Merge pull request #25 from tylerbessire/V2

V2

Author: tylerbessire

.github/pull_request_template.md

@@ -0,0 +1,13 @@
+## Summary
+- [ ] What changed?
+- [ ] Why?
+
+## Testing
+- [ ] Steps
+- [ ] Results
+
+## Risk Assessment
+- [ ] Regressions considered
+- [ ] Rollback plan
+
+[S:PR v1] template=installed pass

arc_solver/solver.py

@@ -19,6 +19,11 @@
     synthesize as synth_baseline,
     predict_two as predict_two_baseline,
 )
+from puma.rft.feature_flags import RFT_ENABLED, build_limits
+from puma.rft.orchestrator import solve_with_rft
+from puma.rft.tracking import HypothesisMemory
+from puma.rft.demo.grid import build_context as build_rft_grid_context
+from puma.rft.explain import explain_last_run
 from .enhanced_search import synthesize_with_enhancements, predict_two_enhanced
 from .hypothesis import HypothesisEngine, Hypothesis
 from .continuous_learning import ContinuousSelfMemory
@@ -79,6 +84,10 @@ def __init__(self, use_enhancements: bool = True,
         self._placeholder_templates: List[PlaceholderTemplate] = []
         self._new_placeholder_templates: List[PlaceholderTemplate] = []
         self._last_hypotheses: List[Hypothesis] = []
+        self.rft_enabled = RFT_ENABLED
+        self._rft_limits = build_limits()
+        self._rft_memory = HypothesisMemory()
+        self._last_rft_status: Optional[str] = None
 
     def solve_task(self, task: Dict[str, List[Dict[str, List[List[int]]]]]) -> Dict[str, List[List[List[int]]]]:
         """Solve a single ARC task using enhanced or baseline methods."""
@@ -177,6 +186,9 @@ def solve_task(self, task: Dict[str, List[Dict[str, List[List[int]]]]]) -> Dict[
         self._record_continuous_experience(task_id, train_pairs, best_hypothesis, solved_training, result)
         if solved_training:
             self.stats['tasks_solved'] += 1
+        if self.rft_enabled:
+            attempt1, attempt2 = self._maybe_run_rft_adapter(task, attempt1, attempt2)
+            result = {"attempt_1": attempt1, "attempt_2": attempt2}
         return result
 
     def _get_predictions(
@@ -216,6 +228,50 @@ def _get_predictions(
             self.logger.info("Using baseline prediction")
         return baseline
 
+    def _maybe_run_rft_adapter(
+        self,
+        task: Dict[str, Any],
+        attempt1: List[List[List[int]]],
+        attempt2: List[List[List[int]]],
+    ) -> Tuple[List[List[List[int]]], List[List[List[int]]]]:
+        """Optionally run the RFT orchestrator when the demo flag is present."""
+
+        metadata = task.get("metadata") or {}
+        rft_config = metadata.get("rft_demo")
+        if not rft_config or "grid" not in rft_config:
+            return attempt1, attempt2
+        grid_spec = rft_config["grid"]
+        grid = grid_spec.get("grid")
+        targets = grid_spec.get("targets", [])
+        if not isinstance(grid, list) or not grid:
+            return attempt1, attempt2
+        try:
+            context, rules = build_rft_grid_context(grid, targets, limits=self._rft_limits)
+        except Exception:
+            return attempt1, attempt2
+        context, status = solve_with_rft(
+            context,
+            rules,
+            self._rft_memory,
+            outer_budget=self._rft_limits.outer_budget,
+        )
+        self._last_rft_status = status
+        if self.enable_logging:
+            self.logger.debug("RFT status=%s\n%s", status, explain_last_run(context))
+        if status != "DONE" or not grid_spec.get("inject_output", False):
+            return attempt1, attempt2
+        solved_grid = context.state.grid
+        if solved_grid and isinstance(solved_grid[0][0], str):
+            palette: Dict[str, int] = {"black": 0}
+            next_id = 1
+            for row in solved_grid:
+                for color in row:
+                    if color not in palette:
+                        palette[color] = next_id
+                        next_id += 1
+            solved_grid = [[palette[color] for color in row] for row in solved_grid]
+        return [solved_grid for _ in attempt1], [solved_grid for _ in attempt2]
+
     def _postprocess_predictions(
         self,
         train_pairs: List[Tuple[Array, Array]],

puma/__init__.py

@@ -0,0 +1,10 @@
+"""PUMA meta-learning components."""
+from importlib import import_module
+
+__all__ = ["rft"]
+
+
+def __getattr__(name):
+    if name == "rft":
+        return import_module("puma.rft")
+    raise AttributeError(name)

puma/rft/README.md

@@ -0,0 +1,104 @@
+# PUMA RFT Module
+
+[S:DOC v2] scope=puma.rft overview=complete pass
+
+The Relational Frame Theory (RFT) module adds pliance-based rule following and
+tracking-driven rule variation to the PUMA solver while remaining fully
+feature-flagged behind `PUMA_RFT`.
+
+## Architecture
+
+```
++-------------------+      +----------------------+      +------------------+
+| feature_flags     | ---> | pliance (determinism)| ---> | explain          |
++-------------------+      +----------------------+      +------------------+
+           |                         |                               ^
+           v                         v                               |
++-------------------+      +----------------------+      +------------------+
+| demo/grid harness | ---> | tracking (variation) | ---> | orchestrator     |
++-------------------+      +----------------------+      +------------------+
+```
+
+* `types.py` exposes typed primitives (rules, context, hypotheses).
+* `pliance.py` executes prioritized rules deterministically.
+* `tracking.py` evaluates adaptive hypotheses and promotes successful ones.
+* `orchestrator.py` alternates pliance and tracking with metrics and guards.
+* `explain.py` records rich traces surfaced via `explain_last_run()`.
+
+## Usage
+
+1. Enable the feature flag: `export PUMA_RFT=1`.
+2. Run the demo tests: `pytest tests/rft/test_rft_integration.py -q`.
+3. Inspect the most recent trace:
+
+   ```python
+   from puma.rft.explain import explain_last_run
+   print(explain_last_run())
+   ```
+
+To exercise the adapter inside the ARC solver:
+
+```bash
+PUMA_RFT=1 python -c 'from arc_solver import ARCSolver; solver = ARCSolver();
+print(solver.solve_task({"train": [...], "test": [...], "metadata":
+{"rft_demo": {"grid": {"grid": [[0,1,0]], "targets": [[0,0],[0,1],[0,2]],
+"inject_output": true}}}}))'
+```
+
+### Demo promotion proof
+
+Copy-paste the snippet below to verify a promotion and trace:
+
+```bash
+PUMA_RFT=1 python - <<'PY'
+from puma.rft.demo.grid import build_context
+from puma.rft.orchestrator import solve_with_rft
+from puma.rft.tracking import HypothesisMemory
+from puma.rft.feature_flags import build_limits
+from puma.rft.explain import explain_last_run
+
+grid = [
+    [0, 1],
+    [2, 1],
+    [2, 1],
+]
+targets = [(r, 0) for r in range(len(grid))]
+context, rules = build_context(grid, targets, build_limits())
+memory = HypothesisMemory()
+context, status = solve_with_rft(context, rules, memory, outer_budget=4)
+print(status)
+print(explain_last_run(context))
+PY
+```
+
+Expected output (abridged):
+
+```
+DONE
+Status: GOAL
+Top hypotheses:
+  - propagate 0 by (1,0) regardless of color (score=1.00)
+Promoted rules:
+  - track_1_0_broad (score=1.00)
+Metrics: ... tracking_promotions=1 ...
+```
+
+## Observability
+
+Structured debug logs are emitted under the `puma.rft.*` namespace. Metrics are
+collected on the context (e.g., `pliance_rule_fires`, `tracking_hypotheses_evaluated`)
+and surfaced via `explain_last_run`.
+
+## Tracking score
+
+Hypothesis tracking uses the explicit scoring rule described in the spec:
+
+```
+progress = satisfied_targets / total_targets
+simplicity = len(active_conditions)
+score = progress - 0.1 * simplicity
+```
+
+Scores greater than the configured threshold (`RFT_DEFAULT_LIMITS["THRESH"]`) are
+promoted. Simplicity penalties ensure compact rules are favoured when progress is
+equivalent.

puma/rft/__init__.py

@@ -0,0 +1,37 @@
+"""Relational Frame Theory inspired solving primitives for PUMA."""
+# [S:DESIGN v1] package=rft_core pass
+
+from .feature_flags import RFT_ENABLED, RFT_DEFAULT_LIMITS, build_limits
+from .types import Entity, Relation, Rule, Context, Hypothesis, Limits
+from .pliance import pliance_solve
+from .tracking import (
+    generate_variations,
+    evaluate_hypothesis,
+    promote_to_rule,
+    HypothesisMemory,
+    tracking_step,
+)
+from .orchestrator import solve_with_rft, insert_after_stuck_rule
+from .explain import explain_last_run, record_trace
+
+__all__ = [
+    "RFT_ENABLED",
+    "RFT_DEFAULT_LIMITS",
+    "build_limits",
+    "Entity",
+    "Relation",
+    "Rule",
+    "Context",
+    "Hypothesis",
+    "Limits",
+    "pliance_solve",
+    "generate_variations",
+    "evaluate_hypothesis",
+    "promote_to_rule",
+    "HypothesisMemory",
+    "tracking_step",
+    "solve_with_rft",
+    "insert_after_stuck_rule",
+    "explain_last_run",
+    "record_trace",
+]