← Về thư mục
📄 / / root / ceo / academic-research-skills / docs / design / 2026-05-15-issue-111-slr-lineage-emission-design.md

Design — #111 slr_lineage emission on systematic-review → academic-paper full handoff

Date: 2026-05-15 Issue: #111 (parented to #108 PR #110 commit 70c8678) Status: implemented (commit 7a2f789 on feat/111-slr-lineage-emission; pending PR) Scope: pipeline plumbing only. NO change to #108 referee, anchor table, protocol doc, or v3.2 venue track.

1. Problem

disclosure mode with --policy-anchor=prisma-trAIce reads slr_lineage: bool at renderer time per the G2 invariant track gate (policy_anchor_disclosure_protocol.md §3.1). When the user runs the documented deep-research systematic-review → academic-paper full → disclosure --policy-anchor=prisma-trAIce pipeline path, the renderer currently sees slr_lineage=false because no upstream component sets it. The user must manually supply mode=systematic-review to dispatch — defeating the intended automatic SLR dispatch.

disclosure is a finishing step invoked standalone after pipeline completion (academic-paper SKILL.md line 278); the renderer cannot read live state_tracker state. The only cross-session, cross-skill carrier is the Material Passport (Schema 9).

2. Decision Doc anchors (frozen, do not re-litigate)

3. Boundary distinction (load-bearing)

Schema G1 status
literature_corpus_entry.schema.json (corpus entry, per-source row) FROZEN — no fields added
Schema 9 Material Passport (run-level provenance) PERMITTED — extended in v3.6.3 (reset_boundary[]), v3.6.4 (literature_corpus[]), v3.6.7 (audit_artifact[]) per the same pattern

Adding slr_lineage to Schema 9 Material Passport top-level is not a corpus entry schema mutation and is not a G1 invariant violation. The lint and audit trail must treat the distinction explicitly so future audit rounds don't false-flag.

4. Design — three pieces

4.1 Schema 9 extension

Add one top-level optional field to Material Passport (Schema 9):

Field Type Description
slr_lineage boolean true iff any pipeline stage in this run history was deep-research systematic-review mode. Set by pipeline_orchestrator_agent at the documented handoff point. Omittable; absence is semantically equivalent to false (renderer treats unset as cold-start).

Backward compat: passports written by pre-#111 runs lack the field; renderer defaults to cold-start path (requires explicit mode= per §3.1 G2 invariant fallback rule, identical to current behavior).

4.2 Emission point

pipeline_orchestrator_agent.md emits slr_lineage at the Stage 1 → Stage 2 handoff (the only transition where new deep-research lineage can enter the run). After Stage 1/2/3 schema validation, the orchestrator computes:

slr_lineage_out = bool(incoming_passport.slr_lineage) or any(
    stage.skill == "deep-research" and stage.mode in {"systematic-review", "slr"}
    for stage in state_tracker.stages.values()
)

The OR with incoming_passport.slr_lineage is load-bearing (codex round-1 [P2] closure): a resume_from_passport=<hash> session has an empty state_tracker.stages (reconstructed from ledger only); recomputing from stages alone would overwrite a persisted true and defeat #111's auto-dispatch goal. A monotonic flag never flips back to false. Reference helper: scripts/slr_lineage.py emit(stages, incoming_slr_lineage).

Reset-boundary path (codex round-2 [P2] closure, refined round-3 [P2] closure): under ARS_PASSPORT_RESET=1 + systematic-review mode, the v3.6.3 reset protocol (academic-pipeline/references/passport_as_reset_boundary.md) freezes a passport at the FULL checkpoint and halts before the Stage 1 → Stage 2 handoff. Handoff-only emission would therefore miss the only write opportunity, and the resuming fresh session would see an empty state_tracker.stages + flag-less incoming passport → OR resolves false → PRISMA-trAIce blocks. Fix: the orchestrator prose at §"Run-level lineage emission" specifies that emit() runs before any passport write (handoff + reset-boundary). Honest accounting (round-3 closure): slr_lineage lives at passport top-level — NOT inside the reset_boundary[] ledger entry — so the v3.6.3 boundary hash (JCS over ledger entries only, see passport_as_reset_boundary.md §"The reset boundary protocol") does not cover this field. Same trust model as the other Schema 9 top-level fields. The need is correctness-at-write, not integrity-after-write.

Import ergonomics (codex round-2 [P2] closure): scripts/slr_lineage.py uses a try: from scripts.policy_anchor_disclosure_referee import SLR_MODES; except ImportError: from policy_anchor_disclosure_referee import SLR_MODES dual-path import. The repo has both namespace-style callers (from scripts.<module> per test_check_sprint_contract.py) and sibling-style callers (sys.path-prepending per test_slr_lineage_emission.py). Both paths resolve to the same module — single source of truth for SLR_MODES preserved.

origin_mode on each artifact's passport remains as today (records the directly-producing skill's mode). slr_lineage is run-level provenance, not artifact-level, so it sits at passport top-level alongside origin_skill / version_label.

Pattern precedent: this mirrors the v3.7.1 Cite-Time Provenance Finalizer (line 605 of pipeline_orchestrator_agent.md) — orchestrator computes a derived signal at a stage transition and persists it for downstream consumers.

4.3 Handoff contract documentation

shared/handoff_schemas.md Schema 9 gains an optional-fields row + a short subsection describing: - semantics (run-level, derived from any-SLR-stage scan) - producer (pipeline_orchestrator_agent at handoff transitions) - consumer (disclosure mode renderer via policy_anchor_disclosure_referee.RendererInputs.slr_lineage) - backward compat (absence = cold-start, identical to pre-#111 behavior) - G1 boundary note (passport-level vs corpus-entry-level distinction)

4.4 e2e fixture

One conformance test under scripts/ (mirroring test_policy_anchor_disclosure.py shape) exercising:

  1. Build a synthetic passport with stages["1"].mode = "systematic-review", stages["2"].mode = "full"
  2. Run orchestrator handoff logic (or its computational equivalent — see §6 implementation note)
  3. Assert outgoing passport carries slr_lineage: true
  4. Pass that passport to policy_anchor_disclosure_referee with --policy-anchor=prisma-trAIce
  5. Assert no G2 TrackGateError, no manual mode= supplied

Plus three negative cases: - Stage 1 mode = full (non-SLR) → passport slr_lineage: false → prisma-trAIce selector errors with G2 (existing behavior preserved) - Pre-#111 passport (no field) → renderer cold-start path (existing behavior preserved) - Pipeline with no Stage 1 (mid-entry from Stage 2) → no systematic-review evidence → slr_lineage: false

5. Files touched

File Change kind
shared/handoff_schemas.md additive — optional field row + subsection
academic-pipeline/agents/pipeline_orchestrator_agent.md additive — emission step in §4 Transition Management
scripts/test_slr_lineage_emission.py (new) e2e conformance fixture
CHANGELOG.md append release note under v3.7.x

Files explicitly NOT touched (matches #111 §Scope out-of-scope)

File Reason
scripts/policy_anchor_disclosure_referee.py #108 referee — contract already correct
academic-paper/references/policy_anchor_disclosure_protocol.md #108 protocol — contract already correct
academic-paper/references/policy_anchor_table.md #108 anchor table
academic-paper/references/disclosure_mode_protocol.md v3.2 venue track + #108 anchor track dispatch — already references slr_lineage as pipeline-supplied
shared/contracts/passport/literature_corpus_entry.schema.json G1 frozen (corpus entry schema, not passport schema)
Any test under scripts/test_policy_anchor_*.py #108 conformance baseline (1053 tests must stay green)

6. Implementation note — orchestrator is LLM prose, not Python

pipeline_orchestrator_agent.md is an LLM-orchestrated agent prompt, not Python. The "emission logic" lives in prose instruction to the LLM at the handoff step.

The e2e fixture under §4.4 cannot dispatch the actual agent; instead it tests:

Discipline: keep the helper minimal. No new orchestration framework, no Python wrapper around the LLM agent. Resolution is one any() over a dict.

7. Regression budget

8. Open questions

None at design time. All §4.4 Decision Doc concerns are frozen by impl-spec; the only branch is §4 / B1 above (Schema 9 top-level field), and it is selected.

9. Out of scope (defer to future issues)