Phase 4 validation (planner + tasker) + deterministic dataset resolver

.specify/feature.json

@@ -1 +1 @@
-{"feature_directory": "specs/013-paper-revision-implementer"}
+{"feature_directory": "specs/014-phase4-plan-tasks-testing"}

CLAUDE.md

@@ -70,5 +70,5 @@ Since this is primarily a research documentation repository without traditional
 <!-- SPECKIT START -->
 For additional context about technologies to be used, project structure,
 shell commands, and other important information, read the current plan:
-[specs/013-paper-revision-implementer/plan.md](specs/013-paper-revision-implementer/plan.md).
+[specs/014-phase4-plan-tasks-testing/plan.md](specs/014-phase4-plan-tasks-testing/plan.md).
 <!-- SPECKIT END -->

agents/prompts/planner.md

@@ -57,19 +57,20 @@ $schema: ...
 - For computational projects, `contracts/` MUST include at least one
   schema (e.g., dataset schema, output schema) that the
   Implementer's tests can validate against.
-- NEVER invent URLs or citations. If the spec/idea has cited URLs,
-  copy them verbatim; do not add new ones, do not fabricate
-  `(verified YYYY-MM-DD)` annotations. The Reference-Validator
-  fetches every cited URL — fabricated URLs flip the verdict to
-  mismatch.
+- For dataset/code/paper references in research.md, cite ONLY the URLs listed in
+  the "# Verified datasets" block of the user message (these have been
+  web-searched and reachability/format-verified for you). NEVER invent or guess
+  a dataset URL. If the block says a dataset has NO verified source, describe the
+  dataset by name but do NOT fabricate a URL.
 - For DATASETS specifically: `research.md`'s "Dataset Strategy"
-  table MUST name only real, programmatically-fetchable sources.
-  If the spec calls for "UCI Electricity" but the canonical UCI
-  endpoint requires browser navigation, plan for the `ucimlrepo`
-  Python package OR substitute a comparable open dataset that has
-  a known-stable raw URL (e.g., NAB benchmark CSVs at
-  `https://raw.githubusercontent.com/numenta/NAB/master/data/realKnownCause/`,
-  or HuggingFace `datasets.load_dataset(...)`).
+  table MUST reference ONLY the sources in the "# Verified datasets"
+  block above — cite each dataset by its verified URL, or load that
+  SAME dataset via a well-known programmatic loader (e.g.
+  `datasets.load_dataset(...)` for a verified HuggingFace dataset, or
+  `ucimlrepo` for a UCI dataset). Do NOT substitute a different dataset
+  and do NOT invent or guess a raw URL. If a dataset the spec needs has
+  NO verified source in the block, state that explicitly rather than
+  fabricating one.
 - For COMPUTATIONAL TASK ORDERING: the plan MUST order phases so
   data is downloaded BEFORE any task that consumes it, models are
   fitted BEFORE any task that evaluates them, and figures are

agents/registry.yaml

@@ -29,7 +29,7 @@ agents:
   fallback_backends:
   - huggingface
   - local
-  default_model: google.gemma-3-27b-it
+  default_model: google.gemma-4-31B-it
   wall_clock_budget_seconds: 300
   paid_opt_in: false
 - name: flesh_out
@@ -218,7 +218,7 @@ agents:
   fallback_backends:
   - huggingface
   - local
-  default_model: google.gemma-3-27b-it
+  default_model: google.gemma-4-31B-it
   tools:
   - citation_fetcher
   wall_clock_budget_seconds: 300
@@ -316,7 +316,7 @@ agents:
   fallback_backends:
   - huggingface
   - local
-  default_model: google.gemma-3-27b-it
+  default_model: qwen.qwen3.5-122b
   wall_clock_budget_seconds: 300
   paid_opt_in: false
 - name: paper_writing
@@ -399,7 +399,7 @@ agents:
   fallback_backends:
   - huggingface
   - local
-  default_model: google.gemma-3-27b-it
+  default_model: google.gemma-4-31B-it
   wall_clock_budget_seconds: 600
   paid_opt_in: false
 - name: latex_fix
@@ -445,7 +445,7 @@ agents:
   fallback_backends:
   - huggingface
   - local
-  default_model: google.gemma-3-27b-it
+  default_model: google.gemma-4-31B-it
   wall_clock_budget_seconds: 300
   paid_opt_in: false
 - name: repository_hygiene
@@ -461,7 +461,7 @@ agents:
   fallback_backends:
   - huggingface
   - local
-  default_model: google.gemma-3-27b-it
+  default_model: google.gemma-4-31B-it
   wall_clock_budget_seconds: 300
   paid_opt_in: false
 - name: task_atomizer
@@ -496,7 +496,7 @@ agents:
   fallback_backends:
   - huggingface
   - local
-  default_model: google.gemma-3-27b-it
+  default_model: google.gemma-4-31B-it
   wall_clock_budget_seconds: 300
   paid_opt_in: false
 - name: paper_reviewer_writing_quality
@@ -818,7 +818,7 @@ agents:
   fallback_backends:
   - huggingface
   - local
-  default_model: google.gemma-3-27b-it
+  default_model: google.gemma-4-31B-it
   tools: []
   wall_clock_budget_seconds: 300
   paid_opt_in: false