[Atlas/landscape] Immunology of aging + immune memory — Allen Immunology domain blocked

Quest: Landscape Analyses

> Goal. Maintain living maps of scientific fields — where research clusters, where the white space is, what the frontiers are. These maps drive quest_gaps (by surfacing empty cells) and quest_inventions (by tagging cells as novel or saturated). Generalizes the existing AI-tools-landscape pattern to every scientific domain SciDEX cares about.
>
> Distinct from ad-hoc review articles: a landscape here is a structured artifact — domain partitioned into cells, each cell with density/recency/controversy metrics, each cell linked to the literature and the world model. It's queried programmatically by other quests.

Parent: [scidex_economy_design_spec.md](scidex_economy_design_spec.md).
Existing AI-tools case: [q-ai-tools-landscape_spec.md](q-ai-tools-landscape_spec.md), [4be33a8a_4095_forge_curate_ai_science_tool_landscape_spec.md](4be33a8a_4095_forge_curate_ai_science_tool_landscape_spec.md).

---

What a landscape analysis looks like

An instance of this artifact class covers one domain (e.g. "CRISPR base editing", "RNA therapeutics for CNS", "small-molecule PROTACs"). It has:

• domain (canonical string; pinned to a world-model subgraph)
• cells: list of {cell_id, label, paper_count, recency_score, controversy_score, saturation, gap_hint}
• boundaries: adjacency edges to neighboring landscapes (so a gap in the boundary region can route to either)
• freshness_date: when the corpus was last ingested
• coverage_completeness (0-1): how much of the named domain is actually mapped
• open_gaps: list of cell_ids with saturation < 0.3 (the white-space frontier)
• top_papers_by_cell: 3-5 representative papers per cell
• frontier_commentary: 2-3 paragraphs of Synthesizer-written narrative on where the field is going

---

Inputs

• Atlas literature index (papers, abstracts, cited-by graph)
• The world-model framework's 7 representations per entity (world_model_framework_spec.md)
• Existing gap rows (a gap in domain X tells us X needs more mapping coverage)
• Previous landscape analysis for the same domain (for longitudinal tracking)

Outputs

Per run, one or more landscape_analysis artifacts. Each admitted artifact feeds:

• quest_gaps — each cell with saturation < 0.3 is emitted as a candidate gap (downstream quest decides if it's actionable)
• quest_inventions and quest_experiments — novelty(cell) lookup
• /showcase/economy dashboard — landscape heatmaps

Task shape

task_type = multi_iter:

• artifact_class = "landscape_analysis"
• required_roles = ["surveyor", "cartographer", "critic"]
• debate_rounds = 3
• max_iterations = 2 (landscapes are expensive to build; don't thrash)
• target_cell = domain
• acceptance_criteria:

1. Generation

Round 1 — Survey. Surveyor agents pull a sized corpus (5k-20k papers depending on domain) from the Atlas literature index and produce an initial clustering. Clusters come with proposed labels (LLM-summarized) and per-cluster paper lists.

Round 2 — Cartography. Cartographer agent takes the clusters and produces:

• A clean partition (no two cells with >20% paper overlap)
• Per-cell metrics (paper_count, median publication date → recency_score, cited-by dispersion → controversy_score, paper_density_per_unit_time → saturation)
• Boundary edges to neighboring landscapes (looked up via domain_adjacency in Atlas)
• Initial gap_hint per under-saturated cell

• Are any important keywords/entities missing? (Cross-ref against the world-model graph for this domain — any high-connectivity entity with no cell assignment is a miss.)
• Is saturation well-calibrated? (Compare to a held-out subsample of papers.)
• Are the labels understandable to a non-expert? (LLM readability check.)

2. Admission

• coverage_completeness ≥ 0.7: ≥70% of the world-model subgraph's high-connectivity entities land in some cell.
• cell_cohesion ≥ 0.6: measured via within-cluster vs between-cluster embedding distance (standard silhouette or Davies-Bouldin threshold).
• freshness_date within 30 days of admission.
• Cross-reference consistency: Sanity-check against the world model; no cell contradicts a high-confidence world-model edge.

3. Refresh cadence

• saturation > 0.5 cells: refresh every 8 weeks (slow-moving fields)
• saturation 0.2-0.5 cells: refresh every 4 weeks (active fields)
• saturation < 0.2 cells: refresh every 2 weeks (frontier fields — highest novelty value)

4. Interactions

• quest_gaps — reads open_gaps from each landscape; the gap factory's scanner component (f456463e9e67_atlas_build_automated_gap_scanner_analy_spec.md) ingests landscape cells as input context.
• quest_inventions — novelty(cell) lookup drives seeding priority.
• quest_experiments — the no_redundant_prior_art admission check consults this landscape's top_papers_by_cell.
• Atlas world model — bidirectional: world model entities get mapped into cells; landscape cells become a view/aggregation over the world-model graph.

5. Showcase

Showcase landscape artifacts demonstrate the full mapping pipeline for a domain of current strategic interest. UI treatment: interactive 2D cell map (umap or similar), click a cell to see its papers + saturation + any inventions/experiments rooted in it.

6. Capacity

• Default: 2 concurrent landscape tasks (expensive).
• One landscape build is ~6-10 agent-hours for the first three rounds, plus ~2-3 hours if iteration kicks in.
• The quest maintains a schedule — fields get queued by refresh-due-date.

7. Open questions

• How do we pick which domains to map first? (Proposed: seed with ~12 high-strategic-value SciDEX domains; user-configurable; add domains as they become relevant to admitted inventions.)
• Should the AI-tools-landscape spec fold into this one? (Proposed: it becomes a specialized sub-case with custom cell labels; shares the refresh and admission machinery.)
• How do we handle cross-domain landscapes ("all of ML-for-biology")? (Proposed: compose multiple landscapes via the boundaries edges; the UI renders a federated view.)

Work Log

2026-04-25 22:25 PT — Task cfecbef1-ea59-48a6-9531-1de8b2095ec7

• Started the Allen Immunology domain slice for immunology-aging-memory after a staleness check confirmed no sibling task or existing artifact on origin/main already covered this domain.
• Grounding plan for this iteration:

• Current world-model cross-checks before implementation:

2026-04-25 19:31 PT — Task cfecbef1-ea59-48a6-9531-1de8b2095ec7

• Converted the Allen Immunology landscape builder from a file-only draft into a DB-backed artifact publisher: scripts/build_landscape_immunology_aging_memory.py now writes the JSON artifact and upserts landscape_analyses for domain immunology-aging-memory.
• Regenerated the artifact at artifacts/landscape_immunology_aging_memory.json and persisted landscape_analyses.id=2 with coverage_completeness=0.857, cell_cohesion=0.63, open_gap_count=12, total_papers=136.
• Added explicit domain_description, generated_at, methodology, per-run coverage metrics, and generated gap IDs so downstream consumers can inspect the artifact through PostgreSQL without scraping repo files.
• Verification: python3 -m py_compile scripts/build_landscape_immunology_aging_memory.py passed; python3 scripts/build_landscape_immunology_aging_memory.py completed successfully and printed the persisted row id.

2026-04-25 23:05 PT — Task cfecbef1-ea59-48a6-9531-1de8b2095ec7

• Promoted the Allen Immunology domain slice from a file-only landscape into a SciDEX-native published artifact: the builder now uses a stable artifact id, upserts artifacts, and emits concrete knowledge_gaps rows linked back to landscape_analyses.id=2.
• Broadened the weakest PubMed survey queries so the tissue-atlas and heterogeneity cells are grounded by non-zero corpus counts instead of relying only on cached representative papers.
• Regenerated artifacts/landscape_immunology_aging_memory.json; current run produced total_papers=326, coverage_completeness=0.857, cell_cohesion=0.63, and emitted_gap_ids=12.
• Verification: