# Mesa Phase 6b - Large-Tier Activation Patching Result Note This document records the Phase 6b result note for the Large-tier activation-patching battery defined in [`PHASE6B_SPEC.md`](PHASE6B_SPEC.md). Phase 6b is a Large sibling of Phase 6 v1 (Medium `net.7` localization of the basin-attractor circuit). The v1 22-cell Medium result is unchanged. Status: Phase 6b v1.1 **complete with pre-registered falsification**. GG6b-loc-D called (destructive-transfer falsifier at all 5 layers in both directions); GG6b-mech not called (per §6 of the spec, mech is not evaluated at destructive layers); GG6b-shape deferred (no locus to decompose). The v1 single-layer locus finding **does not generalize** to the Large field-coupled / under-budget cliff pair. ## 1. Summary Phase 6b ran the Phase 6 v1 axis-b activation-patching protocol on the Large cliff pair (mixed_0_99 recovery vs mixed_0_97 trough, `--value-coef 0.25`, 10M env-steps, `seed=10000`), sweeping the 5 post-Tanh hidden layers (`net.1`, `net.3`, `net.5`, `net.7`, `net.9`) at 64 seeds per cell, both `clean` and `intervened` conditions, both patch directions, and producing per-layer per-direction aggregate metrics. The smoke-v11 (n=8 at net.9) caught a v1 metric pathology (`old_basin_pref` gap ~1.0 made the v1 metric unbounded); the v1.1 amendment introduced an alignment-normalized parallel metric and a conjunctive P4 rule (mean ∧ median ∧ ratio_of_means all ≥ 0.8) plus a destructive-transfer falsifier (`|patched_align − target_align| / |gap_align| > 0.5`). The full-v11 sweep (2026-05-18, ~13.5 min wall-clock) yields one substantive finding: **At every swept layer at Large, activation injection between mixed_0_99 and mixed_0_97 drives the recipient policy's terminal alignment to ~0 (range 0.01–0.09) — below even the recipient's own un-patched baseline.** The patching is *actively destructive of navigation* rather than transferring behavior between sides. This fires the destructive-transfer falsifier at every (layer, direction) cell and routes the verdict to GG6b-loc-D. The destruction is not a metric artifact — it reads off raw `mean_patched_alignment`, a physical per-seed quantity that cannot explode. The verdict therefore stands independent of the conjunctive P4 reading of the normalized `patch_success_align` metric (which would also yield "no layer clears" at the conjunctive rule, but that's secondary). ## 2. Envelope `clean` rows, all 5 swept layers, both patch directions. n=64 per cell. `align_gap = 0.4267` (= `mean_protected_alignment 0.912 − mean_collapsed_alignment 0.486`). Rescue target = protected alignment 0.912; break target = collapsed alignment 0.486. | layer | dir | mean ps_align | median | ratio | patched_align | vs target / gap | verdict | | --- | --- | ---: | ---: | ---: | ---: | ---: | --- | | net.1 | rescue (P→C) | 1.442 | -0.482 | -1.043 | 0.041 | 2.043 | **DESTRUCTIVE** | | net.1 | break (C→P) | -0.162 | 1.407 | 1.983 | 0.066 | 0.983 | **DESTRUCTIVE** | | net.3 | rescue (P→C) | 0.928 | -0.453 | -1.106 | 0.014 | 2.106 | **DESTRUCTIVE** | | net.3 | break (C→P) | 0.065 | 1.485 | 2.070 | 0.029 | 1.070 | **DESTRUCTIVE** | | net.5 | rescue (P→C) | 0.974 | -0.457 | -1.076 | 0.027 | 2.076 | **DESTRUCTIVE** | | net.5 | break (C→P) | -0.026 | 1.285 | 1.924 | 0.091 | 0.924 | **DESTRUCTIVE** | | net.7 | rescue (P→C) | 0.862 | -0.457 | -1.004 | 0.057 | 2.004 | **DESTRUCTIVE** | | net.7 | break (C→P) | 0.076 | 1.464 | 2.103 | 0.015 | 1.103 | **DESTRUCTIVE** | | net.9 | rescue (P→C) | 0.779 | -0.485 | -1.056 | 0.035 | 2.057 | **DESTRUCTIVE** | | net.9 | break (C→P) | 0.058 | 1.460 | 2.103 | 0.015 | 1.103 | **DESTRUCTIVE** | `patched_align` is `mean_patched_alignment`; `vs target / gap` is `|mean_patched_alignment − target_align| / |align_gap|` (the destructive-transfer test from spec §6: > 0.5 fires the falsifier). Every cell exceeds 0.5; ten of ten gap-delta values lie in the range 0.92–2.11. The `intervened` rows are bit-identical to `clean` rows at all five layers (the v1.1 clean/intervened bit-identity bug, see §6 below). Verdicts read off `clean` rows only, per v1.1 acceptance criteria; the destructive finding is condition-independent because it's a raw alignment quantity. ## 3. Artifacts Per-sweep outputs: `results/mesa/phase6b-large-cliff-pair/full-v11/` Files: - `manifest.json` — `cliff_pair: large-v3`, both checkpoints, 64 seeds × 5 layers × 2 conditions per cell. - `axis-b-patch-smoke.csv` — per-(layer, condition, seed) raw rollout metrics with the v1.1 alignment columns. - `axis-b-patch-smoke-aggregate.csv` — per-(layer, condition, direction) aggregates with both metric column families. Earlier sweep artifacts preserved as methodological receipts: - `results/mesa/phase6b-large-cliff-pair/smoke/` — v1 smoke (8 seeds net.9, basin-pref metric), captured the metric pathology that drove the v1.1 amendment. - `results/mesa/phase6b-large-cliff-pair/full/` — v1 full sweep (64 seeds × 5 layers, basin-pref metric only), no verdict called. See [`PHASE6B_SPEC.md`](PHASE6B_SPEC.md) §6.A. - `results/mesa/phase6b-large-cliff-pair/smoke-v11/` — v1.1 smoke (8 seeds net.9, both metrics), surfaced the mean-vs-median disagreement and drove the spec §6 P4-conjunctive-rule refinement. ## 4. GG Verdicts (pre-registered) ### GG6b-loc — **GG6b-loc-D called (Phase 6 v1 protocol doesn't generalize)** The spec required, for any of GG6b-loc-{A, B, C}, that *some* layer clear the conjunctive P4 rule (mean ∧ median ∧ ratio_of_means all ≥ 0.8) in at least one direction without firing the destructive-transfer falsifier. The full-v11 sweep produces: - **Conjunctive P4 clears:** zero cells. No (layer, direction) satisfies mean ≥ 0.8 ∧ median ≥ 0.8 ∧ ratio_of_means ≥ 0.8. Per-statistic clears: mean ≥ 0.8 in **4 cells** (rescue rows for net.1 1.442, net.3 0.928, net.5 0.974, net.7 0.862; net.9 rescue 0.779 misses; all 5 break-row means are near zero), median ≥ 0.8 in **5 cells** (all break rows; break-direction medians 1.285–1.485), ratio_of_means ≥ 0.8 in **5 cells** (all break rows; ratios 1.924–2.103). The mean and median/ratio statistics agree on opposite directions — mean clears on rescue rows where median/ratio do not, median/ratio clear on break rows where mean does not — and the three statistics agree on the conjunction at zero cells. - **Destructive-transfer falsifier:** fires at *all 10 cells* (5 layers × 2 directions). Gap-delta range 0.92–2.11, all well above the 0.5 threshold. Per spec §6, destructive cells contribute to GG6b-loc-D, not A/B/C. With every cell destructive, **GG6b-loc-D is unambiguously called**: the Phase 6 v1 activation-patching mechanism does not transfer between the Large recovery and trough policies. The v1 single-layer locus finding does not generalize. ### GG6b-mech — **not called** Per spec §6, "GG6b-mech is not called at a layer that fired GG6b-loc-D." Every swept layer fired GG6b-loc-D, so the rescue-vs-break asymmetry question never gets a layer to evaluate at. GG6b-mech-{A, B, C} are all undischarged. The smoke-v11 mean-vs-median direction debate (mean 5.42 looking like rescue clears, median 0.05 saying no transfer) is fully resolved by the full-sweep destructive-transfer reading: both directions destroy the recipient's navigation rather than transferring behavior. The asymmetry the mean alone hinted at was small-n outlier noise on top of a destructive substrate. ### GG6b-shape — **deferred** Per spec §6 GG6b-shape-C: "GG6b-loc-A/B did not find a locus, so the substrate-shape question doesn't apply at this layer." There is no locus at any swept layer, so the 5D PCA decomposition question (does the Medium net.7 entangled-subspace finding scale to Large?) is structurally undefined for this cliff pair under this protocol. Phase 6b v2 (or a different cliff-pair selection) picks up the substrate-shape question if a non-destructive layer ever surfaces. ## 5. Why GG6b-loc-D Is Robust (Not Another Metric Artifact) The smoke-v11 audit raised the question: is the v1.1 metric genuinely sound at Large, or is it still producing artifacts? The full-sweep verdict rests on a chain of three independent readings, the first of which is metric-free: 1. **Raw `mean_patched_alignment` (physical, not normalized).** Across all 10 cells, the patched policy terminates at alignment 0.01–0.09 — *below* either reference (collapsed 0.486 / protected 0.912). This is a per-seed mean of a physical quantity; it cannot be exploded by a small denominator. Patching at every swept layer drives the recipient policy into "alignment near zero" territory, which is roughly random-walk terminal performance. 2. **Destructive-transfer falsifier.** Independent of any normalization, `|patched_align − target_align| / |gap_align| > 0.5` fires at all 10 cells with gap-deltas 0.92–2.11. The patched policies are 1–2 gaps away from their intended target on either side. This is the spec §6 falsifier, fired across the entire layer sweep. 3. **Conjunctive P4 disagreement.** The mean, median, and ratio_of_means of `patch_success_align` disagree systematically: rescue rows have positive means (0.78–1.44) but negative medians (-0.45 to -0.49) and negative ratios (-1.00 to -1.11); break rows have near-zero means (-0.16 to 0.08) but positive medians (1.29–1.49) and positive ratios (1.92–2.10). No cell satisfies the conjunction in either direction. This is the v1 P4 discipline inherited verbatim and applied to the v1.1 metric. Three independent readings, three confirmations of the same finding: **the v1 activation-patching protocol does not transfer behavior between the Large mixed_0_99 and mixed_0_97 policies at any layer in the actor's MLP**. The verdict is robust against any one metric being still-broken because it doesn't rest on one metric. ## 6. v3-Consistency Interpretation GG6b-loc-D fits cleanly into the v3 finding shape and sharpens it. The v3 receipt ([`PHASE7_V3_RESULTS.md`](PHASE7_V3_RESULTS.md) §5) established: - The Large U-trough cells (mixed_0_95, mixed_0_97) are `field-coupled, under-budget`, **not** the Medium `reward-coupled` collapse class. - mixed_0_99 (the recovered side at Large) is `field-coupled` with basin-attractor avoidance (GG4-A). - Neither side of the Large cliff pair internalizes the basin (`old_basin_pref` 0.46 / 1.47, both well below the Medium collapse floor 5.56). The Phase 6 v1 mechanism question — "where in the network does the basin attractor live?" — is well-posed at Medium **because** Medium L-Mixed-0.97 has internalized the basin attractor. The cliff between L-Mixed-0.95 (basin-resisting) and L-Mixed-0.97 (basin-internalized) is a real circuit-level boundary, and v1's patching transfers it cleanly at net.7. At Large, *neither side* has internalized a basin attractor. The recovery-vs-trough cliff at Large is something else — both sides are field-coupled, both have low basin-pref, the difference is *navigation competence under high reward weight*, not basin internalization. So the Phase 6 v1 protocol asks the wrong mechanistic question for Large: there is no basin-attractor circuit to transfer because neither side has one. Cross-policy activation injection just breaks the recipient because the policies have different internal representations and no shared mesa-trap substrate to swap. **The Phase 6b verdict is therefore consistent with the v3 verdict and tightens it.** v3 said the Large cliff is not a collapse cliff; 6b says the v1 mechanism-level test that *would have* probed a collapse cliff finds no collapse-cliff-circuit at any layer. The two results agree at substrate level. ## 7. What Phase 6b v1.1 Says, Plainly The earned reading: > At Large capacity (actor parameter_count 4,207,618 ≈ ~4.2M; see > [`PHASE6B_SPEC.md`](PHASE6B_SPEC.md) §4) in the tested shadow-field > navigation family, the v3 cliff between λ=0.99 recovery and > λ=0.97 trough is **not localized to a single MLP layer** under > the Phase 6 v1 activation-patching protocol. At every swept > layer (net.1, net.3, net.5, net.7, net.9), cross-policy > activation injection between the recovery and trough policies > drives the recipient's terminal alignment to ~0 — actively > destroying navigation rather than transferring behavior. This is > consistent with the v3 finding that the Large cliff is not a > collapse-class boundary: neither side internalizes a > basin-attractor, so there is no transferable basin circuit to > probe. Hard scope: - *In-vitro*, mesa shadow-field family only, single-seed Large policies (seed=10000), PPO-specific `vc=0.25` knob. - Phase 6 v1 activation-patching protocol; alternative protocols (e.g. multi-layer co-injection, attention-style key/value editing for non-MLP architectures) are not tested by GG6b-loc-D and could in principle find different mechanism shapes. - The "destructive-transfer" reading describes what happens to the *patched* policy's terminal alignment; it does not claim the unpatched policies are themselves destructive or unstable. - *Not* a claim that Large policies lack mechanistic structure; it's a claim that v1's specific cross-policy-injection test does not surface that structure between this pair. The result is **publishable as a pre-registered falsification**. GG6b-loc-D was written into the spec precisely to catch the "protocol doesn't generalize" outcome cleanly, and the verdict discharges it. ## 8. Open Edges - **clean / intervened bit-identity bug.** Identified in [`PHASE6B_SPEC.md`](PHASE6B_SPEC.md) §6.A on the first sweep (2026-05-18, v1 metric), persists at v1.1 — `intervened` rows are bit-identical to `clean` rows at all five layers across both metrics. v1.1 verdicts read `clean` rows only per acceptance criteria. The bug doesn't change the destructive-transfer reading (which is condition-independent) but should be investigated before any future v2 sweep that wants intervention as a discriminator. The intervention list is built in `run_patched_rollout` and forwarded to the bridge `make` request; the bridge-side `make` handler at the Large path is the likely inspection point. - **Alternative cliff-pair selection.** Phase 6b v1.1 falsifies the v1 protocol for the mixed_0_99 / mixed_0_97 pair specifically. A v2 attempt might select a pair with a genuine basin-pref gap (e.g. signature_terminal_large vs reward_phase3_seg3_large, gap ~6.7) and ask the v1 question of *that* pair — though as the v3 GG5 receipt established, reward_phase3 seg3 is `bootstrap-collapse` (degenerate fixed trajectory, not a competent policy), so cross-policy injection introduces a competence-axis confound. The "right" cliff pair for a v1-analog question at Large may not exist within the v3 envelope. - **Multi-layer co-injection.** v1.1 patches one layer at a time; v1's Medium result also localized to a single layer. A v2 protocol that co-injects activations at multiple layers (e.g. net.7 + net.9 simultaneously) might surface a multi-layer distributed mechanism that the single-layer sweep misses. Not in 6b scope. - **GG6b-shape (5D entangled subspace) at Large.** Structurally undefined for this pair under this protocol; deferred to v2 or to a non-destructive cliff pair if one is identified. ## 9. Cross-References - **Spec source:** [`PHASE6B_SPEC.md`](PHASE6B_SPEC.md) v1 (initial spec) → v1.1 (metric reframe + conjunctive P4 + destructive- transfer falsifier). - **v1 reference receipt:** [`PHASE6_RESULTS.md`](PHASE6_RESULTS.md) v1 (Medium net.7 localization, P→C 0.894/0.944/0.899, C→P 0.934/0.860/0.854) — the v1 conjunctive P4 reading carried forward to v1.1. - **v3 driver:** [`PHASE7_V3_RESULTS.md`](PHASE7_V3_RESULTS.md) §5 (Large cliff is field-coupled-under-budget, not collapse) and §7 (observation-pathway hint that motivated GG6b-mech; not called). - **Harness:** `training/mesa/phase6_probes.py` — `axis-b-smoke` subcommand with `--cliff-pair large-v3` flag. The v1.1 amendment added `terminal_alignment` to `PatchRollout`, the parallel alignment-normalized aggregate columns, and conjunctive console output. ## 10. Versioning - **v1.1 (2026-05-18)** — initial Phase 6b result note. Full-v11 sweep at 64 seeds × 5 layers × 2 conditions, ~13.5 min wall-clock. GG6b-loc-D called (destructive-transfer at every cell, 10 of 10); GG6b-mech not called (per spec §6, mech is not evaluated at destructive layers); GG6b-shape deferred (no locus to decompose). Verdict robustness rests on three independent readings: raw `mean_patched_alignment` (physical), destructive-transfer falsifier (independent of normalization), conjunctive P4 rule on the v1.1 alignment-normalized metric. v3-consistency analysis in §6 frames GG6b-loc-D as a v3-confirming finding rather than a v3-falsifying one: the Phase 6 v1 protocol asks the wrong mechanistic question for a pair where neither side internalizes a basin-attractor. clean / intervened bit-identity bug carried forward as an open edge; does not change this verdict.