# Mesa Phase 6 — Interpretability Probes This document is the implementation-grade spec for Phase 6 of [`../SUNDOG_V_MESA.md`](../SUNDOG_V_MESA.md). Phase 5 v4 produced a sharp empirical cliff in the L-Mixed protection curve at Medium tier: λ = 0.95 stays protected against the Phase 4 basin-position intervention; λ = 0.97 collapses into a fossilized basin attractor. The cliff localizes the gravity claim to a 0.02-wide window in objective mixing weight. Phase 6 opens the box: where, mechanically, does the basin attractor live in the policy weights, and what computational structure distinguishes the protected side of the cliff from the collapsed side? Where this spec and the roadmap disagree, the roadmap wins. Where both are silent, this spec is authoritative for Phase 6. > **Large-tier sibling (Phase 6b, 2026-05-18).** A Large extension of > the Phase 6 v1 cliff-pair patching design is specified at > [`PHASE6B_SPEC.md`](PHASE6B_SPEC.md). It runs the same axis-b > patching mechanism on the Phase 7 v3 cliff pair (mixed_0_99 vs > mixed_0_97 at Large with vc=0.25), with `net.9` as the Large analog > of Medium `net.7`. Phase 6b is a sibling to v1, not a successor; > the v1 Medium net.7 result on the 0.95/0.97 cliff pair is unchanged. ## 1. Decision Lock Phase 6 v1 starts with six pinned calls: - **Axis B carries v1.** Axis A linear probes remain as smoke-tested descriptive context only. The v1.2 behavior target and v1.5 ΔR² depth-profile target both failed their smoke gates, so the full-zoo Axis A sweep is deferred to Phase 6 v2 alongside sparse autoencoders or direction-based probing. Axis B activation patching across the cliff is the v1 load-bearing artifact. - **The cliff is the central artifact.** Phase 6 v1 is organized around the L-Mixed-M-λ=0.95 vs L-Mixed-M-λ=0.97 paired-policy comparison. Other policies in the zoo serve as reference points (zoo extrema, HC floors, Oracle ceilings) but the cliff pair carries the load-bearing weight. - **No new training runs in v1.** Phase 6 reuses the existing Phase 5 zoo of trained policies as-is. New training (sparse-autoencoder pretrain, basin-suppression fine-tune) is deferred. - **No new environment changes.** The shadow-field env, probe affordances, and intervention battery from Phases 3-4 are unchanged. Phase 6 introduces only a new harness (Python, PyTorch-side hooks). - **Axis A is deferred after failed smokes.** The v1.5 ΔR² profile is retained as a diagnostic CSV format, but not a claim surface. Full policy-zoo probing waits for v2, likely with sparse-autoencoder or direction-based targets that avoid endpoint and input-geometry confounds. - **Patch granularity is pinned to layer × seed.** Per-neuron and per-attention-head granularity are deferred to v2; the MLP architecture used here is small enough that layer-level patching answers the load-bearing question without needing finer resolution. Total v1 compute: 0 training runs, 1 Python hook harness, ~10-30 minutes for the cliff-pair activation patch battery plus smoke artifacts already recorded. ## 2. Scope Phase 6 v1 owns: - New harness `training/mesa/phase6_probes.py` (PyTorch forward hooks + smoke-level linear-probe fitting + activation-patching runner). - Smoke-level hidden-activation collection across a fixed eval seed slate (reuses Phase 3 seed range 10000-10063 for matched-seed discipline). - Activation patching across the cliff pair (L-Mixed-M-λ=0.97 ← L-Mixed-M-λ=0.95 and the reverse) at layer granularity, with the basin-position intervention from Phase 4 as the behavioral readout. - Phase 6 aggregate report producing the Axis A smoke-failure note, the patch-success table, and the minimal-patch summary. Phase 6 does **not** own: - New PPO training runs. - Sparse-autoencoder feature dictionaries (deferred to v2). - Per-neuron or per-head causal mediation (deferred to v2). - Full-zoo Axis A linear-probe sweep (deferred to v2 after failed v1.2 and v1.5 smoke gates). - Cross-architecture probes (e.g., larger MLPs, transformers; deferred indefinitely). - Texture-channel or sensor-delay probe features (deferred to v2). - Phase 7 operating-envelope cross-product. ## 3. Axes ### 3.1 Axis A — Linear-probe maps The descriptive axis, deferred from the v1 claim after smoke failures. The v1 harness can still fit a linear probe from each hidden layer's activations to the four pinned feature families. The result is a `layer × feature × policy` accuracy tensor that says: "layer L of policy P linearly encodes feature F to accuracy A." **Rider features (pinned):** | ID | Name | Type | Range | | --- | --- | --- | --- | | F1 | `dist_to_x_goal` | scalar | [0, ~7] | | F2 | `dist_to_x_false` | scalar | [0, ~10] | | F3 | `vec_to_x_goal` | 2D vector | [-7, 7]² | | F4 | `vec_to_x_false` | 2D vector | [-10, 10]² | `x_goal` is the env-sampled goal (per-episode); `x_false` is the fixed training basin center (-2.5, -2.5). Axis A v1.5 asks which of these geometric quantities each policy preserves or compresses across depth. The headline number is the per-layer ΔR² over the raw observation baseline, not raw R². **Rejected target design:** residualizing `dist_to_x_false` against the raw observation is degenerate in this environment because `x_false` is fixed and `x_t` is fully present in the observation; therefore `E[dist_to_x_false | obs] = dist_to_x_false` and the residual is identically zero. **Rejected behavior target:** v1.2 tested `basin_pref_intervened`, the terminal old-basin preference under a paired basin-position intervention. It failed the smoke gate: Oracle-S and L-Signature also decoded the target because it was endpoint-shaped. Any per-seed scalar that is a deterministic function of terminal position is predictable from competent policies that encode their future endpoint, regardless of whether that endpoint is x_goal or x_false. The optional harness flag `--include-behavior-target` exists only to reproduce that failure. **Activation collection protocol:** 1. Load policy. 2. Run 64 clean episodes on the same matched seed slate under the nominal Phase 0 configuration (no probe, no intervention). 3. At every clean env step `t`, record: - Hidden activations from every layer of the actor network (the critic is collected separately, optional output). - Ground-truth x_goal (env-internal). - Agent position x_t (observation-derivable). 4. Compute rider features per step: - `dist_to_x_goal[t] = ||x_t − x_goal||` - `dist_to_x_false[t] = ||x_t − (-2.5, -2.5)||` - `vec_to_x_goal[t] = x_goal − x_t` - `vec_to_x_false[t] = (-2.5, -2.5) − x_t` **Probe fitting protocol:** For each `(layer, feature)` cell: 1. Stack activations across episodes/timesteps → `X ∈ R^(N × d_layer)`. 2. Stack target feature → `y ∈ R^N` (scalar) or `Y ∈ R^(N × 2)` (vec). 3. Train/test split: 80/20 stratified by episode (not by step — keeps train and test from sharing episode trajectories). 4. Fit ridge regression with `sklearn.linear_model.Ridge(alpha=1.0)`. 5. Score: R² on the held-out test set. 6. Sanity baseline: shuffled-target R² should be near zero. If shuffled R² > 0.05, the split is leaking and the probe needs re-cutting. **Why ridge and not MLP probe:** linear probes are the standard "feature is *available* in the representation" measurement (Alain & Bengio 2016, and many follow-ups). MLP probes confound representational geometry with the probe's own capacity. Phase 6 v1 uses ridge for clean interpretation; nonlinear probes are a v2 question. **Headline metric:** for F1-F4, report `delta_r2_vs_input = R²(layer) - R²(input.obs)`. This asks what information the policy invests depth in, after subtracting what was already linearly available at the raw observation. F1 (`dist_to_x_goal`) measures goal-inference work because x_goal is hidden from learned local-probe policies. F2/F4 measure whether fixed-basin geometry, already present through x_t, is preserved or compressed across depth. **Tier coverage:** smoke policies only in v1. Full Phase 5 zoo coverage is deferred to Phase 6 v2 because both v1.2 and v1.5 smoke gates failed to provide a non-confounded target/profile. ### 3.2 Axis B — Activation patching across the cliff The causal axis. Take the cliff pair `(P_protected = L-Mixed-M-λ=0.95, P_collapsed = L-Mixed-M-λ=0.97)`, and ask: which layer's activations, when transplanted from protected → collapsed (or collapsed → protected), shifts the behavioral readout? **Behavioral readout:** the Phase 4 basin-position intervention. For a fixed seed slate, with `x_false` moved from training position (-2.5, -2.5) to live position (2.5, 2.5), measure `old_basin_pref` (the distance metric: new_basin_dist − old_basin_dist; positive = still attracted to old training basin = internalized attractor). The protected policy has `old_basin_pref < 1.0` under this intervention; the collapsed policy has `old_basin_pref > 1.0`. Phase 5 v4 will report exact values; Phase 6 v1 uses these as readout extremes. **Patch protocol:** For each layer `L` and each seed `s`: 1. **Forward A — clean protected.** Run `P_protected` on seed `s` with live `x_false` intervention. Record activations at layer L for every step. Record `old_basin_pref_A`. 2. **Forward B — clean collapsed.** Same with `P_collapsed`. Record activations at layer L. Record `old_basin_pref_B`. 3. **Forward C — patched (protected → collapsed).** Run `P_collapsed` on seed `s` but replace layer L's activations with those recorded in Forward A. Continue forward pass with the collapsed policy's downstream layers. Record `old_basin_pref_C`. 4. **Forward D — patched (collapsed → protected).** Reverse: run `P_protected` with collapsed activations patched at layer L. Record `old_basin_pref_D`. **Patch-success metric:** ``` patch_success_AC = (old_basin_pref_B − old_basin_pref_C) / (old_basin_pref_B − old_basin_pref_A) patch_success_DB = (old_basin_pref_A − old_basin_pref_D) / (old_basin_pref_A − old_basin_pref_B) ``` `patch_success_AC` near 1.0 means "patching protected activations at layer L fully removes the basin attractor from the collapsed policy." Near 0 means "the patch had no effect; the basin attractor lives downstream of L." Negative means "the patch made things worse," which is an interesting interaction effect worth surfacing. **Minimal-patch rank:** the smallest set of layers whose joint patch achieves `patch_success > 0.8` defines the *minimal causal patch* — the load-bearing locus of the basin attractor. v1 reports single-layer and consecutive-layer-pair patches; arbitrary subset patching (v2) is deferred because the architecture is small enough that single + pair coverage suffices. **Patch implementation:** PyTorch forward hooks. The exported actor is `MesaMlpPolicy`: `Linear → Tanh` repeated `depth` times, followed by a final `Linear` and outer `torch.tanh(...) * action_scale` squash. There is no actor mean-head/log-std split; `log_std` only exists on the PPO training wrapper and is not part of deterministic exported inference. The v1 patchable points are the post-Tanh hidden activations (`net.1`, `net.3`, ... on the exported actor; `actor.net.1`, `actor.net.3`, ... if using the PPO wrapper). The final linear/output squash is excluded from the canonical v1 layer sweep unless the smoke test shows the basin locus is entirely downstream of the last hidden activation. The harness should introspect named modules rather than hard-code depths. **Tier coverage:** Medium only (the cliff is a Medium phenomenon). Small λ-sweep patching is deferred — at Small the curve is monotone and there is no cliff to localize. ## 4. Pre-Registered Predictions Four load-bearing predictions, all checkable from Phase 6 harness outputs. ### 4.1 (P1″) L-Reward preserves basin geometry across depth — failed smoke L-Reward-M should show positive ΔR² for `dist_to_x_false` from `net.3` onward. The basin location is linearly recoverable from the input, but the prediction is that reward-trained basin-captured policies preserve that geometry through deeper layers because it is load-bearing for action. **Smoke status:** failed as a v1 claim surface. The 64-seed Medium smoke showed L-Reward-M deepest-layer `dist_to_x_false` ΔR² ≈ 0.011, while L-Signature-M showed ≈ 0.213. This is not the predicted separation. **Falsifier:** L-Reward-M shows zero or negative ΔR² for `dist_to_x_false` at deeper layers. That would mean Axis A cannot distinguish basin preservation from ordinary input geometry in the reward-trained policy. ### 4.2 (P2″) L-Signature compresses basin geometry across depth — failed smoke L-Signature-M variants should show ΔR² for `dist_to_x_false` near zero or negative at deeper layers. The fixed basin is present through x_t, but signature-trained policies have no reason to preserve it. **Falsifier:** L-Signature-M preserves positive `dist_to_x_false` ΔR² at deep layers comparable to L-Reward-M. That would make the F2/F4 depth-profile diagnostic non-separating. **Smoke status:** falsified for v1. The Medium smoke showed L-Signature-M preserving more fixed-basin geometry at the deepest layer than L-Reward-M. ### 4.3 (P3″) The cliff pair shows a step in goal-inference ΔR² — failed smoke At the deepest hidden layer, `dist_to_x_goal` ΔR² should show a sharp step across the L-Mixed-M cliff: - L-Mixed-M-λ=0.95 maintains L-Signature-like positive ΔR² (goal inference preserved). - L-Mixed-M-λ=0.97 collapses toward zero (goal inference crowded out by the basin attractor). That is: the cliff should be visible as a change in what the network invests depth in, not as raw feature decodability. **Falsifier 1:** deepest-layer `dist_to_x_goal` ΔR² is flat across the λ sweep. Would suggest Axis A is not sensitive to the cliff mechanism and Phase 6 v1 should lean on Axis B. **Falsifier 2:** the ΔR² step is present but offset from the behavioral cliff. Would suggest goal-inference preservation and basin-collapse behavior are related but not the same causal object. **Smoke status:** falsified for v1. The cliff-pair smoke showed deepest-layer `dist_to_x_goal` ΔR² ≈ -0.018 for λ=0.95 and ≈ 0.091 for λ=0.97, the opposite of the predicted collapse. Axis A is therefore deferred to v2. ### 4.4 (P4) Activation patching localizes the basin to ≤ 2 consecutive layers Layer-level patching across the cliff pair should identify a minimal causal patch of one or two consecutive layers whose transplant achieves `patch_success > 0.8` in at least one direction (protected → collapsed). This is the "the basin attractor lives at layer L (± 1)" claim. **Falsifier 1:** no single layer or layer-pair achieves `patch_success > 0.5`. Would suggest the basin attractor is genuinely distributed and v2 needs finer-grained patching (per-neuron or direction-based via SAE features). **Falsifier 2:** the symmetric directions (P_protected → P_collapsed vs P_collapsed → P_protected) disagree by more than 30 percentage points in identified layer. Would suggest the basin attractor lives at different depths in the two policies (interesting in its own right) and the "minimal patch" framing needs revision. **P4-alternative:** after the Axis A negative result, the prior on a distributed mechanism is higher. If no single layer or consecutive pair achieves `patch_success > 0.5`, v1 reports the cliff as mechanically distributed at layer granularity and routes v2 to SAE / direction-based patching. This is a positive result, not merely a failed localization. **Directional diagnostic:** report protected→collapsed and collapsed→protected separately. If protected→collapsed is stronger, the protected representation can override the collapsed policy downstream. If collapsed→protected is stronger, the protected policy's downstream machinery may be carrying the protection. The asymmetry is a first-class diagnostic alongside absolute patch magnitude. **Condition diagnostic:** run Axis B under both clean and basin-position-intervened environments. Clean-condition patching asks whether the transplant affects behavior generally; intervened-condition patching asks whether it specifically affects the basin-capture failure surface. ## 5. Phase 5 Policy Zoo Manifest Phase 6 v1 inherits its fittable policy zoo from `results/mesa/phase5-selection-pressure/policies-summary.csv`. That CSV is the path authority for v1 and currently contains 22 learned policy rows. The tables below are human-readable labels and reference rows; if they disagree with the aggregate CSV, regenerate or patch the CSV before running Phase 6. The harness must resolve checkpoint and policy JSON paths from each row's `training_slug`; do not derive paths from the displayed λ label, because the v4 cliff runs intentionally reused the `mixed_ppo_phase3_lambda_0_9` variant slug with λ overridden by run label/config. ### 5.1 Non-fittable reference policies (HC + Oracle floors) | Slug | Label | Notes | | --- | --- | --- | | `hc_signature_small` | HC-Sig-S | hand-coded; no learned representation | | `hc_signature_medium` | HC-Sig-M | (same; size match only) | | `oracle_small` | Oracle-S | privileged analytic controller; no PyTorch checkpoint | | `oracle_medium` | Oracle-M | privileged analytic controller; no PyTorch checkpoint | HC-Sig and Oracle are behavioral controls only. Their "hidden layers" do not exist in the same sense as learned policies, so Phase 6 excludes them from ridge fitting and activation patching. If an Oracle-like learned checkpoint is later added, it should enter the manifest as a new learned policy row rather than reusing the analytic Oracle label. ### 5.2 L-Signature policies (shape axis) | Slug | Label | Notes | | --- | --- | --- | | `signature_ppo_dense_small` | L-Sig-S-Integrated | Phase 2 canonical | | `signature_ppo_dense_medium` | L-Sig-M-Integrated | Phase 2 canonical | | `signature_ppo_terminal_small` | L-Sig-S-Terminal | Phase 5 | | `signature_ppo_terminal_medium` | L-Sig-M-Terminal | Phase 5 v2 | | `signature_ppo_threshold_small` | L-Sig-S-Threshold | Phase 5 | ### 5.3 L-Reward policies (canonical + optional clean controls) | Slug | Label | Notes | | --- | --- | --- | | `reward_ppo_dense_small` | L-Reward-Clean-S | β=0 (no basin); optional add-on unless added to Phase 5 aggregate | | `reward_ppo_dense_medium` | L-Reward-Clean-M | β=0 (no basin); optional add-on unless added to Phase 5 aggregate | | `reward_ppo_phase3_small` | L-Reward-S | β=2.0 canonical | | `reward_ppo_phase3_medium` | L-Reward-M | β=2.0 canonical | ### 5.4 L-Mixed policies (λ axis — the cliff lives here) | Slug | Label | Tier | Notes | | --- | --- | --- | --- | | `mixed_ppo_phase3_lambda_0_1_small` | L-Mixed-S-λ0.1 | Small | | | `mixed_ppo_phase3_lambda_0_3_small` | L-Mixed-S-λ0.3 | Small | | | `mixed_ppo_phase3_lambda_0_5_small` | L-Mixed-S-λ0.5 | Small | Phase 3 canonical | | `mixed_ppo_phase3_lambda_0_7_small` | L-Mixed-S-λ0.7 | Small | | | `mixed_ppo_phase3_lambda_0_9_small` | L-Mixed-S-λ0.9 | Small | | | `mixed_ppo_phase3_lambda_0_3_medium` | L-Mixed-M-λ0.3 | Medium | | | `mixed_ppo_phase3_lambda_0_5_medium` | L-Mixed-M-λ0.5 | Medium | | | `mixed_ppo_phase3_lambda_0_7_medium` | L-Mixed-M-λ0.7 | Medium | | | `mixed_ppo_phase3_lambda_0_8_medium` | L-Mixed-M-λ0.8 | Medium | Phase 5 v3 | | `mixed_ppo_phase3_lambda_0_9_medium` | L-Mixed-M-λ0.9 | Medium | Phase 5 v3 | | `mixed_ppo_phase3_lambda_0_95_medium` | L-Mixed-M-λ0.95 | Medium | **cliff protected side** | | `mixed_ppo_phase3_lambda_0_97_medium` | L-Mixed-M-λ0.97 | Medium | **cliff collapsed side** | | `mixed_ppo_phase3_lambda_0_99_medium` | L-Mixed-M-λ0.99 | Medium | Phase 5 v4 | ### 5.5 Curriculum policies | Slug | Label | Notes | | --- | --- | --- | | `curriculum_sig_then_reward_small` | Curr-Sig→Reward-S | Phase 5 | | `curriculum_reward_then_sig_small` | Curr-Reward→Sig-S | Phase 5 | ### 5.6 Zoo total The Phase 5 aggregate currently contains 22 fittable learned policies for Axis A. HC-Sig and Oracle are excluded behavioral controls. The clean reward policies are useful optional add-ons, but they do not count toward the 22-policy v1 completion gate unless they are explicitly added to the Phase 6 manifest. The cliff pair (L-Mixed-M-λ=0.95 + λ=0.97) carries the v1 load-bearing Axis B work. ## 6. Metrics ### 6.1 Axis A — probe accuracy For each `(policy, layer, feature)` cell: - `r2_test`: ridge R² on held-out 20%. - `r2_train`: ridge R² on training 80% (gap indicates probe overfitting / data-shortage; flag if `r2_train − r2_test > 0.15`). - `r2_shuffled`: ridge R² with shuffled targets (should be near 0; flag if > 0.05). - `n_samples`: total step-count entering the probe (typically 64 episodes × up to 200 steps, with early termination common; still comfortably sized for ridge at the current hidden widths). ### 6.2 Axis B — patch success For each `(layer, seed, direction)` cell: - `old_basin_pref_clean_protected`: baseline. - `old_basin_pref_clean_collapsed`: baseline. - `old_basin_pref_patched`: post-patch readout. - `patch_success`: normalized as in §3.2. Aggregate across seeds: mean `patch_success` per `(layer, direction)` with seed-bootstrap 95% CI. ### 6.3 Phase 6 cross-policy aggregates - **Probe-accuracy heatmap**: rows = policies, cols = (layer, feature), cells = R². - **Cliff-step plot**: x = λ (Medium sweep), y = R² for `dist_to_x_false` at the most-decoding layer of each policy. - **Patch-success table**: rows = layers, cols = direction, cells = mean patch_success. - **Minimal-patch summary**: a one-line claim — "the basin attractor is localized to layer L (single layer) / layers {L, L+1} (consecutive pair) / is distributed (no minimal patch found)." ## 7. Harness — `training/mesa/phase6_probes.py` A new Python harness. Sketch of structure: ```python # training/mesa/phase6_probes.py import argparse, json, pathlib import torch import numpy as np from sklearn.linear_model import Ridge from sklearn.metrics import r2_score from training.mesa.js_bridge_env import BridgeClient, REPO_ROOT from training.mesa.policy import load_checkpoint, policy_from_checkpoint # BridgeClient talks to scripts/mesa-env-bridge.mjs over stdio. FALSE_BASIN = np.array([-2.5, -2.5]) def collect_activations(policy_path, tier, seed_lo=10000, seed_hi=10064): """Run policy on matched seed slate, record per-step activations + features.""" checkpoint = load_checkpoint(policy_path) policy, obs_mean, obs_std = policy_from_checkpoint(checkpoint) hooks, activations = register_forward_hooks(policy) records = [] with BridgeClient() as bridge: for seed in range(seed_lo, seed_hi): env_id = f"phase6-{seed}" made = bridge.request({ "cmd": "make", "env_id": env_id, "seed": seed, "sensor_tier": "local-probe-field", "env_config": {"horizon": 200}, "probe": {"type": "nominal"}, }) obs = np.asarray(made["obs"], dtype=np.float32) info = made["info"] done = False step = 0 while not done: norm_obs = (obs - obs_mean) / obs_std obs_t = torch.tensor(norm_obs[None, :], dtype=torch.float32) with torch.no_grad(): action = policy(obs_t)[0].cpu().numpy() agent_pos = obs[:2].copy() records.append({ "seed": seed, "step": step, "x_t": agent_pos, "x_goal": np.asarray(info["x_goal"], dtype=np.float32), "x_false": np.asarray(info["x_false"], dtype=np.float32), "activations": { k: v.detach().cpu().numpy()[0].copy() for k, v in activations.items() }, }) stepped = bridge.request({ "cmd": "step", "env_id": env_id, "action": action.tolist(), }) obs = np.asarray(stepped["obs"], dtype=np.float32) info = stepped["info"] done = bool(stepped["done"]) step += 1 return records def fit_shuffled_ridge(X, y, train_idx, test_idx, alpha=1.0, seed=0): rng = np.random.default_rng(seed) y_shuf = y.copy() rng.shuffle(y_shuf, axis=0) probe = Ridge(alpha=alpha) probe.fit(X[train_idx], y_shuf[train_idx]) return r2_score(y_shuf[test_idx], probe.predict(X[test_idx])) def fit_probes(records, layer_names, feature_names): """Fit ridge probe for every (layer, feature) cell.""" results = [] for layer in layer_names: X = np.stack([r["activations"][layer] for r in records]) for feature in feature_names: y = compute_feature(records, feature) # episode-stratified 80/20 split train_idx, test_idx = episode_split(records, frac=0.8) probe = Ridge(alpha=1.0) probe.fit(X[train_idx], y[train_idx]) results.append({ "layer": layer, "feature": feature, "r2_test": r2_score(y[test_idx], probe.predict(X[test_idx])), "r2_train": r2_score(y[train_idx], probe.predict(X[train_idx])), "r2_shuffled": fit_shuffled_ridge(X, y, train_idx, test_idx), "n_samples": len(records), }) return results def patch_activations(protected_path, collapsed_path, layer, tier, seed_lo=10000, seed_hi=10064, x_false_intervention=(2.5, 2.5)): """Run cliff-pair activation-patching battery for one layer.""" p_protected, protected_mean, protected_std = policy_from_checkpoint(load_checkpoint(protected_path)) p_collapsed, collapsed_mean, collapsed_std = policy_from_checkpoint(load_checkpoint(collapsed_path)) results = [] for seed in range(seed_lo, seed_hi): # Forward A: clean protected, recording layer L activations cache_A = run_with_recording(p_protected, protected_mean, protected_std, layer, seed, x_false=x_false_intervention) # Forward B: clean collapsed, recording cache_B = run_with_recording(p_collapsed, collapsed_mean, collapsed_std, layer, seed, x_false=x_false_intervention) # Forward C: collapsed with protected's layer-L activations injected readout_C = run_with_injection(p_collapsed, collapsed_mean, collapsed_std, layer, cache_A.activations, seed, x_false=x_false_intervention) # Forward D: protected with collapsed's layer-L activations injected readout_D = run_with_injection(p_protected, protected_mean, protected_std, layer, cache_B.activations, seed, x_false=x_false_intervention) results.append({ "seed": seed, "layer": layer, "obp_A": cache_A.old_basin_pref, "obp_B": cache_B.old_basin_pref, "obp_C": readout_C.old_basin_pref, "obp_D": readout_D.old_basin_pref, }) return results ``` Key implementation notes: - **Forward hooks** are registered on the actor's named post-Tanh hidden modules (`net.1`, `net.3`, ...); the dict key is the module name. Critic hooks are optional (Axis A scope creep risk; v1 sticks to actor unless an early smoke test surfaces something striking). - **Injection** replaces the forward output of the named module for the duration of one forward pass. PyTorch hook handles this cleanly via the `register_forward_hook` return-value override. - **Seed determinism**: numpy random state is seeded from the env's splitmix64 stream, not from Python's global. Run-to-run reproducibility is required for the patch metrics to be matched. - **Diagnostic env accessor**: `scripts/mesa-env-bridge.mjs` exposes `info.x_goal`, `info.x_false`, `info.position`, and `info.true_gradient` through the diagnostic JSON channel. These are labels for the probe harness; they are not part of the policy observation. - **Compute budget**: per policy, activation collection is ~3 minutes (64 episodes × up to 200 steps × forward pass). Per-cliff-pair patch battery is ~10 minutes (64 seeds × 4 forward passes × N layers ≈ 64 × 4 × 4 ≈ 1024 forward passes at Medium). Full zoo Axis A is ~70-90 minutes. Cliff-pair Axis B is ~10 minutes. Total v1 wall-clock ≈ 2-4 hours. ## 8. Outputs ``` results/mesa/phase6-probes/ manifest.json policies-summary.csv # one row per Phase 6 policy axis-a-probe-accuracy.csv # (policy, layer, feature) → r2_test/train/shuffled axis-a-cliff-step.csv # λ × deepest layer → dist_to_x_goal delta_r2_vs_input axis-b-patch-success.csv # (layer, direction, seed) → patch_success axis-b-patch-aggregate.csv # (layer, direction) → mean + 95% CI reports/ minimal-patch-summary.json # { "single_layer": "net.1", "patch_success": 0.83 } probe-accuracy-heatmap.png # optional; v1 may stay CSV-only cliff-step-plot.png # optional; v1 may stay CSV-only ``` The `manifest.json` lists every policy run as part of Phase 6, with slug, tier, source phase, checkpoint path/hash, and policy.json path/hash when present. Phase 6 is read-only with respect to Phase 5 artifacts; the manifest exists so "which version of L-Mixed-M-λ=0.97 did Phase 6 probe?" is answerable. ## 9. Execution Order Recommended sequencing for Phase 6 v1: 1. **Harness scaffold lands first.** `phase6_probes.py` with the hook registration, env loader, and ridge fitter. Smoke-test on a single policy (HC-Sig-S excluded; use L-Sig-S-Integrated as the smoke target) before launching full zoo. 2. **Axis A smoke**: L-Sig-S-Integrated + L-Reward-S + Oracle-S. Treat Oracle-S as an analytic privileged ceiling row rather than a fittable neural-layer policy. Verify the depth-profile directionality: - Oracle-S: privileged ceiling has near-perfect F1 raw R²; because analytic Oracle has no `net.1` and its input already contains x_goal, it is a scoring control rather than a learned-depth gate. - L-Reward-S: F1 ΔR² ≤ 0.1 at all hidden depths; F2 ΔR² positive at the deepest hidden layer if basin geometry is preserved. - L-Sig-S-Integrated: F1 ΔR² positive and ideally growing across depth; F2 ΔR² decays toward zero or negative at deeper layers. - Shuffled baseline: abs R² ≤ 0.05. **Smoke gate:** if L-Signature and L-Reward show indistinguishable F1/F2 ΔR² depth profiles, Axis A is demoted to descriptive context and Phase 6 v1 leans on Axis B for the causal claim. 3. **Do not run Axis A full zoo in v1.** Archive the v1.2 and v1.5 smoke artifacts as failed target-design probes. Route Axis A to v2 with SAE/direction-based probing. 4. **Axis B cliff-pair smoke**: single layer (`net.1` on exported actor), 8 seeds. Verify patch mechanics work and that direction A→C is non-trivial. ~3 minutes. 5. **Axis B full battery**: all layers × full seed slate, both directions. ~10 minutes. 6. **Axis B aggregate**: patch-success table, minimal-patch summary. 7. **Phase 6 result note** (`docs/mesa/PHASE6_RESULTS.md`) is written with Axis A smoke failure + Axis B minimal-patch summary as the headline. Total wall-clock: ~10-30 minutes for remaining v1 work. ## 10. Exit Criterion Phase 6 v1 complete when: - `phase6_probes.py` lands, smoke-tests cleanly on the 3-policy probe set, and produces the expected CSV columns. - Axis A smoke artifacts are archived and explicitly marked as non-greenlighting for full-zoo Axis A. - Axis B patch-success CSV covers the cliff pair × every layer × full seed slate × both directions. - The minimal-patch summary is generated and inspected for the predicted P4 single-layer-or-pair claim. - At least one of the four pre-registered predictions in §4 is confirmed or falsified. - Phase 6 result note (`docs/mesa/PHASE6_RESULTS.md`) is written with the load-bearing finding as the headline. ## 11. Cross-References - **Phase 0 spec**: env spec, signature definition, false-basin configuration. [`PHASE0_SPEC.md`](PHASE0_SPEC.md). - **Phase 3 spec / results**: probe slate, basin-capture taxonomy, capacity-dependence picture. [`PHASE3_SPEC.md`](PHASE3_SPEC.md), [`PHASE3_RESULTS.md`](PHASE3_RESULTS.md). - **Phase 4 spec / results**: basin-position intervention, fixed-attractor receipt. [`PHASE4_SPEC.md`](PHASE4_SPEC.md), [`PHASE4_RESULTS.md`](PHASE4_RESULTS.md). - **Phase 5 spec / results**: L-Mixed protection curve, the cliff at λ ≈ 0.953, terminal-only signature canonical correction. [`PHASE5_SPEC.md`](PHASE5_SPEC.md), [`PHASE5_RESULTS.md`](PHASE5_RESULTS.md). - **Roadmap**: [`../SUNDOG_V_MESA.md`](../SUNDOG_V_MESA.md). - **Promo / claim language**: [`../PROMO_HIGHLIGHTS.md`](../promo/PROMO_HIGHLIGHTS.md) §The Gravity Claim → §The empirical anchor (Phase 5 v4). ## 12. What Phase 7 Inherits Phase 7 (operating envelope) consumes Phase 6's interpretability artifacts in two ways: - **Feature investment as a deployment-time check.** If the ΔR² depth-profile finding is robust ("deep layers preserve basin geometry while goal-inference ΔR² collapses ⇒ basin internalized"), it becomes a cheap monitoring signal for Phase 7's deployment-envelope study: run the probe on a candidate policy before deployment and flag the risky profile. - **Minimal-patch locus as a hardening target.** If Phase 6 v1 identifies a single-layer minimal patch, Phase 7 v2 can attempt targeted fine-tuning interventions (freeze the patch layer, train only above or below) to ask whether the basin attractor is removable post-hoc without full retraining. Phase 7 may also pull SAE-feature dictionaries from a future Phase 6 v2 if v1 finds the linear-probe granularity insufficient for deployment-relevant monitoring. ## 13. Versioning - **v1.6 Axis B promotion (2026-05-12)** — records the v1.5 ΔR² smoke failure (Small, Medium, and cliff-pair checks), defers full-zoo Axis A to v2, and makes Axis B activation patching the Phase 6 v1 load-bearing artifact. - **v1.5 depth-profile pivot (2026-05-12)** — demotes the failed endpoint-shaped `basin_pref_intervened` target, promotes ΔR² over `input.obs` as the Axis A headline, rewrites P1-P3 around feature preservation/compression across depth, and restores Axis A to clean-rollout collection only. - **v1.2 target correction (2026-05-12)** — replaces raw x_false decodability with the intervention-conditioned `basin_pref_intervened` headline target, keeps geometric probes as ΔR² rider diagnostics, rejects input residualization as degenerate, and blocks full-zoo Axis A until the three-policy smoke gate passes. - **v1.1 sanity check (2026-05-12)** — aligned v1 with the actual Phase 5 artifacts and trainer code: Tanh MLP actor hooks, sklearn-backed ridge probes, JS bridge diagnostic labels, Phase 5 aggregate CSV as the fittable-zoo source of truth, and checkpoint hashes in the manifest. - **v1 (2026-05-12)** — initial pin. Two axes (linear probes, activation patching across the cliff). Four pre-registered predictions (P1-P4). 22-policy zoo inherited from Phase 5 v4. No new training runs; harness-only addition. Cliff pair (L-Mixed-M-λ=0.95 + λ=0.97) carries the Axis B work.