@quietloudlab/ai-interaction-atlas 1.0.10 → 1.0.11

package/dist/data.js

@@ -962,6 +962,40 @@ var AI_TASKS = [
       { target_id: "task_estimate", type: "commonly_preceded_by", strength: "medium", reason: "Spatial estimation provides the world-model for planning." },
       { target_id: "task_act", type: "enables", strength: "strong", reason: "Plans are executed as sequences of actions." }
     ]
+  },
+  {
+    id: "task_harvest",
+    layer_id: "layer_internal",
+    name: "Harvest",
+    slug: "harvest",
+    task_type: "ai",
+    elevator_pitch: "Discover recurring patterns from aggregate human decisions over time.",
+    example_usage: "Identifying writing style patterns from content a user consistently accepts without editing.",
+    io_spec: {
+      inputs: {
+        required: [{ id: "data_log", label: "Behavioral Signal Log" }],
+        optional: [{ id: "data_text", label: "Source Content" }, { id: "data_json", label: "Existing Pattern Library" }]
+      },
+      outputs: {
+        primary: { id: "data_json", label: "Discovered Patterns", isArray: true },
+        metadata: [{ id: "data_score", label: "Pattern Confidence" }, { id: "data_text", label: "Source Provenance" }]
+      }
+    },
+    implementation_notes: { maturity: "emerging", typical_latency: "batch", data_requirements: "continuous", human_oversight: "recommended" },
+    ux_notes: { risk: "Hallucinated patterns that don't exist in source data", tip: "Every discovered pattern must link to specific source decisions with provenance \u2014 timestamp, actor, and original context \u2014 so patterns are grounded in evidence, not AI confabulation", anti_patterns: ["Harvesting from insufficient sample sizes", "Treating AI confidence as human validation", "No verification that discovered patterns exist in source material", "Patterns without provenance trail back to source decisions", "Accepting pattern candidates that cannot cite concrete examples from the behavioral log"] },
+    capabilities: [
+      { name: "Sentence Similarity", tag: "sentence-similarity", example: "Analyzing months of accepted blog posts to discover recurring phrase structures, vocabulary preferences, and formatting habits that define a writer's voice" },
+      { name: "Tabular Classification", tag: "tabular-classification", example: "Observing which product recommendations a shopper keeps versus removes over time to discover unstated preferences like avoiding synthetic fabrics" },
+      { name: "Feature Extraction", tag: "feature-extraction", example: "Mining a support team's ticket resolutions to discover that agents who include a one-line summary have 40% fewer re-opens \u2014 a pattern nobody explicitly taught" }
+    ],
+    relations: [
+      { target_id: "task_adapt", type: "enables", strength: "strong", reason: "Harvested patterns are the input that adaptation applies. Harvest discovers what to learn; adapt updates behavior." },
+      { target_id: "task_extract", type: "commonly_preceded_by", strength: "medium", reason: "Individual extraction events produce raw data that harvesting aggregates into patterns over time." },
+      { target_id: "human_edit", type: "requires_input_from", strength: "strong", reason: "Human edits \u2014 what they change and what they leave unchanged \u2014 are the primary behavioral signal." },
+      { target_id: "task_verify", type: "commonly_followed_by", strength: "strong", reason: "AI-discovered patterns are prone to hallucination; verify against source material before applying." },
+      { target_id: "human_review", type: "commonly_followed_by", strength: "medium", reason: "Low-confidence patterns should be staged for human review before affecting system behavior." },
+      { target_id: "task_cluster", type: "distinct_from", strength: "strong", reason: "Cluster groups items by similarity in a vector space at a single point in time; Harvest discovers patterns from sequential human decisions over time. Cluster is spatial (which items are near each other); Harvest is temporal (what behaviors recur across decisions). Cluster needs embeddings; Harvest needs behavioral logs with provenance." }
+    ]
   }
 ];