grounded-answer

Grounded Answer

Use this skill when the task is to answer a question from the published DocMason knowledge base rather than only retrieve evidence.

This is the inner specialist answer workflow behind the user-facing ask entry surface. Ordinary users should not need to name this workflow explicitly before asking business questions.

Front-Door Precondition

• grounded-answer is not a free-standing ordinary front door.
• Start only from canonical ask turn metadata and canonical ask runtime ownership.
• Start only after canonical ask has already handed the live turn here with status = execute and inner_workflow_id = grounded-answer.
• If the current turn is missing that ask-owned handoff, stop and route back to ask.

Required Capabilities

• local file access
• shell or command execution
• ability to inspect structured JSON output
• ability to inspect rendered images when the cited evidence requires visual confirmation

If the agent cannot inspect required rendered evidence, stop and explain that the environment is not capable enough for grounded answering.

Procedure

1. Start from canonical ask turn metadata instead of assuming every direct answer is purely KB-grounded.
   • honor the ask-owned work_brief, resolved decisions, accepted scopes, and affected-output boundary; use them to preserve the intended medium and critical distinctions without reopening decisions this workflow does not own
   • treat answer_state as the top-level four-state answer contract
   • choose an explicit support_basis for the overall answer:
     • kb-grounded
     • external-source-verified
     • model-knowledge
     • mixed
   • honor ask-provided reference_resolution, source_scope_policy, evidence_requirements, and support_contract; when evidence_requirements is present, treat it as the canonical odd-question inspection contract before widening scope or changing evidence basis
   • write a short support ledger before drafting:
     • which source boundary must survive
     • which comparison sources must both survive
     • which published evidence channels are required
     • whether this turn still has the one allowed contract-repair chance
2. Normalize the user question and decompose it only when that improves grounded retrieval.
3. Run docmason retrieve "<query>" --json --compact for the initial question or sub-question when KB evidence is part of the answer path.
   • if you truly need the full nested retrieve payload, redirect full --json to a local file and inspect it selectively instead of streaming it into the live chat context
   • treat compact retrieve as the stable host-facing projection; inspect results, reference_resolution, source_scope_policy, and recommended_hybrid_targets before reaching for nested JSON
   • do not build alternate compact schemas with ad hoc jq assumptions such as .matches
   • prefer published affordance sidecars and already-published evidence channels over ad hoc source inspection
   • inspect reference_resolution first when the user has named a document or locator in user-native terms
   • keep DocMason workspace commands sequential inside the same live answer path; do not overlap retrieve, trace, sync, status, or validate-kb while a lease-owning step is still running
4. Inspect the strongest evidence bundles, matched units, graph expansions, render references, and published-evidence sufficiency judgment.
   • when reference_resolution.status is exact, preserve that narrowing and do not let neighboring documents silently dilute it
   • when reference_resolution.status is approximate but unit_match_status is exact, preserve the approximate notice but still treat the resolved source narrowing as intentional
   • when reference_resolution.status is approximate or unresolved, keep the inline notice and answer wording honest about that boundary
   • when the question is artifact-sensitive, inspect compact retrieval fields first:
     • matched_artifact_ids
     • matched_unit_ids
     • matched_overlay_unit_ids
     • focus_render_assets
     • recommended_hybrid_targets
     • score details such as structure_context_bonus, semantic_overlay_bonus, and compare_coverage_bonus
   • when the answer boundary truly depends on exact artifact metadata such as section_path, caption_text, continuation_group_ids, procedure_hints, or semantic_labels, inspect a file-first full retrieve capture or the published artifact sidecars rather than pasting the nested payload into chat
   • treat those published payloads as an ordered evidence path: first decide whether retrieved text, structure, notes, or media already settle the claim, then inspect cited focus_render_assets, render refs, or page spans when visual confirmation would materially change the answer, and only then escalate
   • when published sufficiency fails because of hard-artifact semantic gaps, the grounded-answer path must enter the governed ask-time multimodal refresh before any raw source inspection
     • use recommended_hybrid_targets as the only legal query-aware narrowing entrypoint
     • if the turn becomes a waiter on that governed refresh, keep the same turn paused and reuse the shared result
     • once the governed refresh picks a source, complete that source's current hybrid candidates, inspect listed render or focus-render assets when present and answer-relevant, reretrieve, and retrace before treating the ask as ready to answer
     • record lightweight settlement audit fields such as render_inspection_used and inspected_render_assets when the work packet exposes visual assets
     • after one covered refresh and post-refresh retrieve/trace, close honestly with the remaining support boundary instead of starting a second same-turn refresh
     • if the governed refresh settles blocked, close the turn as abstained + governed-boundary
5. Run provenance tracing for the strongest support when you need corroboration, contradiction checks, or answer-state clarification:
   • docmason trace --source-id <source_id> --json --compact
   • docmason trace --answer-file <path> --json --compact
   • docmason trace --session-id <session_id> --json --compact
   • if you truly need nested segment support objects, redirect full --json to a local file and inspect it selectively instead of loading the full raw trace payload into the live chat context
   • treat compact trace as the stable host-facing projection; inspect answer_state, reference_resolution, source_scope_policy, issue_codes, and recommended_hybrid_targets before reaching for nested JSON
   • when the same live turn creates more than one ask-owned retrieve session or more than one plausible final trace candidate, keep an explicit artifact ledger while you work:
     • preserve the selected ask-owned session_ids that support the final answer
     • preserve the selected trace_ids that bind the final answer-file version
     • return those selected IDs to the main agent for finalize-time use instead of leaving commit-time disambiguation to complete_ask_turn()
   • if the turn produced only one ask-owned retrieve session and one final trace, ordinary automatic hydration remains sufficient
6. Inspect renders when:
   • the strongest support uses low-confidence extracted text
   • a cited unit has little or no text but does have rendered evidence
   • layout, tables, diagrams, screenshots, or visual style are part of the answer boundary
   • the odd-question plan explicitly prefers render or media
   • artifact supports expose render_page_span, bbox, or normalized_bbox that materially narrow what must be checked
   • artifact or segment supports expose focus_render_assets, which should be preferred over full-page renders when present
   • do not treat every multimodal source as a reason to reopen the raw file; render inspection is for questions whose answer boundary actually depends on visual semantics or a published render-only gap
7. Draft the canonical answer file under runtime/answers/ when conversation context exists.
   • if you need auxiliary drafts or exported scratch artifacts and the user did not specify a path, place them under runtime/agent-work/
8. When the answer is externally verified, persist the lightweight external support manifest before or alongside the final trace so the combined support contract stays machine-readable.
9. Run docmason trace --answer-file <path> --json --compact as the final grounding check.
   • when the answer depends on artifact-level support, inspect compact trace fields first:
     • supporting_artifact_ids
     • segment supporting_artifact_ids
     • segment supporting_overlay_unit_ids
     • segment support-lane counts and compact support counts
   • when you genuinely need nested artifact_supports, semantic_supports, render refs, page spans, or region boxes, inspect a file-first full trace capture or the published artifact surfaces rather than streaming the nested trace payload into chat
   • hand the same answer-file path, plus any selected session_ids / trace_ids, back for hidden finalize; prefer the structured workflow_outcome handoff when the workflow already knows the correct support_basis, selected IDs, or other finalize-owned facts
   • if finalize returns status = execute together with a repairable support_fulfillment, do one contract-aware rewrite and retrace on the same turn, then finalize once more
   • if finalize returns status = execute together with admissibility_repair, use its issue codes and suggested action to rewrite the same answer file, rerun trace, and finalize once more
   • do not render terminal closure metadata yourself; return the exact business answer and workflow-owned support facts to ask, which alone decides whether a separate status line or boundary explanation is user-visible
10. Emit one of these final answer states:

• grounded
• partially-grounded
• unresolved
• abstained

11. Return the final answer, final answer_state, overall support_basis, support boundary, and next steps to the main agent.

Escalation Rules

• If the final answer trace returns partially-grounded or unresolved, do not relabel the result as grounded.
• If the answer is externally verified or intentionally based on stable model knowledge, do not relabel it as a degraded product failure solely because the final answer is not KB-grounded.
• If the final answer trace still requires render inspection, inspect the cited renders before issuing a confident answer.
• If the evidence is contradictory, ambiguous, or insufficient, qualify the answer or abstain.
• If source-reference resolution only succeeded approximately, do not phrase the answer as though the cited document or locator was matched exactly.
• If the answer is comparative, do not finalize it from a support set dominated by one source when the trace or retrieval payload still shows weak comparative coverage.
• Treat trace grounding_reason_codes and coverage ratios as diagnostics for repair and explanation. They do not override the final answer_state.
• Do not treat retrieval alone as a complete grounded-answer workflow.
• If published KB artifacts already satisfy the required evidence channels, do not rerender source files or rummage through original_doc/ as a first reflex.

Completion Signal

• The workflow is complete when the main agent has a final answer text, a final answer state, and an explicit support or uncertainty boundary grounded in retrieval and trace results.

Notes

• This is an inner agent-facing workflow behind ask. It is not a new public docmason answer command.
• New runtime artifacts use the same four-state answer_state contract:
  • grounded
  • partially-grounded
  • unresolved
  • abstained
• support_basis remains an explicit orthogonal field. Interpret answer_state relative to that declared support basis.
• When a narrower KB-specific grounding view is needed, use trace grounding summaries, segment grounding statuses, and kb_answer_state rather than overloading the top-level answer contract.