@sanity/ailf-studio 1.1.0 → 1.1.1

package/dist/index.d.ts

@@ -614,21 +614,21 @@ declare function HelpDrawer(): react_jsx_runtime.JSX.Element | null;
  * @see docs/ARCHITECTURE.md (scoring model)
  */
 declare const GLOSSARY: {
-    readonly overallScore: "A weighted average across all feature areas: Task Completion (50%), Code Correctness (25%), and Doc Coverage (25%).";
-    readonly docLift: "How much the docs help, compared to the model's training data alone. This is the score with docs minus the score without. Higher is better.";
+    readonly overallScore: "A weighted average across all feature areas, using the gold scoring profile: Task Completion (50%), Code Correctness (25%), and Doc Coverage (25%).";
+    readonly docLift: "How much the docs help, compared to the model's training data alone. Calculated as ceiling minus floor, where ceiling includes Doc Coverage and floor does not. Higher is better.";
     readonly actualScore: "How well an AI agent scores when it has to find docs on its own through web search and page fetching. This is the real-world scenario. Only available in full mode.";
     readonly retrievalGap: "The score lost because agents can't find or use all the relevant docs. Calculated as ceiling minus actual. Lower is better; zero means agents find everything.";
     readonly infraEfficiency: "What percentage of the docs' potential quality actually reaches agents (actual ÷ ceiling). 100% means agents find and use all relevant docs perfectly.";
-    readonly floor: "Score without any documentation. This tells you what the model already knows from its training data.";
+    readonly floor: "Output-quality composite without documentation — Task Completion (60%) and Code Correctness (40%) only. Doc Coverage is excluded because it's undefined when no docs are provided. This tells you what the model already knows from its training data.";
     readonly ceiling: "Score with gold-standard docs injected directly into the prompt. This is the best the documentation can do.";
     readonly actual: "Score when an AI agent finds docs on its own through web search and page fetching. This is the real-world experience.";
     readonly retGap: "Quality lost to discoverability (ceiling minus actual). The gap between what the docs could deliver and what agents actually get.";
     readonly efficiency: "What fraction of the docs' quality reaches agents in practice (actual ÷ ceiling, shown as a percentage).";
     readonly invertedRetGap: "⚠️ Inverted retrieval gap: agents that can't find the docs actually score higher, because the docs hurt performance. This usually means there's a doc quality problem.";
-    readonly score: "Weighted score for this feature area: Task Completion × 50% + Code Correctness × 25% + Doc Coverage × 25%.";
+    readonly score: "Ceiling composite for this feature area: Task Completion × 50% + Code Correctness × 25% + Doc Coverage × 25%. The floor uses a different profile (Task × 60% + Code × 40%, no Doc Coverage).";
     readonly taskCompletion: "Can the LLM implement the requested feature? Graded 0–100.";
     readonly codeCorrectness: "Is the generated code idiomatic, correct, and following best practices? Graded 0–100.";
-    readonly docCoverage: "Did the docs provide the information needed to implement the feature? Graded 0–100.";
+    readonly docCoverage: "Did the docs provide the information needed to implement the feature? Graded 0–100. This dimension only contributes to the ceiling composite (with docs) — it's excluded from the floor composite because it's undefined without documentation.";
     readonly tests: "Number of test cases in this feature area.";
     readonly overallDelta: "Change in overall score between the two runs. Positive means the experiment scored higher.";
     readonly actualDelta: "Change in actual (agent-retrieved) score between runs. Positive means agents did better.";
@@ -674,7 +674,7 @@ declare const GLOSSARY: {
     readonly sourceProduction: "Production source — docs fetched from the live production dataset. Scores reflect what real users and AI agents experience today.";
     readonly sourceBranch: "Branch source — docs fetched from a branch or draft dataset. Use this to preview how content changes affect scores before publishing.";
     readonly sourceLocal: "Local source — docs fetched from local files or a local dev server. Useful for testing doc changes before pushing.";
-    readonly reportScore: "The overall weighted score for this evaluation run: Task Completion (50%), Code Correctness (25%), and Doc Coverage (25%), averaged across all feature areas.";
+    readonly reportScore: "The overall ceiling composite for this evaluation run: Task Completion (50%), Code Correctness (25%), and Doc Coverage (25%), averaged across all feature areas.";
     readonly reportMode: "The evaluation mode determines which reference points are measured. Different modes test different aspects of how AI agents interact with documentation.";
     readonly reportTrigger: "What initiated this evaluation run. Knowing the trigger helps you understand whether a score change was from a content edit, a code deploy, or a scheduled check.";
     readonly modeBaseline: "Baseline mode — tests the model with gold-standard docs injected directly. Measures ceiling performance (best the docs can do).";

package/dist/index.js

@@ -2585,8 +2585,8 @@ var taskSchema = defineType5({
           type: "boolean"
         }),
         defineField5({
-          description: 'Rubric mode for baseline. "abbreviated" uses a shorter rubric, "full" uses the same rubric as gold, "none" skips rubric grading.',
-          initialValue: "abbreviated",
+          description: 'Rubric mode for baseline. "full" uses the same rubric as gold, "abbreviated" uses a shorter rubric, "none" skips rubric grading.',
+          initialValue: "full",
           name: "rubric",
           options: {
             list: [
@@ -3305,7 +3305,7 @@ Click into any report for the full breakdown: per-area scores, diagnostics, and
   {
     "id": "scoring-model",
     "title": "Understanding Scores",
-    "body": "## The three dimensions\n\nEvery evaluation task is scored on three dimensions, each graded 0\u2013100:\n\n- **Task Completion (50% weight)** \u2014 Can the AI implement the requested feature?\n  Does the output actually do what was asked?\n- **Code Correctness (25% weight)** \u2014 Is the generated code idiomatic, correct,\n  and following best practices?\n- **Doc Coverage (25% weight)** \u2014 Did the documentation provide the information\n  needed to implement the feature?\n\n## How the overall score is calculated\n\nThe three dimensions combine into a single **AI Literacy Score** per task:\n\n```\nTotal = Task Completion \xD7 0.50 + Code Correctness \xD7 0.25 + Doc Coverage \xD7 0.25\n```\n\nThis weighted composite produces a score from 0\u2013100. Scores are then averaged\nacross all tasks in a feature area to produce a **per-area score**, and across\nall areas to produce the **overall score**.\n\n## What the numbers mean\n\n| Score range  | Interpretation                                                    |\n| ------------ | ----------------------------------------------------------------- |\n| **80\u2013100**   | Docs are working well \u2014 AI agents produce correct implementations |\n| **70\u201379**    | Needs attention \u2014 there may be gaps in specific dimensions        |\n| **Below 70** | Weak \u2014 AI agents consistently struggle with this area             |\n\n## Ceiling decomposition (baseline mode)\n\nWhen running in baseline mode, each task is evaluated twice \u2014 with and without\ndocumentation. This produces:\n\n- **Floor score** \u2014 Score without docs (what the model knows from training data\n  alone)\n- **Ceiling score** \u2014 Score with gold-standard docs injected directly into the\n  prompt\n- **Doc Lift** \u2014 Ceiling minus floor. Positive means docs help; negative means\n  docs hurt.\n- **Doc Quality Gap** \u2014 100 minus ceiling. Room for documentation improvement.\n\n## Three-layer decomposition (full mode)\n\nFull mode adds a third measurement \u2014 what happens when AI agents find docs on\ntheir own:\n\n- **Floor** \u2014 No docs (parametric knowledge only)\n- **Ceiling** \u2014 Gold-standard docs injected (best the docs can do)\n- **Actual** \u2014 Agent-retrieved docs (real-world performance)\n- **Retrieval Gap** \u2014 Ceiling minus actual (quality lost to findability)\n- **Infrastructure Efficiency** \u2014 Actual \xF7 ceiling (what fraction of doc quality\n  reaches agents)\n\n## Cost tracking\n\nEach evaluation also tracks token costs:\n\n- **Provider cost** \u2014 Token usage for generating implementations\n- **Grader cost** \u2014 Token usage for the grading model's assessments\n- **Total cost** \u2014 Both combined, reported in the score summary",
+    "body": "## The three dimensions\n\nEvery evaluation task is scored on three dimensions, each graded 0\u2013100:\n\n- **Task Completion (50% weight)** \u2014 Can the AI implement the requested feature?\n  Does the output actually do what was asked?\n- **Code Correctness (25% weight)** \u2014 Is the generated code idiomatic, correct,\n  and following best practices?\n- **Doc Coverage (25% weight)** \u2014 Did the documentation provide the information\n  needed to implement the feature?\n\n## How the overall score is calculated\n\nThe three dimensions combine into a single **AI Literacy Score** per task using\nnamed scoring profiles from `config/rubrics.yaml`:\n\n```\nGold (with docs):    Total = Task \xD7 0.50 + Code \xD7 0.25 + Docs \xD7 0.25\nBaseline (no docs):  Total = Task \xD7 0.60 + Code \xD7 0.40\n```\n\nThe gold profile includes all three dimensions. The baseline profile excludes\nDoc Coverage because it is undefined when no documentation is provided. This\nensures Doc Lift (ceiling \u2212 floor) is a clean structural measurement of\ndocumentation value.\n\nThe weighted composite produces a score from 0\u2013100. Scores are then averaged\nacross all tasks in a feature area to produce a **per-area score**, and across\nall areas to produce the **overall score**.\n\n## What the numbers mean\n\n| Score range  | Interpretation                                                    |\n| ------------ | ----------------------------------------------------------------- |\n| **80\u2013100**   | Docs are working well \u2014 AI agents produce correct implementations |\n| **70\u201379**    | Needs attention \u2014 there may be gaps in specific dimensions        |\n| **Below 70** | Weak \u2014 AI agents consistently struggle with this area             |\n\n## Ceiling decomposition (baseline mode)\n\nWhen running in baseline mode, each task is evaluated twice \u2014 with and without\ndocumentation. This produces:\n\n- **Floor score** \u2014 Score without docs (what the model knows from training data\n  alone)\n- **Ceiling score** \u2014 Score with gold-standard docs injected directly into the\n  prompt\n- **Doc Lift** \u2014 Ceiling minus floor. Positive means docs help; negative means\n  docs hurt.\n- **Doc Quality Gap** \u2014 100 minus ceiling. Room for documentation improvement.\n\n## Three-layer decomposition (full mode)\n\nFull mode adds a third measurement \u2014 what happens when AI agents find docs on\ntheir own:\n\n- **Floor** \u2014 No docs (parametric knowledge only)\n- **Ceiling** \u2014 Gold-standard docs injected (best the docs can do)\n- **Actual** \u2014 Agent-retrieved docs (real-world performance)\n- **Retrieval Gap** \u2014 Ceiling minus actual (quality lost to findability)\n- **Infrastructure Efficiency** \u2014 Actual \xF7 ceiling (what fraction of doc quality\n  reaches agents)\n\n## Cost tracking\n\nEach evaluation also tracks token costs:\n\n- **Provider cost** \u2014 Token usage for generating implementations\n- **Grader cost** \u2014 Token usage for the grading model's assessments\n- **Total cost** \u2014 Both combined, reported in the score summary",
     "source": "docs/help/scoring-model.md",
     "related": [
       "three-layer",
@@ -3337,7 +3337,7 @@ Click into any report for the full breakdown: per-area scores, diagnostics, and
   {
     "id": "glossary",
     "title": "Glossary",
-    "body": "**Overall Score**\n: A weighted average across all feature areas: Task Completion (50%), Code Correctness (25%), and Doc Coverage (25%).\n\n**Doc Lift**\n: How much the docs help, compared to the model's training data alone. This is the score with docs minus the score without. Higher is better.\n\n**Actual Score**\n: How well an AI agent scores when it has to find docs on its own through web search and page fetching. This is the real-world scenario. Only available in full mode.\n\n**Retrieval Gap**\n: The score lost because agents can't find or use all the relevant docs. Calculated as ceiling minus actual. Lower is better; zero means agents find everything.\n\n**Infra Efficiency**\n: What percentage of the docs' potential quality actually reaches agents (actual \xF7 ceiling). 100% means agents find and use all relevant docs perfectly.\n\n**Floor**\n: Score without any documentation. This tells you what the model already knows from its training data.\n\n**Ceiling**\n: Score with gold-standard docs injected directly into the prompt. This is the best the documentation can do.\n\n**Actual**\n: Score when an AI agent finds docs on its own through web search and page fetching. This is the real-world experience.\n\n**Ret Gap**\n: Quality lost to discoverability (ceiling minus actual). The gap between what the docs could deliver and what agents actually get.\n\n**Efficiency**\n: What fraction of the docs' quality reaches agents in practice (actual \xF7 ceiling, shown as a percentage).\n\n**Inverted Ret Gap**\n: \u26A0\uFE0F Inverted retrieval gap: agents that can't find the docs actually score higher, because the docs hurt performance. This usually means there's a doc quality problem.\n\n**Score**\n: Weighted score for this feature area: Task Completion \xD7 50% + Code Correctness \xD7 25% + Doc Coverage \xD7 25%.\n\n**Task Completion**\n: Can the LLM implement the requested feature? Graded 0\u2013100.\n\n**Code Correctness**\n: Is the generated code idiomatic, correct, and following best practices? Graded 0\u2013100.\n\n**Doc Coverage**\n: Did the docs provide the information needed to implement the feature? Graded 0\u2013100.\n\n**Tests**\n: Number of test cases in this feature area.\n\n**Overall Delta**\n: Change in overall score between the two runs. Positive means the experiment scored higher.\n\n**Actual Delta**\n: Change in actual (agent-retrieved) score between runs. Positive means agents did better.\n\n**Ret Gap Delta**\n: Change in retrieval gap between runs. Negative is good here: it means the gap shrank and agents found more relevant docs.\n\n**Efficiency Delta**\n: Change in infrastructure efficiency between runs. Positive means agents are capturing more of the docs' potential.\n\n**Baseline**\n: The reference run you're comparing against.\n\n**Experiment**\n: The new run you're evaluating.\n\n**Delta**\n: Difference between experiment and baseline. Positive means improvement, negative means regression.\n\n**Change**\n: Whether the change is meaningful: improved, regressed, or unchanged (within the noise threshold).\n\n**Low Scoring Judgments**\n: The grading model's explanations for tests that scored below 70/100.\n\n**Judgment Reason**\n: The grading model's natural language explanation of what went wrong.\n\n**Health Strong**\n: Feature areas scoring 80 or above. The docs are working well for these features \u2014 AI agents produce correct, complete implementations.\n\n**Health Attention**\n: Feature areas scoring 70\u201379. These are okay but could be improved \u2014 there may be gaps in specific dimensions like doc coverage or code correctness.\n\n**Health Weak**\n: Feature areas scoring below 70. The docs are not providing enough support for AI agents to implement these features correctly.\n\n**Negative Doc Lift Metric**\n: Number of areas where the documentation actually hurts AI performance \u2014 the model scores higher without docs than with them. This usually means the docs contain outdated patterns or incorrect examples.\n\n**Weak Areas**\n: Feature areas where the overall score is below 70. These need the most attention \u2014 low scores mean AI agents consistently struggle to implement these features.\n\n**Docs Hurt**\n: Areas where the floor score (no docs) is higher than the ceiling score (with docs). The documentation is actively misleading the model. These docs need to be rewritten or removed.\n\n**Retrieval Issues**\n: Areas where AI agents can find less than 70% of the available doc quality. The docs exist and are good, but agents can't discover them through search. Consider improving page titles, metadata, or search engine indexing.\n\n**Dim Weaknesses**\n: Individual grading dimensions scoring below 50 within an area. These are the specific skills where AI agents fail most \u2014 task completion (can it build the feature?), code correctness (is the code right?), or doc coverage (did it use the docs?).\n\n**Efficiency Anomalies**\n: Areas where agent efficiency exceeds 100% \u2014 meaning agents perform better with self-found docs than with gold-standard docs injected directly. This can indicate doc quality issues (injected docs confuse the model) or agent memorization.\n\n**Doc Lift Wins**\n: Areas where documentation boosts AI performance by 5 or more points. Higher doc lift means the docs are providing crucial information that the model doesn't already know.\n\n**Retrieval Excellence**\n: Areas where AI agents successfully find and use at least 85% of the available doc quality through web search. Good retrieval means your docs are well-indexed and easy for agents to discover.\n\n**Model Breakdown**\n: Break down scores by individual LLM model. The default 'All Models' view shows the cross-model average. Select a specific model to see how it performed independently \u2014 useful for spotting models that struggle with specific feature areas.\n\n**Strengths**\n: What's working well: high-scoring areas, dimensions where the docs are strong, and areas where AI agents successfully find and use the documentation.\n\n**Recommendations**\n: Prioritized remediation plan from gap analysis. Each recommendation identifies a documentation problem, the affected feature area, and the estimated score lift from fixing it.\n\n**Total Potential Lift**\n: Aggregate potential score lift if all identified gaps were fixed. This is a conservative estimate \u2014 each gap targets the median of non-bottlenecked dimensions, not 100.\n\n**Failure Mode**\n: The type of documentation problem: missing-docs (functionality not covered), incorrect-docs (factual errors), outdated-docs (stale API/patterns), or poor-structure (hard to find/understand).\n\n**Estimated Lift**\n: Estimated composite score improvement if this gap is fully fixed. Based on raising bottleneck dimensions to the median of non-bottlenecked dimensions.\n\n**Confidence**\n: How confident the classifier is in this diagnosis. High = strong keyword + structural signal agreement. Medium = partial agreement. Low = weak signals only.\n\n**Agent Behavior Overview**\n: How AI agents interacted with your documentation during evaluation: what they searched for, which pages they visited, and how much time they spent on network requests.\n\n**Search Queries**\n: The exact search queries agents used to find documentation. Helps you understand how agents discover your content and whether your docs appear for relevant queries.\n\n**Doc Slugs Visited**\n: Documentation page slugs that agents actually visited during evaluation. Compare against canonical docs to see if agents found the right pages.\n\n**External Domains**\n: Non-Sanity domains that agents contacted during evaluation. High external domain counts may indicate agents couldn't find what they needed in your docs.\n\n**Avg Doc Pages Visited**\n: Average number of documentation pages visited per test. Higher counts can mean agents need to consult many pages (complex task) or can't find the right one quickly.\n\n**Avg Searches Performed**\n: Average number of web searches performed per test. High search counts can indicate docs are hard to discover through search engines.\n\n**Avg Network Time Ms**\n: Average time spent on network requests per test. Includes page fetches, search queries, and API calls.\n\n**Total Requests**\n: Total number of HTTP requests the agent made during the test, including searches, page visits, and API calls.\n\n**Total Bytes Downloaded**\n: Total bytes downloaded by the agent. Large downloads may indicate the agent is fetching many pages or very large documents.\n\n**Dim Task Completion**\n: Change in task completion between runs. Positive means implementations are more complete.\n\n**Dim Code Correctness**\n: Change in code correctness between runs. Positive means better code quality.\n\n**Dim Doc Coverage**\n: Change in doc coverage between runs. Positive means the docs are providing more useful information.\n\n**Area Delta**\n: Score change for this area compared to the previous evaluation run.\n\n**Source Production**\n: Production source \u2014 docs fetched from the live production dataset. Scores reflect what real users and AI agents experience today.\n\n**Source Branch**\n: Branch source \u2014 docs fetched from a branch or draft dataset. Use this to preview how content changes affect scores before publishing.\n\n**Source Local**\n: Local source \u2014 docs fetched from local files or a local dev server. Useful for testing doc changes before pushing.\n\n**Report Score**\n: The overall weighted score for this evaluation run: Task Completion (50%), Code Correctness (25%), and Doc Coverage (25%), averaged across all feature areas.\n\n**Report Mode**\n: The evaluation mode determines which reference points are measured. Different modes test different aspects of how AI agents interact with documentation.\n\n**Report Trigger**\n: What initiated this evaluation run. Knowing the trigger helps you understand whether a score change was from a content edit, a code deploy, or a scheduled check.\n\n**Mode Baseline**\n: Baseline mode \u2014 tests the model with gold-standard docs injected directly. Measures ceiling performance (best the docs can do).\n\n**Mode Full**\n: Full mode \u2014 runs baseline + agentic. Compares ceiling (injected docs) against actual (agent-retrieved docs) to measure retrieval gap and infrastructure efficiency.\n\n**Mode Agentic**\n: Agentic mode \u2014 the AI agent finds docs on its own via web search. Measures real-world performance: can agents actually discover and use your documentation?\n\n**Mode Observed**\n: Observed mode \u2014 records how agents interact with docs without scoring. Captures search queries, pages visited, and browsing patterns for analysis.\n\n**Mode Debug**\n: Debug mode \u2014 a diagnostic run for pipeline development. May use non-standard configurations or limited task sets.\n\n**Trigger Manual**\n: Manually triggered \u2014 someone ran the evaluation pipeline by hand, either locally or via the Studio UI.\n\n**Trigger Ci**\n: CI-triggered \u2014 the evaluation ran automatically as part of a pull request or merge pipeline.\n\n**Trigger Schedule**\n: Scheduled \u2014 the evaluation ran on a recurring schedule (e.g. nightly or weekly) to track score trends over time.\n\n**Trigger Webhook**\n: Webhook-triggered \u2014 a content change in Sanity triggered the evaluation automatically. Helps catch doc regressions early.\n\n**Trigger Cross Repo**\n: Cross-repo \u2014 triggered from another repository via the dispatch API. Used when external repos want to validate their docs against AILF tasks.",
+    "body": "**Overall Score**\n: A weighted average across all feature areas, using the gold scoring profile: Task Completion (50%), Code Correctness (25%), and Doc Coverage (25%).\n\n**Doc Lift**\n: How much the docs help, compared to the model's training data alone. Calculated as ceiling minus floor, where ceiling includes Doc Coverage and floor does not. Higher is better.\n\n**Actual Score**\n: How well an AI agent scores when it has to find docs on its own through web search and page fetching. This is the real-world scenario. Only available in full mode.\n\n**Retrieval Gap**\n: The score lost because agents can't find or use all the relevant docs. Calculated as ceiling minus actual. Lower is better; zero means agents find everything.\n\n**Infra Efficiency**\n: What percentage of the docs' potential quality actually reaches agents (actual \xF7 ceiling). 100% means agents find and use all relevant docs perfectly.\n\n**Floor**\n: Output-quality composite without documentation \u2014 Task Completion (60%) and Code Correctness (40%) only. Doc Coverage is excluded because it's undefined when no docs are provided. This tells you what the model already knows from its training data.\n\n**Ceiling**\n: Score with gold-standard docs injected directly into the prompt. This is the best the documentation can do.\n\n**Actual**\n: Score when an AI agent finds docs on its own through web search and page fetching. This is the real-world experience.\n\n**Ret Gap**\n: Quality lost to discoverability (ceiling minus actual). The gap between what the docs could deliver and what agents actually get.\n\n**Efficiency**\n: What fraction of the docs' quality reaches agents in practice (actual \xF7 ceiling, shown as a percentage).\n\n**Inverted Ret Gap**\n: \u26A0\uFE0F Inverted retrieval gap: agents that can't find the docs actually score higher, because the docs hurt performance. This usually means there's a doc quality problem.\n\n**Score**\n: Ceiling composite for this feature area: Task Completion \xD7 50% + Code Correctness \xD7 25% + Doc Coverage \xD7 25%. The floor uses a different profile (Task \xD7 60% + Code \xD7 40%, no Doc Coverage).\n\n**Task Completion**\n: Can the LLM implement the requested feature? Graded 0\u2013100.\n\n**Code Correctness**\n: Is the generated code idiomatic, correct, and following best practices? Graded 0\u2013100.\n\n**Doc Coverage**\n: Did the docs provide the information needed to implement the feature? Graded 0\u2013100. This dimension only contributes to the ceiling composite (with docs) \u2014 it's excluded from the floor composite because it's undefined without documentation.\n\n**Tests**\n: Number of test cases in this feature area.\n\n**Overall Delta**\n: Change in overall score between the two runs. Positive means the experiment scored higher.\n\n**Actual Delta**\n: Change in actual (agent-retrieved) score between runs. Positive means agents did better.\n\n**Ret Gap Delta**\n: Change in retrieval gap between runs. Negative is good here: it means the gap shrank and agents found more relevant docs.\n\n**Efficiency Delta**\n: Change in infrastructure efficiency between runs. Positive means agents are capturing more of the docs' potential.\n\n**Baseline**\n: The reference run you're comparing against.\n\n**Experiment**\n: The new run you're evaluating.\n\n**Delta**\n: Difference between experiment and baseline. Positive means improvement, negative means regression.\n\n**Change**\n: Whether the change is meaningful: improved, regressed, or unchanged (within the noise threshold).\n\n**Low Scoring Judgments**\n: The grading model's explanations for tests that scored below 70/100.\n\n**Judgment Reason**\n: The grading model's natural language explanation of what went wrong.\n\n**Health Strong**\n: Feature areas scoring 80 or above. The docs are working well for these features \u2014 AI agents produce correct, complete implementations.\n\n**Health Attention**\n: Feature areas scoring 70\u201379. These are okay but could be improved \u2014 there may be gaps in specific dimensions like doc coverage or code correctness.\n\n**Health Weak**\n: Feature areas scoring below 70. The docs are not providing enough support for AI agents to implement these features correctly.\n\n**Negative Doc Lift Metric**\n: Number of areas where the documentation actually hurts AI performance \u2014 the model scores higher without docs than with them. This usually means the docs contain outdated patterns or incorrect examples.\n\n**Weak Areas**\n: Feature areas where the overall score is below 70. These need the most attention \u2014 low scores mean AI agents consistently struggle to implement these features.\n\n**Docs Hurt**\n: Areas where the floor score (no docs) is higher than the ceiling score (with docs). The documentation is actively misleading the model. These docs need to be rewritten or removed.\n\n**Retrieval Issues**\n: Areas where AI agents can find less than 70% of the available doc quality. The docs exist and are good, but agents can't discover them through search. Consider improving page titles, metadata, or search engine indexing.\n\n**Dim Weaknesses**\n: Individual grading dimensions scoring below 50 within an area. These are the specific skills where AI agents fail most \u2014 task completion (can it build the feature?), code correctness (is the code right?), or doc coverage (did it use the docs?).\n\n**Efficiency Anomalies**\n: Areas where agent efficiency exceeds 100% \u2014 meaning agents perform better with self-found docs than with gold-standard docs injected directly. This can indicate doc quality issues (injected docs confuse the model) or agent memorization.\n\n**Doc Lift Wins**\n: Areas where documentation boosts AI performance by 5 or more points. Higher doc lift means the docs are providing crucial information that the model doesn't already know.\n\n**Retrieval Excellence**\n: Areas where AI agents successfully find and use at least 85% of the available doc quality through web search. Good retrieval means your docs are well-indexed and easy for agents to discover.\n\n**Model Breakdown**\n: Break down scores by individual LLM model. The default 'All Models' view shows the cross-model average. Select a specific model to see how it performed independently \u2014 useful for spotting models that struggle with specific feature areas.\n\n**Strengths**\n: What's working well: high-scoring areas, dimensions where the docs are strong, and areas where AI agents successfully find and use the documentation.\n\n**Recommendations**\n: Prioritized remediation plan from gap analysis. Each recommendation identifies a documentation problem, the affected feature area, and the estimated score lift from fixing it.\n\n**Total Potential Lift**\n: Aggregate potential score lift if all identified gaps were fixed. This is a conservative estimate \u2014 each gap targets the median of non-bottlenecked dimensions, not 100.\n\n**Failure Mode**\n: The type of documentation problem: missing-docs (functionality not covered), incorrect-docs (factual errors), outdated-docs (stale API/patterns), or poor-structure (hard to find/understand).\n\n**Estimated Lift**\n: Estimated composite score improvement if this gap is fully fixed. Based on raising bottleneck dimensions to the median of non-bottlenecked dimensions.\n\n**Confidence**\n: How confident the classifier is in this diagnosis. High = strong keyword + structural signal agreement. Medium = partial agreement. Low = weak signals only.\n\n**Agent Behavior Overview**\n: How AI agents interacted with your documentation during evaluation: what they searched for, which pages they visited, and how much time they spent on network requests.\n\n**Search Queries**\n: The exact search queries agents used to find documentation. Helps you understand how agents discover your content and whether your docs appear for relevant queries.\n\n**Doc Slugs Visited**\n: Documentation page slugs that agents actually visited during evaluation. Compare against canonical docs to see if agents found the right pages.\n\n**External Domains**\n: Non-Sanity domains that agents contacted during evaluation. High external domain counts may indicate agents couldn't find what they needed in your docs.\n\n**Avg Doc Pages Visited**\n: Average number of documentation pages visited per test. Higher counts can mean agents need to consult many pages (complex task) or can't find the right one quickly.\n\n**Avg Searches Performed**\n: Average number of web searches performed per test. High search counts can indicate docs are hard to discover through search engines.\n\n**Avg Network Time Ms**\n: Average time spent on network requests per test. Includes page fetches, search queries, and API calls.\n\n**Total Requests**\n: Total number of HTTP requests the agent made during the test, including searches, page visits, and API calls.\n\n**Total Bytes Downloaded**\n: Total bytes downloaded by the agent. Large downloads may indicate the agent is fetching many pages or very large documents.\n\n**Dim Task Completion**\n: Change in task completion between runs. Positive means implementations are more complete.\n\n**Dim Code Correctness**\n: Change in code correctness between runs. Positive means better code quality.\n\n**Dim Doc Coverage**\n: Change in doc coverage between runs. Positive means the docs are providing more useful information.\n\n**Area Delta**\n: Score change for this area compared to the previous evaluation run.\n\n**Source Production**\n: Production source \u2014 docs fetched from the live production dataset. Scores reflect what real users and AI agents experience today.\n\n**Source Branch**\n: Branch source \u2014 docs fetched from a branch or draft dataset. Use this to preview how content changes affect scores before publishing.\n\n**Source Local**\n: Local source \u2014 docs fetched from local files or a local dev server. Useful for testing doc changes before pushing.\n\n**Report Score**\n: The overall ceiling composite for this evaluation run: Task Completion (50%), Code Correctness (25%), and Doc Coverage (25%), averaged across all feature areas.\n\n**Report Mode**\n: The evaluation mode determines which reference points are measured. Different modes test different aspects of how AI agents interact with documentation.\n\n**Report Trigger**\n: What initiated this evaluation run. Knowing the trigger helps you understand whether a score change was from a content edit, a code deploy, or a scheduled check.\n\n**Mode Baseline**\n: Baseline mode \u2014 tests the model with gold-standard docs injected directly. Measures ceiling performance (best the docs can do).\n\n**Mode Full**\n: Full mode \u2014 runs baseline + agentic. Compares ceiling (injected docs) against actual (agent-retrieved docs) to measure retrieval gap and infrastructure efficiency.\n\n**Mode Agentic**\n: Agentic mode \u2014 the AI agent finds docs on its own via web search. Measures real-world performance: can agents actually discover and use your documentation?\n\n**Mode Observed**\n: Observed mode \u2014 records how agents interact with docs without scoring. Captures search queries, pages visited, and browsing patterns for analysis.\n\n**Mode Debug**\n: Debug mode \u2014 a diagnostic run for pipeline development. May use non-standard configurations or limited task sets.\n\n**Trigger Manual**\n: Manually triggered \u2014 someone ran the evaluation pipeline by hand, either locally or via the Studio UI.\n\n**Trigger Ci**\n: CI-triggered \u2014 the evaluation ran automatically as part of a pull request or merge pipeline.\n\n**Trigger Schedule**\n: Scheduled \u2014 the evaluation ran on a recurring schedule (e.g. nightly or weekly) to track score trends over time.\n\n**Trigger Webhook**\n: Webhook-triggered \u2014 a content change in Sanity triggered the evaluation automatically. Helps catch doc regressions early.\n\n**Trigger Cross Repo**\n: Cross-repo \u2014 triggered from another repository via the dispatch API. Used when external repos want to validate their docs against AILF tasks.",
     "source": "packages/studio/src/glossary.ts",
     "tags": [
       "reference",
@@ -3386,23 +3386,23 @@ import { useClient as useClient3 } from "sanity";
 // src/glossary.ts
 var GLOSSARY = {
   // -- Overview stats -------------------------------------------------------
-  overallScore: "A weighted average across all feature areas: Task Completion (50%), Code Correctness (25%), and Doc Coverage (25%).",
-  docLift: "How much the docs help, compared to the model's training data alone. This is the score with docs minus the score without. Higher is better.",
+  overallScore: "A weighted average across all feature areas, using the gold scoring profile: Task Completion (50%), Code Correctness (25%), and Doc Coverage (25%).",
+  docLift: "How much the docs help, compared to the model's training data alone. Calculated as ceiling minus floor, where ceiling includes Doc Coverage and floor does not. Higher is better.",
   actualScore: "How well an AI agent scores when it has to find docs on its own through web search and page fetching. This is the real-world scenario. Only available in full mode.",
   retrievalGap: "The score lost because agents can't find or use all the relevant docs. Calculated as ceiling minus actual. Lower is better; zero means agents find everything.",
   infraEfficiency: "What percentage of the docs' potential quality actually reaches agents (actual \xF7 ceiling). 100% means agents find and use all relevant docs perfectly.",
   // -- Three-layer decomposition columns ------------------------------------
-  floor: "Score without any documentation. This tells you what the model already knows from its training data.",
+  floor: "Output-quality composite without documentation \u2014 Task Completion (60%) and Code Correctness (40%) only. Doc Coverage is excluded because it's undefined when no docs are provided. This tells you what the model already knows from its training data.",
   ceiling: "Score with gold-standard docs injected directly into the prompt. This is the best the documentation can do.",
   actual: "Score when an AI agent finds docs on its own through web search and page fetching. This is the real-world experience.",
   retGap: "Quality lost to discoverability (ceiling minus actual). The gap between what the docs could deliver and what agents actually get.",
   efficiency: "What fraction of the docs' quality reaches agents in practice (actual \xF7 ceiling, shown as a percentage).",
   invertedRetGap: "\u26A0\uFE0F Inverted retrieval gap: agents that can't find the docs actually score higher, because the docs hurt performance. This usually means there's a doc quality problem.",
   // -- Per-area score columns -----------------------------------------------
-  score: "Weighted score for this feature area: Task Completion \xD7 50% + Code Correctness \xD7 25% + Doc Coverage \xD7 25%.",
+  score: "Ceiling composite for this feature area: Task Completion \xD7 50% + Code Correctness \xD7 25% + Doc Coverage \xD7 25%. The floor uses a different profile (Task \xD7 60% + Code \xD7 40%, no Doc Coverage).",
   taskCompletion: "Can the LLM implement the requested feature? Graded 0\u2013100.",
   codeCorrectness: "Is the generated code idiomatic, correct, and following best practices? Graded 0\u2013100.",
-  docCoverage: "Did the docs provide the information needed to implement the feature? Graded 0\u2013100.",
+  docCoverage: "Did the docs provide the information needed to implement the feature? Graded 0\u2013100. This dimension only contributes to the ceiling composite (with docs) \u2014 it's excluded from the floor composite because it's undefined without documentation.",
   tests: "Number of test cases in this feature area.",
   // -- Comparison deltas ----------------------------------------------------
   overallDelta: "Change in overall score between the two runs. Positive means the experiment scored higher.",
@@ -3460,7 +3460,7 @@ var GLOSSARY = {
   sourceBranch: "Branch source \u2014 docs fetched from a branch or draft dataset. Use this to preview how content changes affect scores before publishing.",
   sourceLocal: "Local source \u2014 docs fetched from local files or a local dev server. Useful for testing doc changes before pushing.",
   // -- Report list columns ----------------------------------------------------
-  reportScore: "The overall weighted score for this evaluation run: Task Completion (50%), Code Correctness (25%), and Doc Coverage (25%), averaged across all feature areas.",
+  reportScore: "The overall ceiling composite for this evaluation run: Task Completion (50%), Code Correctness (25%), and Doc Coverage (25%), averaged across all feature areas.",
   reportMode: "The evaluation mode determines which reference points are measured. Different modes test different aspects of how AI agents interact with documentation.",
   reportTrigger: "What initiated this evaluation run. Knowing the trigger helps you understand whether a score change was from a content edit, a code deploy, or a scheduled check.",
   // -- Mode values -----------------------------------------------------------

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sanity/ailf-studio",
-  "version": "1.1.0",
+  "version": "1.1.1",
   "description": "AI Literacy Framework — Sanity Studio dashboard plugin",
   "type": "module",
   "license": "MIT",