill coordinates seed intake, evidence retrieval, entity planning, and child skill routing.
- Do not treat search results as final field evidence. Capture field-level evidence, limitations, and conflicts first.
- Do not copy external research skills into this repository. Resolve fast-moving evidence skills with
npx skills.
- Do not write database rows. Remote write remains a later CLI/Foundry handoff after deterministic gates pass.
Required Inputs
- Foundry task directory under
.foundry/workspaces/<task-id>/.
- Execution seed or a clarification request if the seed is not executable.
- Target profile/account guard and intended use.
- Source starting points: files, URLs, reports, APIs, database ids, citations, or a bounded source-discovery request.
Seed Gate
Before retrieval, freeze a seed-manifest.json with:
- object/process/product name;
- target entity plan;
- functional intent or declared unknown;
- geography and time scope or declared unknown;
- source starting points or source-discovery scope;
- intended use and quality target;
- profile/account guard.
If source starting points and source-discovery scope are both absent, stop and ask for clarification.
Runtime Evidence Skills
For SCI paper or academic journal evidence, resolve the latest external skill:
npx skills use https://github.com/tiangong-ai/skills \
--skill tiangong-kb-sci-search \
--full-depth
tiangong-kb-sci-search searches only the sci channel. Keep report, patent, standard, official, web, and database evidence as separate channels.
Before using the skill, write a runtime resolution record in the Foundry workspace:
.foundry/workspaces/<task-id>/runtime-skills/runtime-skill-resolution.json
Include the command, source repo, resolved refs/heads/main commit, skill name, evidence channel, timestamp, and output artifact paths.
Workflow
- Freeze seed and source starting points.
- Build goal/scope and evidence questions by field.
- Retrieve and capture evidence by channel.
- Build an evidence dossier:
evidence/sources.jsonlevidence/chunks.jsonlevidence/field-evidence.jsonlevidence/conflicts.jsonl
- Fetch SDK-backed context packs for target TIDAS types.
- Author candidate support, flow, process, and lifecyclemodel rows only from captured evidence or explicit proxy/waiver records.
- Validate and QA rows through the CLI.
- Build and drive the same CLI curation queue used by packaged imports.
- Route semantic blockers to child skills such as
flow-governance-review, process-automated-builder, or Foundry authoring tasks.
- Verify queue scope, run dry-run/publish handoff, and require readback verification before completion.
Before any remote account read or write in steps 8-10, run tiangong-lca auth status --json; before commit, require tiangong-lca auth doctor-auth --json. A login-required result is a human handoff to auth login in a trusted terminal. Never request a username, password, credential, authorization code, token, or legacy API key. Keep a separate private session per account/project/client; headless tokens may come only from an approved orchestrator and never enter argv or artifacts.
Stop Rules
Stop when:
- seed is not executable;
- critical field evidence is absent or conflicting and no scoped proxy/waiver is allowed;
- runtime evidence skill resolution is missing for SCI retrieval;
- a row would require invented units, amounts, locations, years, sources, or classifications;
- schema/QA/curation/queue verify blocks with no runnable next task;
- remote write would bypass CLI dry-run, commit handoff, or readback verification.
Done Criteria
- Seed, source manifest, runtime skill resolution records, evidence dossier, entity plan, rows, schema/QA, curation queue, task checkpoints, mapping/provenance, dry-run, commit handoff, closeout, and readback verification are complete for the requested scope.
- Every authored field points to evidence, profile policy, remote reuse decision, or explicit waiver.
