feat(partition): per-block diagnostic floats + payoff measurement

CHANGELOG.md

@@ -2,7 +2,34 @@
 
 ## Unreleased
 
-(nothing yet)
+### Added
+
+- **Per-block diagnostic floats on `BlockReport`** (closes the Phase 2
+  v1 deferral). `rank_ratio`, `post_A`, and `informative_metric` now
+  populate per block; `forge_mse` stays caller-provided, same as the
+  top-level CompressionReport's. Empty-block degenerate case leaves
+  diagnostics as `None`. 2 new tests; full suite 1079 → 1081.
+
+### Documented (scope honesty)
+
+- **`add-encoding-partition` is polygram-only.** An A/B experiment on
+  a real SAE (jbloom GPT-2 first 64 features) with sae-forge's full
+  `ForgePipeline.run_synthetic` measured: substrate cost reduces
+  **8-16×** under partition (validated), but `forge_faithfulness_kl`
+  is **identical** between uniform and partitioned compressions
+  (10.0484 in both runs). The polygram encoding-family choice does
+  not propagate through sae-forge's current forge path — sae-forge
+  reads W_dec from the safetensors and the partition doesn't change
+  W_dec. The proposal's projected forge-side payoff (10-30% KL lift)
+  is **unproven** and requires separate sae-forge work that is not
+  specified anywhere today.
+  - Artefacts: `runs/real_partition_experiment.py` (reproducible
+    against a fresh venv with the jbloom SAE cached) +
+    `runs/real_partition_experiment.json` (measurement output).
+  - This is a useful negative result, not a failure of the feature.
+    The partition still works as a polygram-Dictionary substrate cost
+    reducer + analyst diagnostic tool. Just don't expect the forge
+    to honour the encoding choice today.
 
 ## 0.14.0 — 2026-05-21
 

polygram/compression/compressor.py

@@ -1231,10 +1231,13 @@ def _build_block_report(
             if m in fid_to_local:
                 assignments[fid_to_local[m]] = local_cid
 
-    # Per-block scale_compression_ratio (analogous to the top-level
-    # helper but scoped to this block's W_dec slice + per-block
-    # clusters). The local plan has local-indexed clusters; rebuild
-    # one for the helper.
+    # Per-block diagnostics (Phase 2 enhancement — completes Phase 2
+    # v1's deferred per-block rank_ratio / post_A / informative_metric).
+    # Rebuild a local-indexed plan for the existing diagnostic helpers,
+    # which take a CompressionPlan + W_dec slice and operate by row.
+    rank_ratio: float | None = None
+    post_A: float | None = None
+    informative_metric_value = None
     if block_clusters:
         local_plan = CompressionPlan(
             clusters=tuple(
@@ -1254,6 +1257,17 @@ def _build_block_report(
         scale_ratio = _compute_scale_compression_ratio(
             sub_source_w_dec, local_plan, sub_merged_norms
         )
+        # rank_ratio scoped to the block's rewritten sub-W_dec — the
+        # block's cluster representatives' decoder rows. Same numerical-
+        # rank-vs-d_model semantic as the top-level metric.
+        rank_ratio = _compute_rank_ratio(sub_rewritten_w_dec, local_plan)
+        # post_A scoped to the block's source sub-W_dec.
+        post_A = _compute_post_A(sub_source_w_dec, local_plan)
+        # informative_metric derived from rank_ratio per the existing
+        # _informative_metric rule (post_A < 0.95, both 0.95-1.05,
+        # forge_mse > 1.05).
+        if rank_ratio is not None:
+            informative_metric_value = _informative_metric(rank_ratio)
     else:
         scale_ratio = 1.0
 
@@ -1268,10 +1282,10 @@ def _build_block_report(
         n_clusters=n_clusters,
         cluster_assignments=tuple(assignments),
         scale_compression_ratio=scale_ratio,
-        rank_ratio=None,   # Phase 2 v1: deferred
-        post_A=None,        # Phase 2 v1: deferred
+        rank_ratio=rank_ratio,
+        post_A=post_A,
         forge_mse=None,
-        informative_metric=None,
+        informative_metric=informative_metric_value,
     )
 
 

runs/partition_payoff_measurement.json

@@ -0,0 +1,54 @@
+{
+  "n_features": 32,
+  "d_model": 64,
+  "n_confirmed_pairs": 16,
+  "uniform": {
+    "encoding": "Rung5(n_amp_qubits=4)",
+    "n_features_kept": 16,
+    "n_clusters": 16,
+    "scale_compression_ratio": 0.49999999254941935,
+    "substrate_cost_kept_reps_slots": 2048,
+    "substrate_cost_full_input_slots": 4096,
+    "wall_ms": 106.94289207458496
+  },
+  "partition": {
+    "blocks": [
+      {
+        "block_id": "heavy",
+        "encoding": "Rung5({'n_amp_qubits': 4})",
+        "slot_cost": 128,
+        "n_features": 4,
+        "n_kept": 2,
+        "n_zeroed": 2,
+        "n_clusters": 2,
+        "scale_compression_ratio": 0.5,
+        "rank_ratio": 0.03125,
+        "post_A": -1.1920928244535389e-07,
+        "informative_metric": "post_A"
+      },
+      {
+        "block_id": "tail",
+        "encoding": "MPSRung1({})",
+        "slot_cost": 8,
+        "n_features": 28,
+        "n_kept": 14,
+        "n_zeroed": 14,
+        "n_clusters": 14,
+        "scale_compression_ratio": 0.4999999914850507,
+        "rank_ratio": 0.21875,
+        "post_A": -8.514949412230521e-09,
+        "informative_metric": "post_A"
+      }
+    ],
+    "n_features_kept": 16,
+    "n_clusters": 16,
+    "scale_compression_ratio": 0.49999999254941935,
+    "substrate_cost_kept_reps_slots": 368,
+    "substrate_cost_full_input_slots": 736,
+    "wall_ms": 17.501115798950195
+  },
+  "reductions": {
+    "kept_reps": 5.565217391304348,
+    "full_input_sae": 5.565217391304348
+  }
+}

runs/real_partition_experiment.json

@@ -0,0 +1,74 @@
+{
+  "sae_repo": "jbloom/GPT2-Small-SAEs-Reformatted",
+  "sae_file": "blocks.8.hook_resid_pre/sae_weights.safetensors",
+  "n_features": 64,
+  "d_model": 768,
+  "heavy_k": 4,
+  "confirmer": "DecoderGeometryConfirmer",
+  "threshold": 0.1,
+  "n_confirmed_pairs": 4,
+  "heaviness_score": "decoder_norm\u00b2 \u00d7 pair_count",
+  "top_k_heavy_fids": [
+    3,
+    19,
+    22,
+    42
+  ],
+  "uniform": {
+    "encoding": "Rung5(n_amp_qubits=4)",
+    "n_features_kept": 4,
+    "n_features_zeroed": 4,
+    "n_clusters": 4,
+    "scale_compression_ratio": 0.49999797352125386,
+    "rank_ratio": 0.005208333333333333,
+    "substrate_kept_slots": 512,
+    "substrate_full_slots": 8192,
+    "wall_s": 0.09181714057922363,
+    "forge_faithfulness_kl": 10.0484037399292,
+    "forge_n_params": 124439808
+  },
+  "partition": {
+    "blocks": [
+      {
+        "block_id": "heavy",
+        "encoding": "Rung5({'n_amp_qubits': 4})",
+        "slot_cost": 128,
+        "n_features": 4,
+        "n_kept": 0,
+        "n_zeroed": 0,
+        "n_clusters": 0,
+        "scale_compression_ratio": 1.0,
+        "rank_ratio": null,
+        "post_A": null,
+        "informative_metric": null
+      },
+      {
+        "block_id": "tail",
+        "encoding": "MPSRung1({})",
+        "slot_cost": 8,
+        "n_features": 60,
+        "n_kept": 4,
+        "n_zeroed": 4,
+        "n_clusters": 4,
+        "scale_compression_ratio": 0.49999797352125386,
+        "rank_ratio": 0.005208333333333333,
+        "post_A": 8.940070594931626e-08,
+        "informative_metric": "post_A"
+      }
+    ],
+    "n_features_kept": 4,
+    "n_features_zeroed": 4,
+    "n_clusters": 4,
+    "scale_compression_ratio": 0.49999797352125386,
+    "rank_ratio": 0.005208333333333333,
+    "substrate_kept_slots": 32,
+    "substrate_full_slots": 992,
+    "wall_s": 0.03039097785949707,
+    "forge_faithfulness_kl": 10.0484037399292,
+    "forge_n_params": 124439808
+  },
+  "substrate_reductions": {
+    "kept_reps": 16.0,
+    "full_input_sae": 8.258064516129032
+  }
+}