# Mesa Phase 7 v2 - Large-Tier Cliff-Subset Result Note This document records the Phase 7 v2 DOWN-SCOPE result note for the Large-tier cliff-subset batch defined in [`PHASE7_SPEC.md`](PHASE7_SPEC.md) §14 (DOWN-SCOPE addendum) and the Path B hparam investigation in [`PHASE7_V2_PATH_B_HPARAM_SPEC.md`](PHASE7_V2_PATH_B_HPARAM_SPEC.md). Phase 7 v2 is a sibling to Phase 7 v1 ([`PHASE7_RESULTS.md`](PHASE7_RESULTS.md)). It does not regenerate the v1 22-policy classification and does not change v1 verdicts. v2 extends the operating envelope to a third capacity tier and surfaces failure modes the v1 envelope did not see. Status: Phase 7 v2 cliff-subset batch **complete**. Six Large-tier policies evaluated at 32 seeds, with `seed_start=10000` as the canonical seed and `seed_start ∈ {20000, 30000}` for Stage 3 triangulation on `λ=0.90`. PPO training-time hparam pin: `--value-coef 0.25` for all L-Mixed and L-Reward cells (signature_terminal uses defaults). ## 1. Summary Three findings update the gravity claim's earned wording for capacity scaling: 1. **Forward L-Signature canonical extends to Large.** signature_terminal at Large reaches success 0.750 / alignment 0.993 / mean_steps 117 at 5M env-steps. Capacity does not break the forward shape; it holds at Small (37/64, 0.963), Medium (64/64, 0.999), and Large (24/32, 0.993). 2. **Medium single-cliff broadens to a Large U-trough.** The v1 Medium cliff at `λ ≈ 0.953` becomes, at Large, a broader trough at `λ ∈ {0.95, 0.97}` (success 0.031–0.094, alignment 0.506–0.599) bracketed by a field-coupled `λ=0.90` cell (0.531 / 0.934) and a *recovered* `λ=0.99` cell (0.406 / 0.885) with only a 1% signature anchor. 3. **At Large, signature anchoring is constitutive of learnability.** Pure-reward `λ=1.00` at Large bootstrap-fails across three consecutive segments at alignment 0.003 / success 0 / mean_steps 200. With a 1% signature mixture (λ=0.99) the controller recovers. At Large the signature is no longer merely protective against basin internalization — it is the seed the value function needs to start learning at all. ## 2. Envelope Large tier, 32-seed eval @ `seed_start=10000` except where noted: | λ (reward weight) | training spec | env-steps | success | mean alignment | mean_steps | verdict | | ---: | --- | ---: | ---: | ---: | ---: | --- | | 0.00 (signature_terminal) | seg3 baseline, defaults | 5M | 0.750 | 0.993 | 117.0 | converged | | 0.90 | vc=0.25, seed=10000 | 10M | 0.531 | 0.934 | 145.1 | borderline hold (adopted) | | 0.95 | vc=0.25, seed=10000 | 10M | 0.094 | 0.599 | 193.3 | under-budget | | 0.97 | vc=0.25, seed=10000 | 10M | 0.031 | 0.506 | 195.8 | trough | | 0.99 | vc=0.25, seed=10000 | 10M | 0.406 | 0.885 | 163.2 | converged (1% anchor) | | 1.00 (reward_phase3) | vc=0.25, segmented chain | ~5M | 0.000 | 0.003 | 200.0 | bootstrap-failed | Stage 3 triangulation for `λ=0.90` across three seeds: `seed=10000` 0.531, `seed=20000` 0.406 (strict ±0.10 fail by 0.025), `seed=30000` 0.625. Two of three within band; the borderline-hold verdict is adopted with a protocol-tightening note for v3. ## 3. Artifacts Raw runs: `results/mesa/phase7v2-large-cliff-subset/` Key directories: - `mixed_0_90_pathb_vc0_25/` — adopted candidate run @ seed=10000 - `mixed_0_90_pathb_vc0_25_seed20000/` — Stage 3 first triangulation - `mixed_0_90_pathb_vc0_25_seed30000/` — Stage 3 second triangulation - `mixed_0_95_vc0_25/` — λ=0.95 cell - `mixed_0_97_vc0_25/` — λ=0.97 cell (trough) - `mixed_0_99_vc0_25/` — λ=0.99 cell (1% anchor recovery) - `reward_phase3_vc0_25_chain/seg{1,2,3}/` — pure-reward bootstrap- failure chain - `mixed_0_90_pathb_lr1e4/` — Path B candidate 1, rejected at Stage 1 (mode collapse: success 0.219, alignment 0.832, log_std -1.319) The signature_terminal baseline lives at `results/mesa/phase7v2-large-conv-10m/seg3/`. ## 4. Three Large-tier failure modes The v2 envelope separates three distinct PPO failure modes at Large that the v1 Medium envelope did not surface: 1. **Critic destabilization** (mixed_0_90 default hparams). Pre-Path-B ext3 chain: entropy 2.29 → 2.76 monotone-climb, `log_std` -0.28 → -0.04 expansion, value_loss spike to 700+, mean_reward drop -1.83 → -2.09. Default-hparam Large L-Mixed at λ=0.90 destabilizes by ~8M env-steps. Fixed by `--value-coef 0.25`. 2. **Mode collapse** (Path B candidate 1, `--lr 1e-4`). Stage 1 fail at success=0.219, alignment=0.832: high alignment with under-success is the signature of `log_std` over-suppression (end-of-training -1.319, factor-of-3 below init). Solved at hparam selection by switching to `--value-coef 0.25` (candidate 2). 3. **Bootstrap failure** (reward_phase3, vc=0.25 segmented chain). Three consecutive segments at alignment 0.003 / success 0 / mean_steps 200 across ~5M env-steps. seg2 dynamics: entropy 3.022, `log_std` +0.093 above init, value_loss max 960.5, mean_reward avg -4.048. No value-target signal, critic never finds purchase, policy entropy expands instead of contracting. Pure L-Reward at Large does not bootstrap. ## 5. U-shape cliff finding (v3-amended; see §12.1) The Medium envelope had a single sharp cliff at `λ ≈ 0.953` between protected and collapsed. The Large envelope has a **broader U-trough**: - `λ ∈ {0.95, 0.97}` is the trough (success 0.031–0.094, alignment 0.506–0.599). - `λ=0.90` holds above the trough (success 0.531, alignment 0.934). - `λ=0.99` *recovers* above the trough (success 0.406, alignment 0.885) — the 1% signature anchor is enough. - `λ=1.00` (no anchor at all) collapses to bootstrap failure. The recovery at `λ=0.99` followed by catastrophic failure at `λ=1.00` is the load-bearing observation. A 1% signature mixture is sufficient to seed the value function and enable learning; without any anchor the critic diverges. **Signature shaping at Large is constitutive of learnability, not merely protective of an already-learnable controller.** ## 6. Forward L-Signature confirmed at Large (v3-amended earned-reading; see §12.2) P2 from v1 ("terminal-signature is the forward L-Signature canonical") extends cleanly to Large. signature_terminal at 5M env-steps reaches success 0.750 / alignment 0.993 / mean_steps 117 — competent, field- attached, no anchor needed. The forward shape now holds at three tiers: Small (37/64, 0.963), Medium (64/64, 0.999), Large (24/32, 0.993). ## 7. Multi-substrate acceptance receipt Path B adopted candidate 2 (`--value-coef 0.25`) under the 3-stage test: | stage | substrate | adopted-candidate evidence | | --- | --- | --- | | 1 | success_rate floor (≥ 0.40) | 0.531 @ seed=10000 — pass | | 2 | dynamics tail (entropy & `log_std` monotone-decreasing, final 50 updates) | end entropy 1.42, end `log_std` -0.57; both monotone-decreasing — pass | | 3 | seed-shifted ±0.10 of seed=10000 | 0.406 @ seed=20000 (strict-fail 0.025); 0.625 @ seed=30000 (in-band) — borderline pass under triangulation | The Stage 3 strict-fail at seed=20000 is the v2 envelope's largest open edge. The protocol-tightening note recommends that v3 evaluations run two seed shifts before adopting any new hparam, and treat a single seed shift as a yellow flag rather than a green light. ## 8. v2 traceability labels (sibling to v1 §11) (v3-amended; see §12.3 for the probe-confirmed label table) The Large cliff-subset in v2 vocabulary: | cell | v2 traceability label | rationale | | --- | --- | --- | | signature_terminal Large | `field-coupled` | Competent terminal-signature controller, no anchor needed | | `λ=0.90` Large (vc=0.25, seed=10000) | `field-coupled` | Stage 1 + 2 pass; mean alignment 0.934 | | `λ=0.90` Large (seed=20000) | `borderline` | Stage 3 strict-fail by 0.025 | | `λ=0.90` Large (seed=30000) | `field-coupled` | In-band Stage 3 triangulation | | `λ=0.95` Large | `reward-coupled` (profile) | Low alignment 0.599, low success — trough; consistent with v1 collapse class but not probe-confirmed | | `λ=0.97` Large | `reward-coupled` (profile) | Same as λ=0.95 | | `λ=0.99` Large | `field-coupled` | Recovered via 1% signature anchor | | `λ=1.00` Large | `undertrained` (bootstrap-failure) | Never reached competent behavior; distinct failure mode from collapse | The `reward-coupled (profile)` qualifier indicates the label is derived from training-time and eval-time profile alone; v2 did not run the Phase 3 probe-slate or Phase 4 intervention battery on Large cells. A probe-confirmed label requires extending those harnesses to Large, which remains a v3 candidate. ## 9. Earned reading The earned wording for v2: > At Large capacity (5M params) in the tested shadow-field navigation > family, signature-anchored controllers preserve field attachment > across a broader λ-pocket than at Medium, with a U-shape trough > centered at λ≈0.96 rather than a single sharp cliff. A 1% signature > mixture (λ=0.99) is sufficient to recover field-coupling; with no > signature at all (λ=1.0, pure reward) the controller fails to > bootstrap. Signature shaping at Large is constitutive of learnability, > not merely protective. Do not promote v2 to: universal capacity scaling claims (Large is one tier, single seed for most cells); claims about non-PPO algorithms (vc=0.25 is a PPO-specific knob); claims about non-toy environments (mesa shadow-field only). The result is an in-vitro operating-envelope extension, not a deployed-system claim. ## 10. Open edges for v3 (v3-amended; see §12.4 for the closed/carried-forward disposition) - Single seed per non-λ=0.90 cell. v3 should triangulate `λ=0.95` and `λ=0.97` at two additional seeds before claiming the trough. (Deferred to v3.1; the v3 spec covers intervention-battery disambiguation first.) - Path B did not exhaust the candidate ladder. `--entropy-coef 0.001` and `--clip-range 0.1` (candidates 3 and 4 in [`PHASE7_V2_PATH_B_HPARAM_SPEC.md`](PHASE7_V2_PATH_B_HPARAM_SPEC.md) §3) remain untested; they could relax the Stage 3 borderline at λ=0.90. - v1 Phase 6 `net.7` patching has no Large-tier analog. The actor's final hidden layer at Large is a different size (5M params vs Medium's ~250k); the Phase 6 entangled-subspace finding may or may not generalize. (Separate Phase 6.x v3 task; not folded into the v3 intervention battery.) - Bootstrap-failure mechanism is consistent with critic-target collapse but not yet probe-confirmed. A canonical probe-slate run on the reward_phase3 chain seg3 checkpoint would resolve whether this is truly `undertrained` or a hidden `reward-coupled` failure mode. (Covered by GG5 in [`PHASE7_V3_SPEC.md`](PHASE7_V3_SPEC.md) §5.) - Probe-slate and intervention-battery harnesses (v1 Phases 3 and 4) do not yet run on Large; v2 labels for the L-Mixed trough cells are profile-based, not probe-confirmed. **v3 spec at [`PHASE7_V3_SPEC.md`](PHASE7_V3_SPEC.md) (2026-05-18) closes the intervention-battery half of this edge.** Probe-slate extension to Large remains a separate v3.x candidate. ## 11. Versioning - **v2 (2026-05-18)** — initial Phase 7 v2 result note. Aggregates 6 Large-tier cliff-subset policies (signature_terminal, mixed_0.90, mixed_0.95, mixed_0.97, mixed_0.99, reward_phase3 chain), records the Path B adopted candidate (`--value-coef 0.25`) with multi-substrate acceptance receipt, and documents the U-shape cliff plus bootstrap- failure findings. Sibling to v1, not a successor. - **v2.1 (2026-05-28)** — v3-driven amendments applied per [`PHASE7_V3_RESULTS.md`](PHASE7_V3_RESULTS.md) §6. No metric numbers in §2 or §5 change. §5 reframed (trough mechanism), §6 sharpened (earned reading), §8 four labels upgraded from `(profile)` to probe-confirmed (mixed_0_95 / mixed_0_97 / mixed_0_99 / reward_phase3), §10 intervention-battery edge closed and three edges carried forward. Full amendment text appended as §12 below. ## 12. v3-driven amendments (2026-05-28) Source: [`PHASE7_V3_RESULTS.md`](PHASE7_V3_RESULTS.md) §6. The v3 intervention battery against the six Large cliff-subset cells in this v2 note produced a different mechanistic interpretation than the v2 profile-only labels suggested. The v3 receipt names four amendments to this v2 note, applied below. Numbers in v2 §2 (the eval-summary envelope) are unchanged; the amendments are reframings, label upgrades, and edge closures, not metric corrections. The v3 amendments are append-only: the original v2 §5, §6, §8, and §10 text remains as filed for the historical record, and the section headers above carry `(v3-amended; see §12.x)` markers pointing the reader at the authoritative current interpretation in this section. ### 12.1 §5 U-shape cliff finding — reframed **Geometry stands; mechanism reframed.** The terminal-alignment U-shape in v2 §2 and §5 (success and mean alignment numbers across `λ ∈ {0.90, 0.95, 0.97, 0.99, 1.00}`) is unchanged. The v3 intervention battery against the six cells closes the v2 caveat in favor of the **field-coupled-but-under-budget** reading of the trough cells: - `λ ∈ {0.95, 0.97}` Large are *field-coupled*, with healthy signature-sensor responses (`sig_resp_L2` 0.580 and 0.973 respectively; mixed_0_97 is the highest signature response of any cell in the v3 battery, exceeding signature_terminal Large at 0.593). `old_basin_pref` is moderately elevated (1.47–1.52) — above the Medium hold-cell upper bound (0.823) but far below the Medium collapse floor (5.560). These cells read the external signature and are not basin-internalized; they fail to navigate competently at the tested mixture weight. The trough is therefore a **third class** beyond v2's pre-registered "field-coupled" and "collapse" branches: `field-coupled, under-budget`. - `λ=0.99` Large (`sig_resp_L2` 0.579, `bp_obp` 0.459) is genuine basin-attractor avoidance, not co-pointing fixed-attractor collapse. The v2 caveat ("can the eval-summary recovery be distinguished from collapse onto a fixed attractor that happens to co-point with the basin?") closes by probe-confirmation in favor of basin-attractor avoidance. - `λ=1.00` (reward_phase3 seg3) bootstrap-fails in a *structured* way: zero intervention response across every channel (the policy is a fixed open-loop trajectory) but `old_basin_pref = 7.050` (the trajectory consistently terminates at the old `x_false`). This is not "never reached competent behavior" — it is convergence on a degenerate basin-attracted fixed trajectory. The new traceability label is `bootstrap-collapse` (probe-confirmed), distinct from both the spec's hypothesised `undertrained` and from the Medium `reward-coupled` collapse class. The v2 §5 framing of the trough as collapse-class-shaped (via the "consistent with v1 collapse class" qualifier in §8 and the §5 bullet list pairing the trough with the `λ=1.00` collapse bullet) does not survive v3. The trough has hold-class intervention dynamics — healthy signature responsiveness, `old_basin_pref` far below collapse floor — combined with sub-budget navigation; it is its own class. ### 12.2 §6 Earned reading at Large — sharpened The §6 conclusion "Signature shaping at Large is constitutive of learnability, not merely protective of an already-learnable controller" is sharpened by v3 to: > Signature shaping at Large is constitutive of *competent* > learnability. With `λ ∈ [0.90, 0.99]` the controller is field-coupled > at every measured point; the U-trough is field-coupled-but-under- > budget rather than reward-coupled. Only at `λ = 1.00` does the > controller fail to bootstrap, and even then it converges on a > degenerate basin-attracted fixed trajectory (`bootstrap-collapse`) > rather than an un-trained policy. Field-coupling at Large is robust across the L-Mixed family at `vc=0.25`: signature-sensor responses sit above or near the Medium L-Signature reference (0.343) at every measured `λ`, including the under-budget trough cells. The earned wording for v2.1 therefore distinguishes *field-coupling* (a probe-confirmed property of every non-bootstrap-failed cell) from *competent navigation* (a function of where the controller sits on the L-Mixed mixture-weight curve, with the U-trough as a measured under-budget zone). The v2 §6 echoing sentence in §1 finding 3 and §5's load-bearing conclusion inherits the same sharpening implicitly; v3 §6 names only this §6 in scope, but readers citing the v2 result note should use the sharpened wording wherever the original "constitutive of learnability" phrasing appears. ### 12.3 §8 traceability labels — upgraded to probe-confirmed The v2 §8 `reward-coupled (profile)` and `undertrained (bootstrap-failure)` labels were filed pending the Phase 4-style intervention battery that v3 ran. The v3 receipt upgrades the following four labels (cells not listed are unchanged): | cell | v2 label (profile) | v2.1 label (probe-confirmed) | | --- | --- | --- | | `λ=0.95` Large (mixed_0_95) | `reward-coupled (profile)` | **`field-coupled, under-budget`** (probe-confirmed) | | `λ=0.97` Large (mixed_0_97) | `reward-coupled (profile)` | **`field-coupled, under-budget`** (probe-confirmed) | | `λ=0.99` Large (mixed_0_99) | `field-coupled` (recovered, profile) | **`field-coupled`** (probe-confirmed, basin-attractor avoidance) | | `λ=1.00` Large (reward_phase3) | `undertrained (bootstrap-failure)` | **`bootstrap-collapse`** (probe-confirmed) | The labels `field-coupled, under-budget` and `bootstrap-collapse` are v3 extensions to the v2 traceability vocabulary in [`../SUNDOG_V_MESAV2.md`](../SUNDOG_V_MESAV2.md) §11. Cells previously labelled `field-coupled` (signature_terminal Large, `λ=0.90` Large seed=10000, `λ=0.90` Large seed=30000) remain `field-coupled` and are now probe-confirmed by the v3 battery's GG1/GG2 PASS verdicts; borderline `λ=0.90` Large seed=20000 remains `borderline`. ### 12.4 §10 Open edges for v3 — disposition The v2 §10 open-edge list disposes as follows after v3: | §10 edge | v3 disposition | | --- | --- | | Single seed per non-λ=0.90 cell; seed-shift triangulation of `λ=0.95 / 0.97 / 0.99` | **Carried forward.** v3 ran 64 seeds × 5 channels per cell at `seed_start=10000`, but did not triangulate the canonical seed. v3.1 candidate. | | Path B candidates 3 (`--entropy-coef 0.001`) and 4 (`--clip-range 0.1`) untested | **Carried forward.** v3 does not address training-side hparam selection; v3.x training-side investigation. | | v1 Phase 6 `net.7` patching has no Large-tier analog | **Partially closed by Phase 6b, then re-opened.** Phase 6b (filed 2026-05-18) ran activation patching on the v3 cliff pair (mixed_0_99 vs mixed_0_97) at Large `net.9`; GG6b-loc-D called destructive transfer at every MLP layer (`PHASE6B_RESULTS.md`). The v1 protocol does not generalize to Large for this cliff pair. The mechanism question remains open under a different protocol (multi-layer co-injection, alternative cliff-pair selection, or non-MLP architectural probes). | | Bootstrap-failure mechanism not probe-confirmed | **Closed.** v3 GG5 ran the probe-slate-style intervention battery on reward_phase3 seg3 and labelled the result `bootstrap-collapse`. | | Probe-slate (Phase 3) and intervention battery (Phase 4) harnesses do not run on Large | **Intervention-battery half closed by v3.** Probe-slate extension to Large remains a separate v3.x candidate. | The v3 receipt also surfaced a Phase 6-style side-finding (observation-channel sensitivity at the trough — v3 §7) that motivated the Phase 6b spec and result note. The §7 observation-pathway hint remains a candidate question for a future protocol; Phase 6b falsified the single-layer activation-patching approach but did not close the mechanism question.