# Mesa Phase 6b - Large-Tier Activation Patching on the v3 Cliff Pair This document is the implementation-grade spec for Phase 6b, which extends Phase 6 v1 (Medium cliff-pair activation patching localized to `net.7`) to a Large-tier cliff pair surfaced by [`PHASE7_V3_RESULTS.md`](PHASE7_V3_RESULTS.md). Phase 6b is a *Large sibling* of Phase 6 v1, not a successor; the v1 22-cell net.7 result on Medium is unchanged. Where this spec and [`PHASE6_SPEC.md`](PHASE6_SPEC.md) disagree, the v1 spec wins for patching semantics. Where both are silent, this spec is authoritative for the Large extension. > **v1.1 amendment (2026-05-18) — metric reframe.** The v1 Phase 6b > spec inherited the v1 patch_success metric, which normalizes on > the `old_basin_pref` gap between protected and collapsed cells. > That metric is well-defined on the Medium v1 cliff pair (collapse > gap ~5+), but on the Large v3 cliff pair the gap is ~1.0 and the > normalized "fraction" is unbounded — patching swings of ±6 > produce `patch_success` values in the hundreds. The v1 P4 > threshold (0.8) is meaningless on this scale. The first full > sweep at the broken v1 metric is preserved as a methodological > receipt but **no GG verdict was called against the broken metric** > (see §6.A below). > > v1.1 introduces a parallel `patch_success_align` metric > normalized on the **terminal-alignment gap** between protected > (mixed_0_99 align 0.912) and collapsed (mixed_0_97 align 0.486), > ≈ 0.426. The alignment gap is the actual behavioral separation > the v3 receipt established between recovered and trough cells. > §6 predictions are rewritten against this v1.1 metric; > alignment-patch_success is the canonical v1.1 reading. v1's > basin-pref-based patch_success is kept as a transparency column > in the CSV but is not used for verdict calls. ## 1. Decision Lock Six pinned calls: - **One cliff pair, mirroring v1's design.** The primary cliff pair is **mixed_0_99 vs mixed_0_97** at Large with `--value-coef 0.25`, both at 10M env-steps, `seed_start = 10000`. mixed_0_99 is `field-coupled` (the "recovered" side); mixed_0_97 is `field-coupled, under-budget` (the v3 trough side). Adjacent in λ, one cell on each side of the v3 behavioral boundary — direct Phase 6 v1 analog of the Medium 0.95 / 0.97 pair. - **Inherit Phase 6 v1 patching semantics exactly.** Same activation- capture-and-inject mechanism via `register_forward_hook`, same per-step injection at the layer of interest, same `patch_success` metric, same clean / intervened conditions. Phase 6b is a v1 *application* at a new tier, not a v1 redesign. - **Layer of interest is `net.9` (analog of Medium `net.7`).** Large actor architecture is depth=5, hidden_size=1024 (vs Medium depth=4, hidden_size=256). Final hidden Tanh activation index in the `Sequential` module list is **`net.9`** for Large (`0..9 = L,T,L,T, L,T,L,T,L,T`; `net.10` is the policy head Linear). All five Tanh layers — `net.1`, `net.3`, `net.5`, `net.7`, `net.9` — are run in the layer sweep, with `net.9` as the pre-registered terminal candidate matching v1's net.7 verdict. - **Pre-registered mechanism hypothesis from v3 §7.** The Phase 7 v3 side-finding (observation-channel response at trough cells ~3× the canonical signature controller, observation-channel `old_basin_pref` ~7× higher) gives Phase 6b a sharp entry hypothesis: **the field-coupled-under-budget pathway is over-leveraged on position-observation input.** If true, the patched cliff effect should depend on layers downstream of where position-observation is most strongly read. This is named explicitly in §6 (GG6b-mech) and is independent of the GG6b-localization prediction. - **Single-seed slate, mirroring v1.** Both cliff-pair cells are single-seed at `seed_start = 10000` (carried over from v2). v3.1 triangulation is a separate work item; Phase 6b does not own it. This is a tractability decision, not a methodology preference; Phase 6b v2 (or a v3.1 retrofit) would extend to multi-seed. - **Harness extension is mandatory, scope is bounded.** The Phase 6 v1 harness (`training/mesa/phase6_probes.py`) hard-codes Medium `CLIFF_PROTECTED` / `CLIFF_COLLAPSED` at module top-level. Phase 6b adds a Large pair as a parallel constant and exposes a `--cliff-pair {medium-v1,large-v3}` flag with `medium-v1` as the default. v1 code path is preserved unchanged. No other harness modifications are authorized in Phase 6b. ## 2. Purpose [`PHASE7_V3_RESULTS.md`](PHASE7_V3_RESULTS.md) §5 introduced a new traceability class — `field-coupled, under-budget` — and named the behavioral boundary at Large between recovery (λ=0.99) and trough (λ=0.97). The v3 receipt established that this is *not* the Medium collapse class: the trough policies still read the signature. The mechanistic question Phase 6b asks: **does the cliff between field-coupled and field-coupled-under-budget at Large localize to a specific layer, the way Medium's protected-vs-collapsed cliff localized to `net.7`?** Three named hypotheses (§6 below): - **GG6b-localization** — the cliff localizes at Large net.9 with patch_success crossing the v1 P4 threshold (0.8) in at least one direction. - **GG6b-mech** — the v3 §7 observation-sensitivity hypothesis holds mechanistically: the under-budget pathway is dominated by position- observation features at some layer downstream of input. - **GG6b-substrate-shape** — the entangled-5D substrate finding from Phase 6 v3 (97.4% variance, 5 PCs, K=5 minimum) is *not* the same shape at Large net.9. Hidden dim is 4× larger; the compression ratio may differ. The third hypothesis is pre-registered as a v2-style falsifier rather than a confirm-or-fail dichotomy. ## 3. Target Cliff Pair Two checkpoints (`.pt` paths relative to repo root): | side | tier | λ | v3 traceability label | checkpoint | | --- | --- | ---: | --- | --- | | recovered | Large | 0.99 | `field-coupled` (probe-confirmed, basin-attractor avoidance) | `results/mesa/phase7v2-large-cliff-subset/mixed_0_99_vc0_25/checkpoints/mixed_ppo_phase3_lambda_0_9_large_seed_0_mixed_0_99_vc0_25.pt` | | trough | Large | 0.97 | `field-coupled, under-budget` | `results/mesa/phase7v2-large-cliff-subset/mixed_0_97_vc0_25/checkpoints/mixed_ppo_phase3_lambda_0_9_large_seed_0_mixed_0_97_vc0_25.pt` | Both checkpoints are verified present at spec-write time. Both were trained with `--value-coef 0.25` at 10M env-steps. The Phase 6 harness loads `.pt` checkpoints (not `.policy.json`) via `policy_from_checkpoint(load_checkpoint(...))`; both target paths are in the same on-disk format as the Medium v1 cliff pair. ## 4. Architecture Audit `training/mesa/policy.py`: ```python CAPACITY_CONFIGS: dict[str, PolicyConfig] = { "small": PolicyConfig(tier="Small", hidden_size=64, depth=2), "medium": PolicyConfig(tier="Medium", hidden_size=256, depth=4), "large": PolicyConfig(tier="Large", hidden_size=1024, depth=5), } ``` Actor architecture is `nn.Sequential` of alternating `Linear` and `Tanh` modules, with a final `Linear` projection to `act_dim=2`. Named-module layout per tier (`policy.named_modules()` returns keys under the `net` attribute): | index | Medium (depth=4) | Large (depth=5) | | ---: | --- | --- | | net.0 | Linear (6 → 256) | Linear (6 → 1024) | | net.1 | Tanh | Tanh | | net.2 | Linear (256 → 256) | Linear (1024 → 1024) | | net.3 | Tanh | Tanh | | net.4 | Linear (256 → 256) | Linear (1024 → 1024) | | net.5 | Tanh | Tanh | | net.6 | Linear (256 → 256) | Linear (1024 → 1024) | | net.7 | **Tanh** (final hidden, v1 cliff locus) | Tanh | | net.8 | Linear (256 → 2) | Linear (1024 → 1024) | | net.9 | — | **Tanh** (final hidden, Large analog) | | net.10 | — | Linear (1024 → 2) | Phase 6 v1 found the cliff localized to Medium **net.7** (final hidden Tanh before the policy head). The Large structural analog is **net.9** (final hidden Tanh, 1024-dim vs Medium's 256-dim). Actor parameter counts (from the canonical `.policy.json` exports for each tier): | tier | hidden_size | depth | actor parameter_count | | --- | ---: | ---: | ---: | | Medium | 256 | 4 | 199,682 | | Large | 1024 | 5 | 4,207,618 | Large is **~21× Medium** in actor parameter count (not the rough 20× implied by hidden_size² × depth alone — the depth=5 vs depth=4 step adds another Linear layer on top of the 4× hidden-dim scaling). ## 5. Harness Extension `training/mesa/phase6_probes.py` requires the following bounded changes: **v1 (landed 2026-05-18):** - Added `CLIFF_PROTECTED_LARGE` and `CLIFF_COLLAPSED_LARGE` PolicySpec constants pointing to the §3 checkpoints. - Added `--cliff-pair` to the `axis-b-smoke` argparser with choices `{medium-v1, large-v3}` and default `medium-v1`. - In `run_axis_b_patch`, branched on `args.cliff_pair` to pick which pair to load. v1 medium-v1 code path preserved unchanged. - Updated manifest emission to use the selected `protected_spec` / `collapsed_spec` rather than the hardcoded v1 constants, with a `cliff_pair` field recording the selection. **v1.1 (queued, not yet landed):** - Add `terminal_alignment` to `PatchRollout` (computed from rollout summary, mirroring `mesa-intervention-battery.mjs` semantics). - Add `safe_patch_success_align` function paralleling `safe_patch_success` but normalized on alignment gap (denominator is `collapsed_align − protected_align` for P→C; symmetric for C→P). Same `nan` guard on small denominators. - Compute both metrics per (layer, condition, seed); write both parallel column sets to the aggregate CSV. New columns: `mean_patch_success_align`, `median_patch_success_align`, `patch_success_align_ratio_of_means`, `mean_protected_alignment`, `mean_collapsed_alignment`, `mean_patched_alignment`. v1 columns (`mean_patch_success`, etc.) retained for transparency. - Investigate the `clean` / `intervened` bit-identity observed in the §6.A first-sweep receipt before re-running. The intervention list is built in `run_patched_rollout` and forwarded to the bridge `make` request; the bridge's `make` handler is the likely inspection point. Until clean/intervened produces meaningful divergence, v1.1 GG6b verdicts use only the `clean` rows. No other changes. The harness's `register_forward_hook` mechanism is already layer-agnostic — it accepts any name in `dict(policy.named_modules())`. A separate `axis-b-large-smoke` subcommand alias is not added; the existing `axis-b-smoke` command takes the new flag. ## 6. Pre-Registered Predictions (v1.1) **Metric:** all GG6b predictions in v1.1 read against `patch_success_align`, the alignment-normalized parallel metric: ``` patch_success_align (P→C direction) = (collapsed_align − patched_align) / (collapsed_align − protected_align) patch_success_align (C→P direction) = (protected_align − patched_align) / (protected_align − collapsed_align) ``` where `*_align` is `mean_on_terminal_alignment` from the rollout output. The denominator is the alignment gap (~0.426 for the Large cliff pair, see preamble) — by construction the metric is bounded [~0, 1] for patched outputs that interpolate between source and target, > 1 if patching overshoots toward the target, < 0 if it diverges away. **P4 verdict rule (conjunctive, inherited from v1):** "clears P4" at a (layer, condition, direction) means **all three** of `mean_patch_success_align`, `median_patch_success_align`, and `patch_success_align_ratio_of_means` clear 0.8. This mirrors v1's published P4 reading (PHASE6_RESULTS.md net.7: P→C `0.894 / 0.944 / 0.899`, C→P `0.934 / 0.860 / 0.854` — all three statistics clearing in both directions). Mean alone is not sufficient: the v1.1 smoke (2026-05-18, n=8) showed mean values can be outlier-dominated and disagree with median by orders of magnitude (mean 5.42 vs median 0.05) without indicating a real transfer. **Destructive-transfer falsifier (v1.1, new).** Independent of the P4 conjunctive rule, a (layer, condition, direction) is classified as **destructive** rather than transferring when `|mean_patched_alignment − target_alignment| / mean_baseline_gap_align > 0.5`, where `target_alignment` is `mean_protected_alignment` for P→C (rescue) and `mean_collapsed_alignment` for C→P (break). This catches the case where patching drives the policy *away from both* references — e.g., toward `alignment ≈ 0` — rather than interpolating between source and target. A destructive layer contributes to GG6b-loc-D regardless of how the conjunctive P4 statistics look at that layer; it does not contribute to GG6b-loc-A/B/C. ### GG6b-localization (v1.1) — net.9 is the Large cliff locus Pre-registered: **at least one of the two patch directions clears the conjunctive P4 rule (mean ∧ median ∧ ratio_of_means all ≥ 0.8) at net.9 without triggering the destructive-transfer falsifier**, and no earlier Large layer (net.1, net.3, net.5, net.7) clears it in either direction. - **GG6b-loc-A (confirm)** — net.9 clears the conjunctive P4 rule in at least one direction, earlier layers do not, and the locus layer is not destructive. The Large recovery-vs-trough cliff localizes structurally analogous to Medium net.7's basin cliff. - **GG6b-loc-B (falsify — earlier locus)** — some earlier layer (net.1, net.3, net.5, or net.7) clears the conjunctive P4 rule without being destructive, and net.9 does not. The Large cliff lives upstream of the final hidden, consistent with the v3 §7 position-observation pathway hint. - **GG6b-loc-C (falsify — no locus)** — no layer clears the conjunctive P4 rule in either direction without being destructive. The Large recovery-vs-trough cliff is distributed rather than localized. - **GG6b-loc-D (falsify — Phase 6 v1 protocol doesn't generalize)** — the patching mechanism itself doesn't transfer to Large at this pair. Triggered by *any* of the following at *most* of the swept layers: (a) `patch_success_align` mean and median disagree by orders of magnitude (mean / median ratio > 10 or < -10), (b) the destructive-transfer falsifier fires (patched alignment falls outside the protected↔collapsed band by more than half a gap-width in either direction), (c) the median is near zero while raw `mean_patched_alignment` differs from both reference alignments by more than the gap (signature of "patching destroys navigation rather than transferring"). The v1.1 smoke (n=8) at net.9 partially hit (a) and (b); the full sweep determines whether this is a net.9-specific artifact or a methodology limit at Large. ### GG6b-mech (v1.1) — observation-pathway dependence **Harness naming clarification** (precondition for the prediction): in `phase6_probes.py`, `patched_protected_to_collapsed` (P→C in CSV columns) is the **collapsed policy with protected activations injected at the patched layer** — the *rescue* direction (move the trough toward recovered behavior by giving it the recovered side's activations). `patched_collapsed_to_protected` (C→P) is the **protected policy with collapsed activations injected** — the *break* direction. From v3 §7: trough cells show ~3× the observation-channel response of the canonical signature controller, suggesting the trough has signature responsiveness but is missing the navigation circuit the recovered side has. Pre-registered against `patch_success_align`: **rescue (P→C) is decisively the larger direction.** Injecting mixed_0_99 activations into the mixed_0_97 policy should restore some of the missing navigation more than injecting mixed_0_97 activations into mixed_0_99 disrupts it. - **GG6b-mech-A (confirm)** — P→C (rescue) clears the conjunctive P4 rule, and C→P (break) does not (or P→C clears more cleanly: mean delta ≥ 0.2 across all three statistics, and C→P fails at least one statistic of the conjunction). Confirms the observation- pathway / signature-pathway asymmetry hinted at by v3 §7. The trough has a missing-piece that the locus layer carries and that's restorable by injecting recovered-side activations. - **GG6b-mech-B (falsify)** — symmetric patch behavior. Both directions clear the conjunctive P4 rule at comparable magnitudes (|P→C − C→P| < 0.2 across all three statistics); the v3 §7 observation-sensitivity finding does not translate to a directional patching asymmetry. The cliff is bidirectional in alignment, like Medium v1 was bidirectional in basin-pref. - **GG6b-mech-C (falsify — inverted asymmetry)** — C→P (break) clears the conjunctive P4 rule and P→C (rescue) does not. This would mean the recovered side is *fragile* to collapsed activation injection while the trough is *unresponsive* to recovered activation injection — a qualitatively different story than v3 §7 implies, and worth its own GG-line so the result isn't conflated with GG6b-mech-A. **GG6b-mech is not called at a layer that fired GG6b-loc-D (destructive-transfer falsifier).** Both directions being destructive at a layer doesn't tell us about the rescue/break asymmetry; it tells us the patching mechanism doesn't transfer there. The v1.1 smoke (n=8) at net.9 is a concrete illustration: under the mean alone, P→C looked like rescue-clears (5.42 ≥ 0.8) and C→P didn't (-4.54), which would point at GG6b-mech-A — but median (P→C 0.05, C→P 1.30), ratio_of_means (-1.00, 2.24), and raw `mean_patched_alignment` (P→C 0.10, C→P 0.007, neither near either target) together show the patching is destructive in both directions, not asymmetric rescue. The full sweep determines whether some non-net.9 layer is non-destructive and asymmetric. ### GG6b-substrate-shape — 5D entangled subspace at Large? Phase 6 v3 found that Medium net.7's basin-attractor circuit compresses to **5 PCA components capturing 97.4% of variance** and reproducing the full-layer patch effect (51× compression from 256 dims). At Large net.9, hidden_dim is 4× larger; the same compression shape may or may not hold. Reads against `patch_success_align`. - **GG6b-shape-A (confirm — same shape)** — 5–10 PCA components at Large locus layer reproduce the patch effect from §3, at variance capture ≥ 90%. The "small handful of generators, irreducibly entangled" finding generalizes across capacity. - **GG6b-shape-B (falsify — wider entanglement)** — a substantially larger PCA basis is required (≥ 30 components to reach 90% variance, or ≥ 20 to reproduce the patch effect). The entangled-substrate generator count scales with capacity. - **GG6b-shape-C (defer)** — GG6b-loc-A/B did not find a locus, so the substrate-shape question doesn't apply at this layer. Phase 6b v2 picks up wherever the locus lives. GG6b-shape is deferred to a second-pass after GG6b-loc is called. Phase 6b v1.1 lands GG6b-loc and GG6b-mech; the PCA decomposition is a v1.2 follow-on if (and only if) GG6b-loc finds a locus. ## 6.A First-Sweep Receipt (v1 metric, no verdict) For methodological transparency, the first full Phase 6b sweep (2026-05-18, ~12 min wall-clock, 64 seeds × 5 layers, v1 basin-pref-based `patch_success`) is preserved at `results/mesa/phase6b-large-cliff-pair/full/`. **No GG verdict was called against this sweep** — the basin-pref-based metric is unbounded for the Large cliff pair (basin-pref gap ~1.0, patching swings ±6 → values in the hundreds) and the v1 P4 threshold is not meaningful in that range. Two findings from the first sweep ARE worth noting independent of the metric question, because they shape v1.1 expectations: - **`clean` and `intervened` rows are bit-identical to 16 sig figs across all 5 layers.** At Medium v1 these diverged because basin- position intervention shifted the collapsed-side policy. At Large, neither side of the cliff pair internalizes the basin (v3 GG4-A bp_obp 0.46 / 1.47), so the intervention is *expected* to have near-zero behavioral effect — but bit-identity is too clean, suggesting the intervention either isn't being applied at the env side or its effect is masked by something upstream of the policy. **This is a v1.1 prerequisite to investigate before the re-run** (see §5 v1.1 bullet). - **First-sweep basin-pref `patch_success` mean values are in the hundreds with negative-direction medians.** Even accepting the metric is broken, this is preliminary evidence pointing away from GG6b-loc-A (net.9 as a clean Medium-net.7 analog) — patching net.9 swings the basin-pref outcome wildly rather than transferring cleanly. v1.1 will tell whether this is metric-specific noise (GG6b-loc-C or -D under v1's metric, possibly different under v1.1's) or substantive. ## 7. Acceptance Criteria (v1.1) Phase 6b v1.1 is complete when: - ✓ The harness `--cliff-pair large-v3` flag is implemented (v1, landed) *and* the alignment-patch_success metric is computed and written to the aggregate CSV (v1.1, landed — see §5 v1.1 bullet). - ⏳ The `clean` / `intervened` bit-identity bug is investigated and either fixed or its non-effect explained mechanistically. v1.1 verdicts default to `clean` rows only until the bug is resolved. *Status:* open; doesn't change the destructive-transfer verdict in v1.1 (which is condition-independent). - ✓ A v1.1 smoke (8 seeds, `--layer net.9`) lands with sensible `patch_success_align` magnitudes (bounded, not in the hundreds). *Smoke-v11 landed 2026-05-18 ~15s wall-clock; metric bounded (single digits); see §13 versioning for the per-statistic numbers that drove the P4-rule sharpening.* - ✓ The v1.1 full layer sweep (`--layers net.1,net.3,net.5,net.7,net.9` at 64 seeds) has run and the parallel-column CSVs are written. *Full-v11 landed 2026-05-18 ~13.5 min wall-clock.* - ✓ Each of GG6b-loc-{A,B,C,D} is called against `patch_success_align`; if A or B confirms, the locus layer is named. *Called: GG6b-loc-D (destructive-transfer at all 10 cells).* - ✓ GG6b-mech is called against the locus layer (if any). *Not called per spec §6 — mech is not evaluated at destructive layers; every layer fired the falsifier.* - ✓ [`PHASE6B_RESULTS.md`](PHASE6B_RESULTS.md) is written with the v1.1 verdicts and a back-reference to the §6.A first-sweep receipt for methodological completeness. ## 8. Compute Envelope Reference point: Phase 6 v1 axis-b at Medium swept the four post-Tanh hidden layers (`net.1`, `net.3`, `net.5`, `net.7`; the final Linear head is excluded from the canonical v1 layer sweep). With the harness's default `--conditions clean,intervened` and four rollouts per inner iteration (clean_protected, clean_collapsed, patched_protected_to_collapsed, patched_collapsed_to_protected), the v1 step count is: ``` 64 seeds × 4 layers × (2 patch directions × 2 conditions) × ~200 steps ≈ ~205k env-steps ``` at Medium hidden_size 256, bridge-bound, ~2–3 hours operator wall- clock. Phase 6b at Large differs from v1 along two axes: - **Step count is ~1.25× v1** — 64 seeds × **5 layers** (net.1, net.3, net.5, net.7, net.9 — one more Tanh layer than v1's four, because Large is depth=5) × 2 patch directions × 2 conditions × 200 steps ≈ **~256k env-steps**. - **Per-step cost is higher at Large** — actor parameter count is ~21× Medium (§4), so the policy forward pass is heavier even though the env-bridge round-trip is unchanged. The bridge round- trip dominates per-step cost at Medium (per Phase 7 v2 §14.4); whether it still dominates at Large during *inference-only* patching is not measured yet. The smoke step (§9) is the calibration measurement — wall-clock on smoke × ~32 gives the full estimate. Pre-measurement estimates (to be confirmed by smoke): - **Smoke** (8 seeds × 1 layer × 2 patch directions × 2 conditions × 200 steps ≈ 6.4k step-equivalents): **~5–10 minutes**. - **Full** (~256k step-equivalents): **~2.5–4 hours** if bridge- bound holds at Large; longer if torch forward becomes a meaningful per-step share. Single-threaded, CPU-only, no GPU. This is **larger** than a single Phase 7 v3 intervention battery (~60–70 minutes) — Phase 6b's step count is ~5× v3's and the per-step cost is comparable or slightly higher. Treat Phase 6b as a multi- hour session, not a same-budget extension of v3. ## 9. Staged Commands (operator) The v1.1 canonical commands as actually executed (wall-clock measured, not estimated): ```powershell # repo root: # SMOKE — Large cliff pair, single layer net.9, 8 seeds. # Measured wall-clock 2026-05-18: ~15 seconds. python -m training.mesa.phase6_probes axis-b-smoke ` --cliff-pair large-v3 ` --layer net.9 ` --seeds 8 ` --out results/mesa/phase6b-large-cliff-pair/smoke-v11 # FULL — Large cliff pair, layer sweep, 64 seeds. # Measured wall-clock 2026-05-18: ~13.5 minutes (10:32 → 10:46). python -m training.mesa.phase6_probes axis-b-smoke ` --cliff-pair large-v3 ` --layers net.1,net.3,net.5,net.7,net.9 ` --seeds 64 ` --out results/mesa/phase6b-large-cliff-pair/full-v11 ``` The smoke-v1 (basin-pref-only, before the v1.1 harness change) and full-v1 receipts are preserved alongside at `.../smoke/` and `.../full/` respectively; see §6.A. Pre-measurement wall-clock estimates in §8 (~5–10 min smoke, ~2.5–4 h full) were conservative by a factor of ~10× on both — bridge round-trip dominates at Large inference-only patching the same way it does at Medium, and the torch forward at hidden_size=1024 is not the bottleneck. (The subcommand keeps the `axis-b-smoke` name from Phase 6 v1 even though the full run is not a smoke; renaming the subcommand is out of Phase 6b scope.) A `Show-PatchSummary` PowerShell helper will be staged in chat alongside the operator commands, mirroring the `Show-IntervSummary` helper from Phase 7 v3. ## 10. Outputs ``` results/mesa/phase6b-large-cliff-pair/ smoke/ axis-b-patch-smoke.csv axis-b-patch-smoke-aggregate.csv full/ axis-b-patch-smoke.csv axis-b-patch-smoke-aggregate.csv ``` CSV column names inherit from Phase 6 v1: `mean_patch_success_protected_to_collapsed`, `mean_patch_success_collapsed_to_protected`, `patch_success_ratio_of_means`, `median_patch_success`, etc. ## 11. Cross-References - **v1 source spec:** [`PHASE6_SPEC.md`](PHASE6_SPEC.md) v1.6. Patching semantics, axis-b harness structure, P4 threshold, ratio-of-means metric inherited unchanged. - **v1 reference receipt:** [`PHASE6_RESULTS.md`](PHASE6_RESULTS.md) v1 (Medium net.7 localization) and [`PHASE6_V31_RESULTS.md`](PHASE6_V31_RESULTS.md) (5D entangled subspace finding) for GG6b-shape calibration. - **v3 driver:** [`PHASE7_V3_RESULTS.md`](PHASE7_V3_RESULTS.md) §5 introduces the `field-coupled, under-budget` class; §7 hands off the observation-pathway hypothesis underlying GG6b-mech. - **Cliff-pair lineage:** [`PHASE7_V2_RESULTS.md`](PHASE7_V2_RESULTS.md) §2 envelope (target checkpoint provenance). - **Architecture source of truth:** `training/mesa/policy.py` `CAPACITY_CONFIGS` constant. - **Harness:** `training/mesa/phase6_probes.py` `axis-b-smoke` subcommand (extended with `--cliff-pair` flag per §5). ## 12. Non-Goals Phase 6b does not own: - Cross-tier patching (Medium → Large or vice versa). The cliff pair is within-tier. - Held-out policy pair generalization studies analogous to Phase 6 v3.5–v3.7 (J1/J2 work). Phase 6b is a single-pair localization pass; cross-policy generalization at Large is a future v2 follow-on once GG6b-loc is called. - Probe-slate (Phase 3) at Large. Open edge from v3 §8; not in 6b scope. - Seed-shift triangulation of the cliff-pair cells. Open edge for v3.1; not in 6b scope. ## 13. Versioning - **v1 (2026-05-18, spec)** — initial Phase 6b spec. Single Large cliff pair (mixed_0_99 vs mixed_0_97 at vc=0.25, 10M, seed=10000), layer sweep across net.1/3/5/7/9 with net.9 pre-registered as the Medium-net.7 analog. Three named GG predictions: GG6b-loc (localization), GG6b-mech (observation-pathway asymmetry from v3 §7), GG6b-shape (5D entanglement at Large net.9; deferred to v1.1). Compute envelope: ~1.5–2.5 hours full, ~3–5 minutes smoke. - **v1 first sweep (2026-05-18, executed)** — harness extension landed, full sweep ran in ~12 min wall-clock (smoke pace was 15s, the §8 multi-hour estimate was conservative). **No GG verdict called.** Metric defect surfaced: v1's basin-pref-based `patch_success` is unbounded on the Large cliff pair (gap ~1.0 vs Medium's ~5.5), producing values in the hundreds. The first sweep is preserved at `results/mesa/phase6b-large-cliff-pair/full/` as a methodological receipt; see §6.A. - **v1.1 (2026-05-18, amendment)** — metric reframe. Predictions rewritten against `patch_success_align`, an alignment-gap normalized parallel metric (denominator: collapsed_align − protected_align ≈ 0.426 for the Large pair). v1's basin-pref metric retained as a transparency column but not used for verdicts. New falsification branch added: GG6b-loc-D for "Phase 6 v1 patching protocol doesn't generalize to Large in either metric." `clean` / `intervened` bit-identity flagged as v1.1 prerequisite to investigate before re-run. - **v1.1 (2026-05-18, harness landed)** — `terminal_alignment` added to `PatchRollout` and captured from `info.metrics.terminalAlignment` on `done=true` (bridge already exposes it; no bridge change needed). `safe_patch_success` reused with alignment values (the formula is metric-agnostic; semantically `safe_patch_success_align`). Parallel CSV columns added: `*_terminal_alignment` per side and patch direction, `patch_success_align_*` per direction, `baseline_gap_align_*`, and aggregate `mean_patch_success_align`, `median_patch_success_align`, `patch_success_align_ratio_of_means`, `mean_baseline_gap_align`, `mean_protected_alignment`, `mean_collapsed_alignment`, `mean_patched_alignment`. Console summary rewritten to lead with `ps_align_*` and trail with the legacy basin-pref metric in brackets. Spec §6 GG6b-mech wording corrected: the rescue direction is P→C (in harness CSV naming), not C→P as the prior amendment text said; GG6b-mech-A now predicts P→C `patch_success_align` decisively higher than C→P, plus a new GG6b-mech-C falsifier branch for inverted asymmetry. `clean` / `intervened` bit-identity investigation still queued. - **v1.1 (2026-05-18, P4 verdict rule sharpened post-smoke)** — Smoke-v11 (8 seeds, net.9) showed alignment-normalized metric is *bounded* (single digits vs hundreds, falsifying GG6b-loc-D's unbounded-metric subcondition) but *mean values are outlier- dominated* with small-n (P→C mean 5.42 vs median 0.05; C→P mean -4.54 vs median 1.30) and raw `mean_patched_alignment` for both directions fell near 0 rather than near either reference target. Spec §6 verdict rule updated: P4 reading is now explicitly conjunctive (mean ∧ median ∧ ratio_of_means all ≥ 0.8), mirroring v1's published net.7 receipt (P→C 0.894/0.944/0.899). A new destructive-transfer falsifier is introduced, classifying any (layer, condition, direction) where patched alignment falls more than half a gap-width outside the protected↔collapsed band; such layers contribute to GG6b-loc-D rather than to A/B/C. GG6b-mech is explicitly not called at destructive layers. Reasoning anchored to the v1.1 smoke results directly so future readers can recapitulate the decision. - **v1.1 (2026-05-18, full-v11 sweep executed, result note filed)** — Full sweep at 64 seeds × 5 layers × 2 conditions, ~13.5 min wall- clock. All 10 cells fired the destructive-transfer falsifier (gap-delta range 0.92–2.11); zero cells cleared the conjunctive P4 rule. **GG6b-loc-D called** (Phase 6 v1 protocol doesn't generalize to Large for this cliff pair); GG6b-mech not called; GG6b-shape deferred. Receipt at [`PHASE6B_RESULTS.md`](PHASE6B_RESULTS.md). The verdict rests on three independent readings (raw `mean_patched_alignment`, destructive-transfer falsifier, conjunctive P4) and is consistent with the v3 finding that the Large recovery/trough cliff is not a basin-attractor collapse cliff — there is no transferable basin circuit to probe between two field-coupled policies.