dsh-math-modeling-agent 0.3.1 → 0.5.0

package/skills/math-modeling-agent/references/report-contract.md

@@ -1,122 +1,130 @@
-# Report Contract
-
-## Attempt report
-
-Every attempt report contains exactly:
-
-1. one round objective;
-2. candidate/direction;
-3. assumption delta;
-4. claims changed;
-5. actual execution or derivation artifacts;
-6. new evidence IDs;
-7. obligations closed;
-8. issues opened and closed;
-9. delta from prior attempt;
-10. valid-progress decision;
-11. budget usage;
-12. current status;
-13. one next action.
-
-Do not copy full prior solutions or private reasoning.
-
-## Terminal report
-
-State the frozen question, final scientific status, answer/recommendation, verified claims and evidence links, conditional/unresolved claims, assumptions and sensitivity, candidate comparison, data and parameter sources, validation, failure cases, limitations, reproducibility commands, and resume conditions when non-SOLVED.
-
-整体感 requirement (v0.3): the terminal report is ONE narrative, not a
-concatenation of sub-question reports — it must show the question-dependency
-chain (which model fed which), the shared global assumptions, and the
-cross-sub-question consistency checks. A reader must be able to see the whole
-problem as one connected argument.
-
-## Chat display — the runlog digest (MUST output, v2.7)
-
-The following is shown DIRECTLY IN THE CHAT, not only on disk.
-
-### Readability standard — write for a human reader, never a bullet list
-
-Every section is an EXPOSITION, not a summary: state WHY the step is taken,
-HOW the reasoning runs (derivation, mechanism, first-principles chain), WHAT
-came out (numbers with units), and HOW it is checked. A reader who knows
-nothing about the problem must be able to follow each section without
-guessing.
-
-Style: step-by-step numbered exposition with plain direct language (this is
-a working document, not a final paper — no paper-style formal phrasing
-required), but with MORE detail than a paper would carry: every parameter
-choice gets its reason, every formula is followed by a plain-language
-explanation of each symbol, every step ends with a transition line
-("这一步保证了…，为下一步…做好准备"), and every conclusion states what it
-would take to be wrong. Minimum expectations per section:
-
-- 问题重述：restate the question in your own words, explaining every
-  parameter's meaning and where it comes from in the problem text.
-- 问题分析：walk the physical/geometric mechanism step by step (e.g. why the
-  nodes move radially, how the focal geometry pins the paraboloid), each
-  step justified by problem text, data, or derivation.
-- 模型假设：for EVERY assumption give the reason it is justified
-  (original wording / data statistic / derivation / literature) and what
-  happens if it fails.
-- 模型建立与求解：write the derivation in full (equations, variable
-  definitions, units, how the equation is solved, numerical method and why
-  it was chosen), not just the final formula.
-- 验证：explain the verification LOGIC for each obligation (what would
-  falsify it, what tolerance is used and why), then the outcome.
-- 鲁棒性：for each perturbation say what was varied, why that parameter,
-  and what would count as a conclusion flip.
-- 评价与改进 / 参考文献：state the comparison criterion and map each
-  reference to the exact method it supports.
-
-### Intake digest (once, at intake completion) — the six-section deep restatement
-
-【题目要点摘录】the problem's own numbers and constraints, quoted, each explained
-【背景解读】engineering/physical mechanism, walked through
-【数据语义】every attachment column by column (meaning/units/coordinate system/magnitudes), with observed statistics
-【隐含信息与歧义清单】numbered; each = LLM reading + derivation/basis + risk (the user corrects entries here)
-【各子问题数学本质判定】every sub-question (any count) classified + why
-【待澄清问题】open questions inviting user background/corrections/references
-
-After the D0 exchange, show the delta applied to problem-brief.md.
-
-### Quality self-check (MUST run before every D4 verdict)
-
-Before presenting a D4 verdict, run the self-check on the current attempt's
-report and show the result in the chat: for each of the eight digest sections,
-confirm whether it answers all four questions — WHY (why this step), HOW
-(how the reasoning runs), WHAT (result with units), CHECK (how it is
-verified, or what would falsify it). Any section failing one of the four is
-flagged as incomplete and the round is NOT presented for verdict until fixed.
-The self-check is a human-facing declaration: the user decides whether the
-exposition is actually sufficient. This is the quality gate; the machine
-gates above only guarantee existence and structure.
-
-### Per-round digest (after EVERY attempt round)
-
-```text
-## Attempt N 摘要
-问题重述：<1-2 sentences, current subproblem>
-问题分析：<2-4 sentences, this round's inputs/insights/delta vs last round>
-模型假设：<沿用 X 条 + 新增 Y 条 + 修订 Z 条；修订注明原因；重中之中，不得省略>
-模型建立与求解：<core equations + method + key numbers>
-验证：<obligations ✓/✗ + evidence levels + most important failure>
-鲁棒性：<this round's sensitivity results, or "未覆盖，计划在 attempt M">
-评价与改进：<candidate comparison / limitations / next improvement>
-参考文献：<new citations or "无新增">
-```
-
-Rules:
-- The restatement is frozen after intake; analysis and assumptions ROLL FORWARD
-  per subproblem and per round (they change, and the digest shows the delta).
-- Private chain-of-thought is never shown; only auditable content.
-- The digest is a hard requirement of the round contract: a round whose digest
-  is missing does not count as a completed round.
-
-## Cleanup checklist (before ANY terminal transition)
-
-1. attempts/<n>/code holds only final reproducible scripts with header comments.
-2. Exploratory/buggy code is moved to _drafts/ or failed/code-drafts/ — never deleted.
-3. No debug prints or machine-specific hardcoded paths.
-4. figures → plots/, tables → data/, deliverable at run root.
-5. reproducibility.json matches the final scripts.
-6. run-state validate passes; every obligation has a terminal status.
+# Report Contract
+
+## Attempt report
+
+Every attempt report contains exactly:
+
+1. one round objective;
+2. candidate/direction;
+3. assumption delta;
+4. claims changed;
+5. actual execution or derivation artifacts;
+6. new evidence IDs;
+7. obligations closed;
+8. issues opened and closed;
+9. delta from prior attempt;
+10. valid-progress decision;
+11. budget usage;
+12. current status;
+13. one next action.
+
+Do not copy full prior solutions or private reasoning.
+
+## Terminal report
+
+State the frozen question, final scientific status, answer/recommendation, verified claims and evidence links, conditional/unresolved claims, assumptions and sensitivity, candidate comparison, data and parameter sources, validation, failure cases, limitations, reproducibility commands, and resume conditions when non-SOLVED.
+
+整体感 requirement (v0.3): the terminal report is ONE narrative, not a
+concatenation of sub-question reports — it must show the question-dependency
+chain (which model fed which), the shared global assumptions, and the
+cross-sub-question consistency checks. A reader must be able to see the whole
+problem as one connected argument.
+
+## Chat display — the runlog digest (MUST output, v2.7)
+
+The following is shown DIRECTLY IN THE CHAT, not only on disk.
+
+### Readability standard — write for a human reader, never a bullet list
+
+Every section is an EXPOSITION, not a summary: state WHY the step is taken,
+HOW the reasoning runs (derivation, mechanism, first-principles chain), WHAT
+came out (numbers with units), and HOW it is checked. A reader who knows
+nothing about the problem must be able to follow each section without
+guessing.
+
+Style: step-by-step numbered exposition with plain direct language (this is
+a working document, not a final paper — no paper-style formal phrasing
+required), but with MORE detail than a paper would carry: every parameter
+choice gets its reason, every formula is followed by a plain-language
+explanation of each symbol, every step ends with a transition line
+("这一步保证了…，为下一步…做好准备"), and every conclusion states what it
+would take to be wrong. Minimum expectations per section:
+
+- 问题重述：restate the question in your own words, explaining every
+  parameter's meaning and where it comes from in the problem text.
+- 问题分析：walk the physical/geometric mechanism step by step (e.g. why the
+  nodes move radially, how the focal geometry pins the paraboloid), each
+  step justified by problem text, data, or derivation.
+- 模型假设：for EVERY assumption give the reason it is justified
+  (original wording / data statistic / derivation / literature) and what
+  happens if it fails.
+- 模型建立与求解：write the derivation in full (equations, variable
+  definitions, units, how the equation is solved, numerical method and why
+  it was chosen), not just the final formula.
+- 验证：explain the verification LOGIC for each obligation (what would
+  falsify it, what tolerance is used and why), then the outcome.
+- 鲁棒性：for each perturbation say what was varied, why that parameter,
+  and what would count as a conclusion flip.
+- 评价与改进 / 参考文献：state the comparison criterion and map each
+  reference to the exact method it supports.
+
+### Intake digest (once, at intake completion) — the six-section deep restatement
+
+【题目要点摘录】the problem's own numbers and constraints, quoted, each explained
+【背景解读】engineering/physical mechanism, walked through
+【数据语义】every attachment column by column (meaning/units/coordinate system/magnitudes), with observed statistics
+【隐含信息与歧义清单】numbered; each = LLM reading + derivation/basis + risk (the user corrects entries here)
+【各子问题数学本质判定】every sub-question (any count) classified + why
+【待澄清问题】open questions inviting user background/corrections/references
+
+After the D0 exchange, show the delta applied to problem-brief.md.
+
+### Quality self-check (MUST run before every D4 verdict)
+
+Before presenting a D4 verdict, run the self-check on the current attempt's
+report and show the result in the chat: for each of the eight digest sections,
+confirm whether it answers all four questions — WHY (why this step), HOW
+(how the reasoning runs), WHAT (result with units), CHECK (how it is
+verified, or what would falsify it). Any section failing one of the four is
+flagged as incomplete and the round is NOT presented for verdict until fixed.
+The self-check is a human-facing declaration: the user decides whether the
+exposition is actually sufficient. This is the quality gate; the machine
+gates above only guarantee existence and structure.
+
+### Per-round digest (after EVERY attempt round)
+
+```text
+## Attempt N 摘要
+问题重述：<1-2 sentences, current subproblem>
+问题分析：<2-4 sentences, this round's inputs/insights/delta vs last round>
+模型假设：<沿用 X 条 + 新增 Y 条 + 修订 Z 条；修订注明原因；重中之中，不得省略>
+模型建立与求解：<core equations + method + key numbers>
+验证：<obligations ✓/✗ + evidence levels + most important failure>
+鲁棒性：<this round's sensitivity results, or "未覆盖，计划在 attempt M">
+评价与改进：<candidate comparison / limitations / next improvement>
+参考文献：<new citations or "无新增">
+```
+
+Rules:
+- The restatement is frozen after intake; analysis and assumptions ROLL FORWARD
+  per subproblem and per round (they change, and the digest shows the delta).
+- Private chain-of-thought is never shown; only auditable content.
+- The digest is a hard requirement of the round contract: a round whose digest
+  is missing does not count as a completed round.
+
+## Cleanup checklist (before ANY terminal transition)
+
+1. attempts/<n>/code holds only final reproducible scripts with header comments.
+2. Exploratory/buggy code is moved to _drafts/ or failed/code-drafts/ — never deleted.
+3. No debug prints or machine-specific hardcoded paths.
+4. figures → plots/, tables → data/, deliverable at run root.
+5. reproducibility.json matches the final scripts.
+6. run-state validate passes; every obligation has a terminal status.
+
+## v0.5 machine contract
+
+Every attempt report contains the nine scientific sections in this order: 问题重述、问题分析、模型假设、符号说明、模型建立与求解、验证、鲁棒性分析、模型的评价与改进、参考文献. The existing 13 administrative fields remain front matter and are not a substitute for scientific exposition.
+
+The report validator emits section coverage, WHY/HOW/WHAT/CHECK diagnostics, symbol errors, unsupported claim language, evidence references and requirement coverage. Missing or stale requirement coverage blocks report/evaluation/terminal transitions; partial coverage may only produce CONDITIONAL.
+
+The paper-evidence directory is an evidence package, not a silently generated manuscript. Every section cites claim/evidence IDs and preserves limitations, failures and unresolved scope.

package/skills/math-modeling-agent/references/run-directory.md

@@ -1,62 +1,65 @@
-# Run Directory Contract (v2)
-
-The run directory is the ONLY place work artifacts live. It must stay clean at
-every moment, not only at the end.
-
-## Layout
-
-```text
-math-modeling-runs/<task-id>/
-├── run.json  ledger.json  events.jsonl        # run-state files (only run-state.mjs writes them)
-├── problem-brief.md  inputs.json              # frozen restatement + input manifest
-├── attempts/<n>/
-│   ├── report.md                              # 13-section contract + 8-section digest
-│   ├── code/                                  # final reproducible scripts ONLY (clean)
-│   ├── data/                                  # results: json/csv/npy/xlsx
-│   ├── plots/                                 # figures
-│   └── _drafts/                               # this attempt's exploratory code (kept, isolated)
-├── failed/
-│   ├── directions/<id>/                       # direction-level failures: wall memo + code + why
-│   ├── code-drafts/<attempt>-<name>/          # implementation-level failures: buggy/timed-out/deprecated
-│   └── issues.md                              # failure ledger index (see below)
-├── research/                                  # literature (only when research happened)
-├── walls/                                     # breakthrough memos (legacy path kept)
-├── result.xlsx                                # the deliverable (when the problem asks for one)
-├── reproducibility.json                       # final command list + env + hashes
-└── final-report.md                            # terminal report
-```
-
-## Failure classification (failure ledger)
-
-Every problem in ANY stage is recorded in `failed/issues.md` with:
-
-```text
-#<seq> | 环节(DATA/TOOL/IMPLEMENTATION/MODEL/VALIDATION/EVIDENCE/RESEARCH_GAP) | 发现时间/attempt
-症状：…
-根因：…
-处置：修了/绕过/放弃(→wall memo id)
-影响：哪些 claim/obligation 被污染，下游如何重验
-```
-
-Direction-level failures additionally get `failed/directions/<id>/wall.md`
-(target, paths tried, exact wall, wall type, breakthrough condition, restart
-checklist) and their code snapshot. Implementation-level failures get
-`failed/code-drafts/<attempt>-<name>/` with a README stating the failure.
-
-## Cleanup gate (MUST run before ANY terminal transition)
-
-1. attempts/<n>/code contains only the final reproducible scripts; each starts
-   with a header comment: purpose, inputs, outputs, reproduction command.
-   Exploratory/buggy versions move to `_drafts/` or `failed/code-drafts/` — never deleted.
-2. No debug prints, unused variables, or hardcoded machine-specific paths
-   (the data-root constant is the only allowed absolute path).
-3. Figures in `plots/`, tables in `data/`; the deliverable at the run root.
-4. reproducibility.json is complete and matches the final scripts.
-5. `run-state.mjs validate <dir>` passes and every obligation has a terminal status.
-6. The cleanup itself is noted in the terminal transition reason.
-
-## Tidiness invariants
-
-- Code is refactored as soon as it stabilizes within an attempt (not only at the end).
-- Every script file maps 1:1 to an attempt artifact or a final deliverable.
-- Nothing outside the run directory is written by the modeling work.
+# Run Directory Contract (v2)
+
+The run directory is the ONLY place work artifacts live. It must stay clean at
+every moment, not only at the end.
+- After intake, calculations read only inputs/raw and inputs/manifest.json. External source paths are provenance only; source changes are diagnostic and do not alter the frozen snapshot.
+
+## Layout
+
+```text
+math-modeling-runs/<task-id>/
+├── run.json  ledger.json  events.jsonl        # run-state files (only run-state.mjs writes them)
+├── problem-brief.md  inputs.json              # frozen restatement + input manifest
+├── inputs/raw/  inputs/manifest.json           # immutable source snapshot + hashes
+├── attempts/<n>/
+│   ├── report.md                              # 13-section contract + 8-section digest
+│   ├── code/                                  # final reproducible scripts ONLY (clean)
+│   ├── data/                                  # results: json/csv/npy/xlsx
+│   ├── plots/                                 # figures
+│   └── _drafts/                               # this attempt's exploratory code (kept, isolated)
+├── failed/
+│   ├── directions/<id>/                       # direction-level failures: wall memo + code + why
+│   ├── code-drafts/<attempt>-<name>/          # implementation-level failures: buggy/timed-out/deprecated
+│   ├── failures.jsonl                          # structured failure records
+│   └── issues.md                              # failure ledger index (see below)
+├── research/                                  # literature (only when research happened)
+├── walls/                                     # breakthrough memos (legacy path kept)
+├── result.xlsx                                # the deliverable (when the problem asks for one)
+├── reproducibility.json                       # final command list + env + hashes
+└── final-report.md                            # terminal report
+```
+
+## Failure classification (failure ledger)
+
+Every problem in ANY stage is recorded in failed/failures.jsonl and indexed in failed/issues.md with:
+
+```text
+#<seq> | 环节(DATA/TOOL/IMPLEMENTATION/MODEL/VALIDATION/EVIDENCE/RESEARCH_GAP) | 发现时间/attempt
+症状：…
+根因：…
+处置：修了/绕过/放弃(→wall memo id)
+影响：哪些 claim/obligation 被污染，下游如何重验
+```
+
+Direction-level failures additionally get `failed/directions/<id>/wall.md`
+(target, paths tried, exact wall, wall type, breakthrough condition, restart
+checklist) and their code snapshot. Implementation-level failures get
+`failed/code-drafts/<attempt>-<name>/` with a README stating the failure.
+
+## Cleanup gate (MUST run before ANY terminal transition)
+
+1. attempts/<n>/code contains only the final reproducible scripts; each starts
+   with a header comment: purpose, inputs, outputs, reproduction command.
+   Exploratory/buggy versions move to `_drafts/` or `failed/code-drafts/` — never deleted.
+2. No debug prints, unused variables, or hardcoded machine-specific paths
+   (the data-root constant is the only allowed absolute path).
+3. Figures in `plots/`, tables in `data/`; the deliverable at the run root.
+4. reproducibility.json is complete and matches the final scripts.
+5. `run-state.mjs validate <dir>` passes and every obligation has a terminal status.
+6. The cleanup itself is noted in the terminal transition reason.
+
+## Tidiness invariants
+
+- Code is refactored as soon as it stabilizes within an attempt (not only at the end).
+- Every script file maps 1:1 to an attempt artifact or a final deliverable.
+- Nothing outside the run directory is written by the modeling work.

package/skills/math-modeling-agent/references/subagent-dispatch.md

@@ -7,8 +7,9 @@ produce reports and evidence. Never let a subagent mutate run-state directly.
 
 | Class | Role (lineage) | When | Count |
 |---|---|---|---|
-| explorer | parallel candidate exploration (legacy Modeler) | CANDIDATES_READY, >= 2 independent directions | 1 per direction |
-| verifier | blind re-verification (legacy Analyzer/Correction) | before terminal: key numeric/feasibility claims | Standard: 1 per task; High-Assurance: 1 per claim |
+| explorer | parallel candidate exploration (original Modeler lineage) | CANDIDATES_READY, >= 2 independent directions | 1 per direction |
+| verifier | blind re-verification (original Analyzer lineage) | before terminal: key numeric/feasibility claims | Standard: 1 per task; High-Assurance: 1 per claim |
+| corrector | apply accepted critique and produce the next attempt (original Correction lineage) | after a failed/conditional verdict | 1 per selected direction |
 | auditor | artifact-only audit (math-modeling-audit skill) | before SOLVED | High-Assurance: mandatory; else optional |
 
 ## Brief template (self-contained; subagents see no parent context)

package/skills/math-modeling-agent/references/tool-policy.md

@@ -22,6 +22,17 @@ e.g. "本阶段用 Python+OSQP 做约束优化；Wolfram 对本题非必需，
 The problem never silently depends on a single tool: every phase names its
 backend, its fallback, and the degradation path when the backend fails.
 
+## Computation backend extension
+
+Before selecting an external computation backend, run
+`scripts/computation/probe_backends.py` (or its PowerShell compatibility entry)
+and use only the live capability result. Persist a backend snapshot with
+`scripts/computation/backend_inventory.py` and an explicit run-local
+`--state-file` when the phase needs a stable inventory. Use
+`scripts/computation/computation_record.py` for reproducible/file or
+high-assurance computation. A probe, inventory, or record validation does not
+prove a mathematical claim.
+
 ## Python
 
 Use `scripts/python-environment.mjs`. Prefer uv, then an existing Python runtime plus venv. Keep the environment under the run directory. Install only required normalized PyPI package names; VCS URLs, local wheels, or arbitrary indexes require explicit user approval. Record lock, seed, command, workdir, timeout, stdout/stderr, exit code, and generated files.

package/skills/math-modeling-agent/references/verification-recipes.md

@@ -0,0 +1,43 @@
+# Verification Recipes (v0.5)
+
+Recipes are selected by claim type before a phase can advance. A verification verdict is local to one recipe; it does not directly set an obligation or claim verdict.
+
+## Status layers
+
+- Verification: PASS, FAIL, INCONCLUSIVE, NOT_RUN.
+- Obligation: OPEN, SATISFIED, BLOCKED, WAIVED, STALE.
+- Claim: SUPPORTED, CONDITIONAL, INCONCLUSIVE, REFUTED, STALE.
+- Run: lifecycle state such as VERIFY, EVALUATE, REVISE, CONDITIONAL or SOLVED.
+
+## Recipes
+
+| Recipe | Required checks | Typical claim |
+|---|---|---|
+| data | schema, ids, references, hash, independent-parser | attachment/data interpretation |
+| numeric | recompute, domain, strict, tolerance, counterexample | numeric value or formula |
+| feasibility | all-constraints, actual-output, strict-and-tolerance | optimization/engineering feasibility |
+| optimization | objective-scope, coverage, initialization, local-global-label | best-found or optimum claim |
+| simulation | baseline, bounds, sampling, independent-implementation | simulation/ratio |
+| report | nine-sections, why-how-what-check, coverage, symbols, citations | paper/report completeness |
+| reproducibility | snapshot, command, environment, hash, clean-rerun | reproduction |
+| branch-coverage | candidate-enumeration, branch-switch | branch stability |
+
+## Independence rule
+
+Two checks count separately only when their calculation path or code provenance differs. Reusing a deterministic result, solver-internal result, branch summary, derived artifact, code path, or reprint does not create a second independent verification. The runner records an independenceKey and rejects duplicate or summary-derived keys.
+
+## Numeric and output rules
+
+A numeric recipe reports strict and tolerance outcomes separately. A display format is not numeric quantization. When an output is rounded, the quantized values are generated first and every original hard constraint is rerun on the quantized artifact.
+
+## Branch rules
+
+If a declared four-branch problem executes only one branch, the recipe returns or raises branch-switch-not-tested; branch_changed=false is not evidence of stability. Branch ranking from an existing solver summary is only a consistency check, not an independent rerun.
+
+## Benchmark rule
+
+An expensive recipe must carry a PASS benchmark artifact before full execution. The benchmark records command, input subset, expected time, environment, stdout/stderr and exit code. Missing or failed benchmark evidence blocks the expensive recipe.
+
+## Scope rule
+
+If a required check was not executed, the recipe is INCONCLUSIVE. If the user explicitly waives the check, the result is WAIVED and may only contribute to a CONDITIONAL run.

package/skills/math-modeling-agent/references/workflow.md

@@ -8,7 +8,7 @@
 
 ## States
 
-TRIAGE → SCOPE_FROZEN → INPUT_PROFILED → CLAIMS_REGISTERED → CANDIDATES_READY → ATTEMPT → EXECUTE → VERIFY.
+TRIAGE → RESEARCH → SCOPE_FROZEN → INPUT_PROFILED → CLAIMS_REGISTERED → CANDIDATES_READY → ATTEMPT → EXECUTE → VERIFY.
 
 VERIFY may move to REVISE, RESEARCH, FORK, an independent audit, or one terminal status:
 SOLVED, PARTIAL, CONDITIONAL, INCONCLUSIVE, REFUTED, INFEASIBLE, UNIDENTIFIABLE, BLOCKED, CANCELLED.
@@ -19,16 +19,17 @@ SOLVED, PARTIAL, CONDITIONAL, INCONCLUSIVE, REFUTED, INFEASIBLE, UNIDENTIFIABLE,
 - **逐问层**（SPi 循环）：子问题分析 → 子问题假设 Si（D2-SPi）→ ATTEMPT 建模求解 → VERIFY 该问验证 → 该问鲁棒性（扰动计划→交互→执行）→ D4 裁决 → REVISE（本问修订）或进入 SPi+1（reason: "SPi 完成，进入 SPi+1"）。
 - **收束层**：跨问联动检查（终态 gate 校验全部子问题 DONE + 依赖满足）→ 全题鲁棒性汇总 → 终态硬门 → SOLVED + final-report（整体叙事，非拼凑）。
 
-## Pre-modeling literature research (v2.6, mandatory)
+## Pre-modeling literature research (mandatory)
 
-CLAIMS_REGISTERED → RESEARCH → CANDIDATES_READY. Before any candidate is
-built, run the D-R exchange (interaction-protocol.md): search principles and
-methods by direction, write research/sources.jsonl, attach a literature column
-to the candidate matrix. Every direction must be traceable to at least one
-source — engineering background, methodology paper, textbook, or official
-documentation; a literature-free direction must be explicitly marked
+TRIAGE → RESEARCH → D-R → SCOPE_FROZEN. After the initial task triage and
+before D0 restatement is frozen, run the D-R exchange
+(interaction-protocol.md): search principles, engineering background, and
+methods by direction; write research/sources.jsonl; attach a literature
+column to the candidate matrix. Every direction must be traceable to at least
+one source. A literature-free direction must be explicitly marked
 "原创方向" with its derivation basis. Never search for solutions to the
-specific contest problem.
+specific contest problem. Mid-solve evidence gaps use a recorded
+RESEARCH backtrack, not an implicit bypass of the initial research phase.
 
 ## Interaction gates (v2)