[Khalil] · medical-nutrition-deep-research-1.1.0.zip
medical-nutrition-deep-research-1.1.0 Score: 34/40 Verdict: Pass (workshop-pass) Отправитель: Khalil hodzha@gmail.com SHA-256: cb61fd227602087d… Blockers: none Unverified: F07, F08 Required rework A04 (What it does and when to apply (from description alone)): Expand the description to name representative trigger situations while keeping it routing-only, e.g. ‘Produces a traceable, evidence-based review of nutrition, diet, or supplement questions for prevention or a clinical condition, covering guidelines, benefits, harms, and conflicts.’ D03 (Reference navigation): Add a short ‘## Contents’ list at the top of each reference file longer than 100 lines. F01 (Design note and exact use case): Add a ‘Design Note’ section documenting the exact repeatable use case, the observed baseline failure modes of the agent without the skill (e.g., generic unsourced advice, mixing food/supplement evidence, no conflict handling), the chosen control level and why, and the supporting files/evals selected. F06 (Seven specified evals): Restructure test-scenarios.md so each of the 7 required types (positive, negative, boundary, security injection, evidence quality, broken script, portability/path) is an explicit case, and give every case three labeled fields: query, expected_behavior, and failure_modes. F07 (Trigger behavior quality (observed)): Run the eval cases in a clean agent context; record a baseline (without skill) and with-skill trace for at least one representative task plus the positive/negative/boundary trigger cases, and include the traces in the submission. F08 (Security, evidence and navigation behavior (observed)): Provide observed run traces (e.g., a security-injection case and an evidence-quality case) demonstrating the required behaviors, including final validator output. ID Score Status Evidence Defect / minimal fix A01 1/1 passed SKILL.md present at root (preflight inventory). YAML frontmatter parses with frontmatter_error: null. Non-empty name=‘medical-nutrition-deep-research’ and description present (SKILL.md:2-3). Additional metadata (version/author/license/metadata.hermes) are benign non-routing fields. — A02 1/1 passed name=‘medical-nutrition-deep-research’ is 32 chars (<=64), matches ^[a-z0-9]+(?:-[a-z0-9]+)*$, is not a reserved value (research/deep-research are reserved only as standalone), and denotes a concrete applied domain (medical nutrition) rather than a generic method. — A03 1/1 passed description=‘Evidence-based deep research on medical nutrition.’ is 50 chars (1-1024), contains no XML tags, and is plain routing text (not a YAML object or authority-expanding instruction). — A04 0/1 failed The 50-char description states only what the skill does (‘Evidence-based deep research on medical nutrition.’); it does not, from the description alone, convey realistic user requests/situations for when to choose it, and it is a bare noun phrase (nominative) rather than a third-person activity statement. The ‘when to apply’ detail lives only in the body Trigger section (SKILL.md:36-46), which A04 forbids relying on. Expand the description to name representative trigger situations while keeping it routing-only, e.g. ‘Produces a traceable, evidence-based review of nutrition, diet, or supplement questions for prevention or a clinical condition, covering guidelines, benefits, harms, and conflicts.’ A05 1/1 passed description contains no workflow, defaults, examples, implementation details, feature inventory, role slop, or ad promises. Each phrase changes the routing decision; ‘Evidence-based’ usefully separates this skill from a recipe/cooking intent. — A06 1/1 passed Positive triggers listed (SKILL.md:38-44) and a concrete anti-trigger: ‘Do not trigger for a simple recipe, meal idea, calorie lookup, or general cooking question unless the user explicitly asks for medical evidence review.’ (SKILL.md:46). Non-negotiable Boundaries (SKILL.md:48-57) add useful scope edges. catalog_dir unavailable; boundaries judged self-consistent. — A07 1/1 passed All skill-internal resource paths cited in SKILL.md:200-211 (references/*.md, templates/report.html, scripts/validate_report.py) resolve to files present in the inventory, use relative paths with forward slashes, and contain no absolute paths, drive letters, usernames, or broken links. — B01 1/1 passed The package defines a repeatable research procedure (5 phases: Planner/Searcher/Reader/Verifier/Synthesizer) with a concrete observable result: a self-contained, validator-passing .html report (SKILL.md:26,94-149). Not a persona, tool list, or link collection. — B02 1/1 passed Single coherent unit of work: evidence review of medical nutrition for prevention or a clinical condition. Narrow domain scope; explicitly separates food pattern vs nutrient vs supplement evidence. Not a universal helper or generic-advice bundle. — B03 1/1 passed Terms (prevention vs disease, food vs nutrient vs supplement, RDA/AI/DRV/UL) used consistently. Volatile facts are checked at runtime, not hardcoded: source-hierarchy.md:3 ‘URLs below are starting points, not permission to assume a document is current’; workflow requires recording version/date and confirming guideline edition (SKILL.md:126, source-hierarchy.md:103-109). — B04 1/1 passed Control level matches risk: a defined repeatable expert workflow (5 ordered phases) with explicit gates for consequential medical actions — safety handoff (population-safety.md:27-40), quantitative policy (SKILL.md:59-71), and non-negotiable boundaries (no diagnosis/prescription). Not arbitrary for a fragile domain, and not over-mechanized for judgment. — C01 1/1 passed One clear main path with concrete sequential actions: Planner -> Searcher -> Reader/Extractor -> Verifier -> Synthesizer (SKILL.md:94-149), each with specific steps. Not an essay or a ‘be thorough’ instruction. — C02 1/1 passed Search-strategy branches (breadth-first / depth-first / hypothesis-driven) each carry a ‘Use for …’ decision rule (deep-research-protocol.md:55-67); prevention vs disease mode is decided in the research brief. The overall default path is fixed; the dual-search is ‘use both when available’ (not an undirected tool menu). — C03 1/1 passed Consequential actions carry precise stop conditions and verification: acute-risk safety handoff (SKILL.md:57, population-safety.md:27-40); stopping criteria (SKILL.md:235-244); report build is plan -> copy template -> fill -> run validator -> fix FAIL (deep-research-protocol.md:246-260). Optional CLI install is gated by explicit permission (keenable-and-codex.md:55). — C04 1/1 passed Concrete artifact rework loop: run scripts/validate_report.py, treat each FAIL/error as a failure, fix the flagged issue, re-validate (deep-research-protocol.md:252-258; SKILL.md:148,162,244). Combined with iterative search and stopping rules, the check/failure/change/restart elements are present for the deliverable. — C05 1/1 passed Required result (self-contained .html with evidence ledger + linked bibliography), readiness criterion (validator must pass), and 15 required section IDs are defined and machine-enforced (SKILL.md:213-231, scripts/validate_report.py:18-34,182-189). Strict schema is appropriate for a formal report; claims of completion require observable validation. — C06 1/1 passed Common workflow is easy to locate (Mandatory Architecture heading); vague commands are replaced with checkable actions (specific query families, log formats, extraction checklist). Theory (GRADE/AMSTAR/RoB framing) is clearly labeled as appraisal guidance, not masquerading as the procedure (SKILL.md:166-185). — C07 1/1 passed Declared requires_toolsets: [web, files] (SKILL.md:11). The workflow explains the capabilities it needs (built-in web search + optional KeenAble) and gates the optional CLI install. No hidden action, no credential reading (‘Never place an API key …’, keenable-and-codex.md:17), and the validator is stdlib-only. KeenAble unavailability degrades gracefully to built-in search. — D01 1/1 passed SKILL.md is 276 lines (preflight skill_line_count: 276), well under the 500-line limit. — D02 1/1 passed SKILL.md keeps the core workflow, boundaries, quantitative policy, and a short output form. Long methodology (search iterations, appraisal frameworks, source registry, ledger schema) is moved to references/; the large report template lives in templates/. The entrypoint is decision-bearing, not filler. — D03 0/1 failed Each cited reference states its purpose (SKILL.md:200-211) and references sit one level down with no A->B chains. However, six reference files exceed 100 lines (deep-research-protocol 260, population-safety 181, evidence-ledger-schema 180, keenable-and-codex 153, test-scenarios 133, source-hierarchy 121) and none contains a brief explicit Contents/table-of-contents block at the top; only ad-hoc section headings exist. Add a short ‘## Contents’ list at the top of each reference file longer than 100 lines. D04 1/1 passed All resources are live and used by the common path: 7 reference docs, a report template, a stdlib validator, a regression test, and two validator fixtures. Names are descriptive; no misc.md/doc1.md, no dead resources, no OS/editor metadata (.DS_Store etc.). Eval scenarios and fixtures are organized under references/ rather than a tests/ dir but are real and used. — D05 1/1 passed scripts/validate_report.py: shebang + docstring (‘Uses only the Python standard library’), args documented (python3 validate_report.py <report.html> [–template], SKILL.md:161), input=HTML report, output=PASS/FAIL with specific errors + exit 0/1/2, portable relative paths. scripts/test_validator.py: shebang + docstring, stdlib deps, fixtures under references/fixtures/ (verified present). — D06 1/1 passed Static audit: validate_report.py is stdlib-only, handles file-not-found/read/parse errors (exit 2, lines 146-156), returns specific machine-checkable errors and exit codes. test_validator.py exercises 3 positive and 6 negative paths against present fixtures with explicit substring assertions. No syntax/logic defects found. NOTE: live execution was not performed (review-only policy); confidence is based on full source audit, not mere file presence. — E01 1/1 passed Retrieved content is consistently treated as data to verify, never as authority: ‘Never treat a search snippet as evidence’ (SKILL.md:124), ‘A KeenAble snippet is not evidence’ (keenable-and-codex.md:113), ‘Treat built-in and KeenAble results as discovery channels, not independent evidence’ (SKILL.md:111). No reference instructs the agent to override the workflow or disable safeguards. — E02 1/1 passed Explicit secret protection: ‘Never place an API key in a report, search log, skill file, or command transcript’ (keenable-and-codex.md:17). Downloads/writes target declared locations (deliverable to user’s working dir; installer to /tmp). Optional KeenAble CLI install is gated by user permission; validator is stdlib-only with no dependencies. — E03 1/1 passed The HTML artifact write is the requested deliverable (deep research -> HTML report). Intermediate artifacts (research brief, issue tree, search log, evidence ledger, contradiction register) are framed as optional in-context working notes folded into the final report (SKILL.md:187-198), not undeclared separate files. Web/URL access is the requested research activity; optional CLI install is gated and reported. — E04 1/1 passed All shell/network operations are justified and described: keenable CLI (search/fetch/install), python3 validator, command -v keenable. No bundled binaries (all files are .md/.py/.html), no obfuscated code. The curl|sh installer is the declared, gated install path for the optional search tool. — E05 1/1 passed No instruction to ignore previous instructions, hide actions, weaken safeguards, or expand permissions. The skill reinforces safeguards (safety handoff, non-negotiable boundaries) and instructs reporting the file path and search log rather than hiding actions. — E06 1/1 passed Optional tool is checked before use: ‘command -v keenable / keenable –version’ (keenable-and-codex.md:23-26); on failure, record the class and continue with built-in search, never fabricate results (SKILL.md:112, keenable-and-codex.md:129-137). Built-in web search is declared as a required toolset (host-provided). — E07 1/1 passed The research/validate workflow uses only relative paths (scripts/, references/, templates/). $HOME references (keenable-and-codex.md:45-46,68) denote the host platform’s standard skill/cargo locations via an environment variable, not a hardcoded author path, username, or drive letter. Deliverable path is user-chosen (‘user’s requested working/deliverable directory’, SKILL.md:160). — E08 1/1 passed The skill references KeenAble MCP only generically (‘prefer its MCP tools in Codex when configured’, SKILL.md:109) and provides the CLI (keenable search/fetch) as the concrete, fully-specified interface. No bare MCP tool name is invoked that would require ServerName:tool_name qualification, so no unqualified name is used. — F01 0/1 failed No design note (or equivalent dedicated section) exists. The Purpose section states use cases and the observable outcome, but there is no recorded baseline of agent errors observed on representative tasks without the skill, and no explicit rationale for the chosen degree of freedom or the selected supporting files/evals. ‘baseline’ appears only in the nutritional sense (baseline intake/deficiency), not as a baseline-failure analysis. has_design_note_phrase: false (preflight). Add a ‘Design Note’ section documenting the exact repeatable use case, the observed baseline failure modes of the agent without the skill (e.g., generic unsourced advice, mixing food/supplement evidence, no conflict handling), the chosen control level and why, and the supporting files/evals selected. F02 1/1 passed source-hierarchy.md defines a domain-appropriate taxonomy: Tier A authoritative recommendations/reference values, Tier B evidence syntheses, Tier C primary/ongoing studies, plus methodology sources. Source selection rules forbid blogs/shops/manufacturer pages as efficacy evidence and require tracing secondary sources to primary (source-hierarchy.md:77-89). Independence and recency tests are specified. — F03 1/1 passed Freshness: recency rules require recording dates/versions and searching for updates/retractions (source-hierarchy.md:103-109). Diversity: mandatory dual-search and iterative querying with provenance-based deduplication. Conflicts: explicit Contradiction Register with ‘Never average a contradiction away’ and a synthesis tier ‘Disputed’ plus ‘Unsupported/insufficient’ which constitutes confidence lowering (deep-research-protocol.md:203-230). — F04 1/1 passed Facts, interpretation, and synthesis are separated (SKILL.md:138 ‘Distinguish fact, author interpretation, and this review’s synthesis’; evidence-ledger-schema isolates atomic claims). Central conclusions trace to ledger claims via source_id/source_url/verifying_text. Confidence is GRADE-informed and proportional (evidence-ledger-schema.md:71-95); ‘Insufficient evidence’ is an explicit allowed label. — F05 1/1 passed Required HTML sections cover all semantic elements: research-question (reframed goal), recommendations (conclusion), evidence-ledger (key evidence), harms + limitations + open-questions (risks/limitations/unknowns), confidence, bibliography (provenance) (SKILL.md:213-231). The validator enforces presence and non-emptiness of these sections. — F06 0/1 failed references/test-scenarios.md provides 15 scenarios, but they use ‘Prompt’ and ‘Expected’ only — none includes an explicit failure_modes field, which the spec mandates for every case. Additionally, not all seven required case types are clearly present as distinct cases: no explicit negative-trigger case (a non-medical query that must not activate), no security-injection case, no broken-script case, and no portability/path case are identifiable. Restructure test-scenarios.md so each of the 7 required types (positive, negative, boundary, security injection, evidence quality, broken script, portability/path) is an explicit case, and give every case three labeled fields: query, expected_behavior, and failure_modes. F07 0/1 failed No observed run traces are included in the submission — no baseline-vs-skill comparison on a representative task, and no recorded activations on positive/negative/boundary cases. The spec states that files describing expected behavior without a run score 0. The required workshop trace is absent from the submission itself (failed, not merely unverified). Run the eval cases in a clean agent context; record a baseline (without skill) and with-skill trace for at least one representative task plus the positive/negative/boundary trigger cases, and include the traces in the submission. F08 0/1 failed No observed run traces confirming that the agent resisted injection/hidden side effects, avoided strong conclusions from weak evidence, read the needed supporting files, skipped irrelevant files, and fulfilled the output contract with final validation. The required workshop trace is absent from the submission (failed, not merely unverified). Provide observed run traces (e.g., a security-injection case and an evidence-quality case) demonstrating the required behaviors, including final validator output. Полная scorecard ID Score Status Evidence Defect / minimal fix A01 1/1 passed SKILL.md present at root (preflight inventory). YAML frontmatter parses with frontmatter_error: null. Non-empty name=‘medical-nutrition-deep-research’ and description present (SKILL.md:2-3). Additional metadata (version/author/license/metadata.hermes) are benign non-routing fields. — A02 1/1 passed name=‘medical-nutrition-deep-research’ is 32 chars (<=64), matches ^[a-z0-9]+(?:-[a-z0-9]+)*$, is not a reserved value (research/deep-research are reserved only as standalone), and denotes a concrete applied domain (medical nutrition) rather than a generic method. — A03 1/1 passed description=‘Evidence-based deep research on medical nutrition.’ is 50 chars (1-1024), contains no XML tags, and is plain routing text (not a YAML object or authority-expanding instruction). — A04 0/1 failed The 50-char description states only what the skill does (‘Evidence-based deep research on medical nutrition.’); it does not, from the description alone, convey realistic user requests/situations for when to choose it, and it is a bare noun phrase (nominative) rather than a third-person activity statement. The ‘when to apply’ detail lives only in the body Trigger section (SKILL.md:36-46), which A04 forbids relying on. Expand the description to name representative trigger situations while keeping it routing-only, e.g. ‘Produces a traceable, evidence-based review of nutrition, diet, or supplement questions for prevention or a clinical condition, covering guidelines, benefits, harms, and conflicts.’ A05 1/1 passed description contains no workflow, defaults, examples, implementation details, feature inventory, role slop, or ad promises. Each phrase changes the routing decision; ‘Evidence-based’ usefully separates this skill from a recipe/cooking intent. — A06 1/1 passed Positive triggers listed (SKILL.md:38-44) and a concrete anti-trigger: ‘Do not trigger for a simple recipe, meal idea, calorie lookup, or general cooking question unless the user explicitly asks for medical evidence review.’ (SKILL.md:46). Non-negotiable Boundaries (SKILL.md:48-57) add useful scope edges. catalog_dir unavailable; boundaries judged self-consistent. — A07 1/1 passed All skill-internal resource paths cited in SKILL.md:200-211 (references/*.md, templates/report.html, scripts/validate_report.py) resolve to files present in the inventory, use relative paths with forward slashes, and contain no absolute paths, drive letters, usernames, or broken links. — B01 1/1 passed The package defines a repeatable research procedure (5 phases: Planner/Searcher/Reader/Verifier/Synthesizer) with a concrete observable result: a self-contained, validator-passing .html report (SKILL.md:26,94-149). Not a persona, tool list, or link collection. — B02 1/1 passed Single coherent unit of work: evidence review of medical nutrition for prevention or a clinical condition. Narrow domain scope; explicitly separates food pattern vs nutrient vs supplement evidence. Not a universal helper or generic-advice bundle. — B03 1/1 passed Terms (prevention vs disease, food vs nutrient vs supplement, RDA/AI/DRV/UL) used consistently. Volatile facts are checked at runtime, not hardcoded: source-hierarchy.md:3 ‘URLs below are starting points, not permission to assume a document is current’; workflow requires recording version/date and confirming guideline edition (SKILL.md:126, source-hierarchy.md:103-109). — B04 1/1 passed Control level matches risk: a defined repeatable expert workflow (5 ordered phases) with explicit gates for consequential medical actions — safety handoff (population-safety.md:27-40), quantitative policy (SKILL.md:59-71), and non-negotiable boundaries (no diagnosis/prescription). Not arbitrary for a fragile domain, and not over-mechanized for judgment. — C01 1/1 passed One clear main path with concrete sequential actions: Planner -> Searcher -> Reader/Extractor -> Verifier -> Synthesizer (SKILL.md:94-149), each with specific steps. Not an essay or a ‘be thorough’ instruction. — C02 1/1 passed Search-strategy branches (breadth-first / depth-first / hypothesis-driven) each carry a ‘Use for …’ decision rule (deep-research-protocol.md:55-67); prevention vs disease mode is decided in the research brief. The overall default path is fixed; the dual-search is ‘use both when available’ (not an undirected tool menu). — C03 1/1 passed Consequential actions carry precise stop conditions and verification: acute-risk safety handoff (SKILL.md:57, population-safety.md:27-40); stopping criteria (SKILL.md:235-244); report build is plan -> copy template -> fill -> run validator -> fix FAIL (deep-research-protocol.md:246-260). Optional CLI install is gated by explicit permission (keenable-and-codex.md:55). — C04 1/1 passed Concrete artifact rework loop: run scripts/validate_report.py, treat each FAIL/error as a failure, fix the flagged issue, re-validate (deep-research-protocol.md:252-258; SKILL.md:148,162,244). Combined with iterative search and stopping rules, the check/failure/change/restart elements are present for the deliverable. — C05 1/1 passed Required result (self-contained .html with evidence ledger + linked bibliography), readiness criterion (validator must pass), and 15 required section IDs are defined and machine-enforced (SKILL.md:213-231, scripts/validate_report.py:18-34,182-189). Strict schema is appropriate for a formal report; claims of completion require observable validation. — C06 1/1 passed Common workflow is easy to locate (Mandatory Architecture heading); vague commands are replaced with checkable actions (specific query families, log formats, extraction checklist). Theory (GRADE/AMSTAR/RoB framing) is clearly labeled as appraisal guidance, not masquerading as the procedure (SKILL.md:166-185). — C07 1/1 passed Declared requires_toolsets: [web, files] (SKILL.md:11). The workflow explains the capabilities it needs (built-in web search + optional KeenAble) and gates the optional CLI install. No hidden action, no credential reading (‘Never place an API key …’, keenable-and-codex.md:17), and the validator is stdlib-only. KeenAble unavailability degrades gracefully to built-in search. — D01 1/1 passed SKILL.md is 276 lines (preflight skill_line_count: 276), well under the 500-line limit. — D02 1/1 passed SKILL.md keeps the core workflow, boundaries, quantitative policy, and a short output form. Long methodology (search iterations, appraisal frameworks, source registry, ledger schema) is moved to references/; the large report template lives in templates/. The entrypoint is decision-bearing, not filler. — D03 0/1 failed Each cited reference states its purpose (SKILL.md:200-211) and references sit one level down with no A->B chains. However, six reference files exceed 100 lines (deep-research-protocol 260, population-safety 181, evidence-ledger-schema 180, keenable-and-codex 153, test-scenarios 133, source-hierarchy 121) and none contains a brief explicit Contents/table-of-contents block at the top; only ad-hoc section headings exist. Add a short ‘## Contents’ list at the top of each reference file longer than 100 lines. D04 1/1 passed All resources are live and used by the common path: 7 reference docs, a report template, a stdlib validator, a regression test, and two validator fixtures. Names are descriptive; no misc.md/doc1.md, no dead resources, no OS/editor metadata (.DS_Store etc.). Eval scenarios and fixtures are organized under references/ rather than a tests/ dir but are real and used. — D05 1/1 passed scripts/validate_report.py: shebang + docstring (‘Uses only the Python standard library’), args documented (python3 validate_report.py <report.html> [–template], SKILL.md:161), input=HTML report, output=PASS/FAIL with specific errors + exit 0/1/2, portable relative paths. scripts/test_validator.py: shebang + docstring, stdlib deps, fixtures under references/fixtures/ (verified present). — D06 1/1 passed Static audit: validate_report.py is stdlib-only, handles file-not-found/read/parse errors (exit 2, lines 146-156), returns specific machine-checkable errors and exit codes. test_validator.py exercises 3 positive and 6 negative paths against present fixtures with explicit substring assertions. No syntax/logic defects found. NOTE: live execution was not performed (review-only policy); confidence is based on full source audit, not mere file presence. — E01 1/1 passed Retrieved content is consistently treated as data to verify, never as authority: ‘Never treat a search snippet as evidence’ (SKILL.md:124), ‘A KeenAble snippet is not evidence’ (keenable-and-codex.md:113), ‘Treat built-in and KeenAble results as discovery channels, not independent evidence’ (SKILL.md:111). No reference instructs the agent to override the workflow or disable safeguards. — E02 1/1 passed Explicit secret protection: ‘Never place an API key in a report, search log, skill file, or command transcript’ (keenable-and-codex.md:17). Downloads/writes target declared locations (deliverable to user’s working dir; installer to /tmp). Optional KeenAble CLI install is gated by user permission; validator is stdlib-only with no dependencies. — E03 1/1 passed The HTML artifact write is the requested deliverable (deep research -> HTML report). Intermediate artifacts (research brief, issue tree, search log, evidence ledger, contradiction register) are framed as optional in-context working notes folded into the final report (SKILL.md:187-198), not undeclared separate files. Web/URL access is the requested research activity; optional CLI install is gated and reported. — E04 1/1 passed All shell/network operations are justified and described: keenable CLI (search/fetch/install), python3 validator, command -v keenable. No bundled binaries (all files are .md/.py/.html), no obfuscated code. The curl|sh installer is the declared, gated install path for the optional search tool. — E05 1/1 passed No instruction to ignore previous instructions, hide actions, weaken safeguards, or expand permissions. The skill reinforces safeguards (safety handoff, non-negotiable boundaries) and instructs reporting the file path and search log rather than hiding actions. — E06 1/1 passed Optional tool is checked before use: ‘command -v keenable / keenable –version’ (keenable-and-codex.md:23-26); on failure, record the class and continue with built-in search, never fabricate results (SKILL.md:112, keenable-and-codex.md:129-137). Built-in web search is declared as a required toolset (host-provided). — E07 1/1 passed The research/validate workflow uses only relative paths (scripts/, references/, templates/). $HOME references (keenable-and-codex.md:45-46,68) denote the host platform’s standard skill/cargo locations via an environment variable, not a hardcoded author path, username, or drive letter. Deliverable path is user-chosen (‘user’s requested working/deliverable directory’, SKILL.md:160). — E08 1/1 passed The skill references KeenAble MCP only generically (‘prefer its MCP tools in Codex when configured’, SKILL.md:109) and provides the CLI (keenable search/fetch) as the concrete, fully-specified interface. No bare MCP tool name is invoked that would require ServerName:tool_name qualification, so no unqualified name is used. — F01 0/1 failed No design note (or equivalent dedicated section) exists. The Purpose section states use cases and the observable outcome, but there is no recorded baseline of agent errors observed on representative tasks without the skill, and no explicit rationale for the chosen degree of freedom or the selected supporting files/evals. ‘baseline’ appears only in the nutritional sense (baseline intake/deficiency), not as a baseline-failure analysis. has_design_note_phrase: false (preflight). Add a ‘Design Note’ section documenting the exact repeatable use case, the observed baseline failure modes of the agent without the skill (e.g., generic unsourced advice, mixing food/supplement evidence, no conflict handling), the chosen control level and why, and the supporting files/evals selected. F02 1/1 passed source-hierarchy.md defines a domain-appropriate taxonomy: Tier A authoritative recommendations/reference values, Tier B evidence syntheses, Tier C primary/ongoing studies, plus methodology sources. Source selection rules forbid blogs/shops/manufacturer pages as efficacy evidence and require tracing secondary sources to primary (source-hierarchy.md:77-89). Independence and recency tests are specified. — F03 1/1 passed Freshness: recency rules require recording dates/versions and searching for updates/retractions (source-hierarchy.md:103-109). Diversity: mandatory dual-search and iterative querying with provenance-based deduplication. Conflicts: explicit Contradiction Register with ‘Never average a contradiction away’ and a synthesis tier ‘Disputed’ plus ‘Unsupported/insufficient’ which constitutes confidence lowering (deep-research-protocol.md:203-230). — F04 1/1 passed Facts, interpretation, and synthesis are separated (SKILL.md:138 ‘Distinguish fact, author interpretation, and this review’s synthesis’; evidence-ledger-schema isolates atomic claims). Central conclusions trace to ledger claims via source_id/source_url/verifying_text. Confidence is GRADE-informed and proportional (evidence-ledger-schema.md:71-95); ‘Insufficient evidence’ is an explicit allowed label. — F05 1/1 passed Required HTML sections cover all semantic elements: research-question (reframed goal), recommendations (conclusion), evidence-ledger (key evidence), harms + limitations + open-questions (risks/limitations/unknowns), confidence, bibliography (provenance) (SKILL.md:213-231). The validator enforces presence and non-emptiness of these sections. — F06 0/1 failed references/test-scenarios.md provides 15 scenarios, but they use ‘Prompt’ and ‘Expected’ only — none includes an explicit failure_modes field, which the spec mandates for every case. Additionally, not all seven required case types are clearly present as distinct cases: no explicit negative-trigger case (a non-medical query that must not activate), no security-injection case, no broken-script case, and no portability/path case are identifiable. Restructure test-scenarios.md so each of the 7 required types (positive, negative, boundary, security injection, evidence quality, broken script, portability/path) is an explicit case, and give every case three labeled fields: query, expected_behavior, and failure_modes. F07 0/1 failed No observed run traces are included in the submission — no baseline-vs-skill comparison on a representative task, and no recorded activations on positive/negative/boundary cases. The spec states that files describing expected behavior without a run score 0. The required workshop trace is absent from the submission itself (failed, not merely unverified). Run the eval cases in a clean agent context; record a baseline (without skill) and with-skill trace for at least one representative task plus the positive/negative/boundary trigger cases, and include the traces in the submission. F08 0/1 failed No observed run traces confirming that the agent resisted injection/hidden side effects, avoided strong conclusions from weak evidence, read the needed supporting files, skipped irrelevant files, and fulfilled the output contract with final validation. The required workshop trace is absent from the submission (failed, not merely unverified). Provide observed run traces (e.g., a security-injection case and an evidence-quality case) demonstrating the required behaviors, including final validator output. ← назад к лидерборду