Lattice gaps

Lattice Gaps

Overview

Lattice Gaps consumes evidence prepared by Lattice Find Papers and then performs compact gap discovery or direct Anti-Gap verification. The default behavior is a short, high-signal gap audit: evidence boundary, top drift/problem signals, up to three candidate gaps, minimal Anti-Gap checks, and the next smallest action. Full evidence trees and long matrices are optional and only produced when the user explicitly asks for full evidence mapping.

This skill is downstream of Lattice Find Papers. Lattice Find Papers retrieves and screens literature, imports screened DOI records into Zotero when available, generates runs/<run_id>/request_PDF/doi_list.md, writes find_papers_outputs/tables/zotero_import_manifest.csv, and collects user-downloaded PDFs. Lattice Gaps reads those outputs. It does not import DOI records, export batches, create Zotero collections, or manage Zotero libraries.

Modes

• compact discovery mode default: use when the user wants to find gaps from project evidence. Keep the output short enough for manual review.
• full evidence mapping mode: use only when the user explicitly asks for full data tree, problem tree, method tree, data-problem matrix, or a systematic evidence map.
• verify-only mode: use when the user already has one or more proposed gaps and asks for Anti-Gap, 反研究空白, falsification, narrowing, or whether the gap is real. Skip data-tree/problem-tree discovery unless needed to understand the proposed gap.

Input Contract

Default input path:

runs/<run_id>/request_PDF/

Use this resolution order:

1. If the user gives a Lattice Research Skill run directory, inspect <run_dir>/request_PDF/ and the run metadata.
2. If the user gives a folder path, inspect that folder for PDFs, full text, DOI lists, BibTeX/RIS/CSV files, notes, and request-paper outputs.
3. If the user gives only a run id and a likely runs/ root exists in the workspace, resolve runs/<run_id>/request_PDF/.
4. If no explicit folder is given, inspect the current project folder for request_PDF/, request papers, requested_papers, doi_list.md, and recent Lattice Find Papers outputs.
5. Read find_papers_outputs/tables/zotero_import_manifest.csv when available.
6. Probe Zotero read access after project-folder inspection, then read only matching project items, relevant DOI/title matches, or the user-specified Zotero collection.

Read doi_list.md and zotero_import_manifest.csv as seed and status files, not as proof. DOI metadata can locate Zotero items or request targets, but it cannot support a full-text claim by itself. If only DOI/metadata exists, send the item back to Lattice Find Papers for Zotero import, Request PDF, or manual full-text handling.

Evidence Sources And Evidence Levels

• Local project scope: inspect the project folder and prior Request PDF folders first. Treat local PDFs and supplements as the strongest starting evidence.
• Zotero read scope: use the Zotero skill/MCP after local inspection. Run readiness checks first; search only by manifest DOI, title, project terms, user-specified collection, or author-year. Use Zotero attachments and indexed full text only when available and relevant to the project.
• Missing evidence scope: if local files and Zotero do not provide full text, figures/tables, or supplements needed for data/experiment claims, classify the item as an evidence blocker and return it to Lattice Find Papers/Request PDF. Do not perform DOI import inside Lattice Gaps.
• Anti-Gap verification scope: default to the minimum search needed to challenge each candidate. Check whether the same correction, re-modeling, parameter update, benchmark, or review already exists. Expand into broader academic search only when a candidate survives the first pass or the user asks for a full audit.
• Final claim scope: do not claim a gap is globally proven absent. State the searched sources, terms, date, and remaining uncertainty.

Evidence levels:

• L0_metadata: DOI, title, bibliographic record, citation metadata.
• L1_abstract: abstract, snippet, indexing page, or secondary summary.
• L2_indexed_text: Zotero indexed full text, HTML text, or partial readable text without figures/tables.
• L3_full_text: readable PDF/full text with methods, results, figures, and tables.
• L4_supplement: supplementary methods, tables, raw parameter files, datasets, or code.

Parameter values and model assumptions should be treated as strong evidence only at L3_full_text or higher unless the value is explicitly visible in an authoritative online table, dataset, standard, or preprint full text.

Hard Boundaries

• Do not skip project-folder inspection before going to Zotero or online search.
• Do not scan a whole Zotero library without project constraints; searches must be bounded by DOI list, title, project name, material/system, mechanism, parameter, or author-year.
• Do not write to Zotero, import DOI records, export batches, or create collections. Those actions belong to Lattice Find Papers.
• Do not use abstracts as full-text evidence.
• Do not treat online metadata-only fallback as a data/experiment gap proof. It can refute, narrow, or locate evidence during Anti-Gap verification, but it cannot replace full text, figures/tables, supplements, datasets, or methods.
• Do not produce detailed per-paper literature annotations; use only compact evidence anchors such as filename, page, figure, table, or section.
• Do not output full data trees, problem trees, relation matrices, or long per-paper tables in compact discovery mode.
• Do not generate more than three candidate gaps by default. Rank and keep only the highest-value signals.
• Do not treat "unified framework", "multi-factor coupling", "data normalization", "database/platform", or "more variables" as a gap unless it survives open-domain Anti-Gap verification.
• Do not treat a numeric difference as a parameter gap until unit, parameter definition, material state, boundary condition, and measurement/modeling method have been checked.

Workflow

1. Build the project evidence set.

   • Inspect the project folder and prior request-paper folders first.
   • List PDFs, supplements, full-text files, doi_list.md, zotero_import_manifest.csv, BibTeX/RIS/CSV files, and Lattice Find Papers outputs.
   • Extract DOI/title/author-year seeds and mark which items have local full text, Zotero PDF, indexed full text, metadata only, or manual-download status.
   • Separate primary papers from secondary/background papers using the user's topic, filenames, DOI list, and available metadata.

2. Probe Zotero and read matching evidence.

   • Use the Zotero skill/MCP readiness check before Zotero operations.
   • Determine whether Zotero is readable.
   • Search Zotero by DOI/title/project terms and record whether each match has attachment PDF, indexed full text, metadata only, or no match.
   • Use metadata-only or indexed-full-text-only items as evidence-level markers, not as full-text proof.

3. Decide mode.

   • If the user provides proposed gaps and asks for verification, switch to verify-only mode and jump to Anti-Gap verification.
   • If the user asks for discovery, use compact discovery mode unless they explicitly request full mapping.
   • If evidence is insufficient for discovery, output the missing PDFs/supplements/datasets and route back to Lattice Find Papers or Request PDF.

4. Scan sources by role and evidence level.

   • Primary papers: read Results, Figures/Tables, Methods/SI, Discussion, and limitations.
   • Secondary papers: use only for context, method comparison, or boundary conditions.
   • Zotero indexed full text: use as partial text only; do not infer figure/table data unless the attached PDF or table is readable.
   • Online metadata/abstracts: use only for Anti-Gap refutation, narrowing, or locating items that should return to Lattice Find Papers/Request PDF; do not use them for full-text conclusions.
   • Record evidence anchors, not long quotations.

5. Build a compact evidence sketch.

   • Identify only the top 3-5 data/method/experiment nodes that matter for the user's research question.
   • For defect-focused topics, prioritize defect type, size, depth, density, distribution, measurement method, scale, condition, and outcome.
   • For parameter-focused topics, prioritize parameter name, value/range, unit, material/system, state variable, boundary condition, measurement/model method, year, and evidence level.
   • Keep this sketch internal unless it directly supports a candidate gap.

6. Use full trees only when requested.

   • In full evidence mapping mode, build data tree, problem tree, method tree, and relation matrix.
   • Start from observed objects and data families.
   • For defect-focused topics, organize by defect type, size, depth, length, shape, density, distribution, measurement method, scale, condition, and outcome.
   • For parameter-focused topics, organize by parameter name, value/range, unit, material/system, state variable, boundary condition, measurement/model method, year, and evidence level.
   • Keep variable names measurable and comparable.

7. Build the problem tree only if needed.

   • Split the topic into scientific questions such as starting point, propagation, energy, boundary condition, scale transfer, mechanism, and method visibility.
   • Convert broad topics into testable questions.
   • In compact mode, keep only the 1-3 questions directly tied to candidate gaps.

8. Build the data-problem relation matrix only in full mode.

   • Link data nodes to problem nodes using relation types: supports, contradicts, correlates_with, causes_claimed, modulates, measures, not_reported, incomparable, and only_tested_under.
   • Mark whether each relation is direct evidence, inference, missing evidence, or blocker.

9. Identify candidate gaps inside the evidence set.

   • Missing node: a core variable or mechanism is absent from all readable evidence sources.
   • Weak edge: a relation is claimed but not measured.
   • Contradiction: comparable conditions lead to conflicting claims.
   • Boundary gap: evidence exists only under narrow materials, scales, or test conditions.
   • Method gap: current methods cannot observe the claimed process.
   • Reporting blocker: missing parameters make comparison impossible.
   • Parameter assumption drift: earlier work uses a fixed/default parameter or narrow range, while later evidence reports a substantially different or state-dependent value that could change model or experimental conclusions.
   • Rank candidates by disruptive potential, evidence level, condition comparability, and model/experiment impact.
   • Keep at most three candidates in compact mode.

10. Run the parameter assumption drift audit and pseudo-gap red-flag filter when parameter values matter.

• Extract old values as ranges when the literature uses ranges, not as a strawman single value. Example: treat Li modulus as 30-50 MPa if that is the practical range, not only 50 MPa.
• Normalize units and parameter definitions; distinguish Young's modulus, shear modulus, storage modulus, hardness, fitted effective modulus, and model calibration parameters.
• Compare material state and boundary conditions: volume/SOC, temperature, pressure, strain rate, phase, microstructure, loading path, interface condition, and measurement/model method.
• Flag >=2x inconsistency as a strong reference signal when conditions are comparable or the condition-dependence itself is the scientific issue.
• Flag >=10x inconsistency, new state-dependence, or direction-changing thresholds as disruptive/collapse signals.
• Ask whether replacing the old parameter range with the new value would change stress, strain, failure threshold, interface stability, crack initiation, dendrite growth, or another model conclusion.
• Compare wording with references/PSEUDO_GAP_LIBRARY.md; pseudo-gap detection is a red-flag filter, not the final judgment.
• If a text file is provided, optionally run scripts/detect_pseudo_gaps.py to catch common pseudo-gap phrases before manual evidence review.
• Rewrite any surviving gap as: "Under boundary condition X, prior models/experiments assumed parameter P in range A, while later evidence reports P as B or state-dependent; it remains untested whether updating P changes conclusion Y."

11. Apply embedded Anti-Gap verification.

• Read references/PSEUDO_GAP_LIBRARY.md.
• Restate each candidate/proposed gap narrowly. Do not broaden it.
• In compact mode, run only three refutation checks first: same re-modeling/re-interpretation, same parameter or method update, and review/benchmark/protocol evidence that already closes the gap.
• Build searches from mechanism terms, variable names, material/system names, boundary conditions, methods, synonyms, abbreviations, adjacent-field terminology, and likely review/benchmark terms when broader checking is needed.
• For parameter drift, search old parameter names/ranges, new reported values, state variables, model sensitivity terms, "revisited", "updated parameter", "finite element", "calibration", "benchmark", and the specific material/interface.
• Search outside the initial evidence set only enough to find prior studies, datasets, validated methods, negative results, reviews, benchmarks, standards, protocols, or preprints that could refute the candidate.
• Record searched sources, exact terms, and search date.
• Classify as survives narrowly, too broad, pseudo-gap, evidence blocker, not yet falsified after open search, or refuted by external evidence.
• Rewrite surviving gaps as mechanism-variable-boundary-condition claims.

Verify-Only Mode

Use this mode when the user already has proposed gaps.

1. Restate the proposed gap exactly in one sentence.
2. Identify what would refute it: same mechanism, parameter, dataset, method, material system, boundary condition, or completed re-modeling/re-interpretation.
3. Check references/PSEUDO_GAP_LIBRARY.md for red flags.
4. If a text file is provided, optionally run scripts/detect_pseudo_gaps.py before manual review.
5. Inspect any provided local PDFs/Zotero evidence if available.
6. Run only the external checks needed to falsify or narrow the proposed gap.
7. Return a judgment: refuted, too broad, pseudo-gap, evidence blocker, not yet falsified after open search, or survives narrowly.

Do not invent new gaps in verify-only mode unless the user explicitly asks for new gap discovery.

Output Format

Respond in Chinese by default. Compact discovery and verify-only mode must use this short format:

1. 证据边界: what was read, what is missing, and the current evidence level.
2. 高价值信号: at most five signals; use one compact table only when it improves readability.
3. 候选 Gap: at most three candidates. For each, give one-sentence definition, why it may hold, and biggest weakness.
4. Anti-Gap 快查: minimum refutation checks and result: 保留, 降级, 证据不足, or 已被反证.
5. 下一步最小动作: 1-3 concrete actions, such as one missing PDF, one supplement, one sensitivity analysis, or one re-modeling check.

Full evidence mapping mode may add data tree, problem tree, method tree, relation matrix, and detailed per-paper evidence tables, but only after the user explicitly asks for that expanded output.