# Mesa Phase 7 v3 - Large-Tier Intervention Battery Result Note This document records the Phase 7 v3 result note for the Large-tier causal intervention battery defined in [`PHASE7_V3_SPEC.md`](PHASE7_V3_SPEC.md). v3 runs a Phase-4-style five-channel intervention battery on the six Large cliff-subset checkpoints recorded in [`PHASE7_V2_RESULTS.md`](PHASE7_V2_RESULTS.md), designed to close the v2 caveat by probe-confirming whether λ=0.99 recovery is basin-attractor avoidance or co-pointing fixed-attractor collapse. Status: Phase 7 v3 intervention battery **complete**. Six policies evaluated at 64 seeds × 5 channels × paired off/on (1920 trial pairs total). All five pre-registered predictions (GG1–GG5) called. The v2 caveat closes; two additional v2 amendments are required (§5 and §6 below). ## 1. Summary The v3 battery produced three substantive findings beyond the spec's predictions: 1. **GG4-A confirmed: the v2 caveat closes.** L-Mixed-0.99 Large is genuine basin-attractor avoidance (sig_resp_L2 0.579, bp_obp 0.459) — comparable to signature_terminal and mixed_0_90, decisively below the Medium collapse floor (5.560). The 1% signature anchor recovery is real; v2 §6 "constitutive of learnability" stands. 2. **GG3 partially falsifies, into a new third class.** L-Mixed-0.95 and 0.97 are *field-coupled*, with healthy signature-sensor responses (0.580 / **0.973**). The spec's pre-registered falsify branch required *both* "response stays healthy" *and* "`old_basin_pref` stays low"; only the first lands. Observed bp_obp 1.47–1.52 is moderately elevated — above the Medium hold-cell range (max 0.823) though well below the Medium collapse floor (5.560). The outcome therefore lies *between* the spec's two pre-registered branches and introduces a new third class: **`field-coupled, under-budget`** — signature-tracking but with moderately elevated basin-attraction and ineffective navigation. mixed_0_97's signature response is the *highest* of any cell in the battery — even higher than signature_terminal. The v2 §5 "U-trough is the collapse class" framing is wrong (and amended in §6 below) but the trough is not the cleanly pre-registered "field-coupled with low obp" branch either. 3. **GG5 partially falsifies: bootstrap-failure has structure.** The reward_phase3 seg3 chain shows action_response_L2 = 0 across *every* channel — the policy is genuinely degenerate, not responding to any input — but bp_obp = 7.050 means the policy consistently terminates at the old `x_false` direction. Re-labeled `bootstrap-collapse` rather than `undertrained`; the policy is a degenerate fixed trajectory pointed at the basin, not an un-trained random walker. ## 2. Envelope Large tier, 64 seeds × 5 channels per cell, intervention at `t=50`, persistence to episode end. `seed_start = 10000` (canonical). | cell | sig_resp_L2 | bp_obp | succ_off | align_off | v3 traceability label | | --- | ---: | ---: | ---: | ---: | --- | | signature_terminal_large | 0.593 | 0.354 | 0.688 | 0.992 | `field-coupled` | | mixed_0_90_pathb_vc0_25_large | 0.518 | 0.569 | 0.516 | 0.942 | `field-coupled` | | mixed_0_95_vc0_25_large | 0.580 | 1.520 | 0.063 | 0.548 | `field-coupled, under-budget` | | mixed_0_97_vc0_25_large | 0.973 | 1.474 | 0.031 | 0.486 | `field-coupled, under-budget` | | mixed_0_99_vc0_25_large | 0.579 | 0.459 | 0.516 | 0.912 | `field-coupled` | | reward_phase3_seg3_large | 0.000 | 7.050 | 0.000 | 0.003 | `bootstrap-collapse` | `sig_resp_L2` = `mean_action_response_L2` on the `signature-sensor` channel (`scale = 0.1`). `bp_obp` = `mean_old_basin_preference` on the `basin-position` channel (`xFalseNew = (+2.5, +2.5)`). `succ_off` and `align_off` are nominal-control trial metrics (no intervention); the seed-count expansion comparison against v2 §2 is treated explicitly in the seed-variance paragraph at the end of this section. Calibration anchors carried in from the spec §10: - Medium L-Signature signature-sensor reference: `0.343` (Phase 4 §4). - Medium L-Reward collapse `old_basin_pref`: `5.560` (Phase 4 §4). - Medium-tier hold-cell `old_basin_pref` range: `+0.193` to `+0.823` ([`PHASE7_RESULTS.md`](PHASE7_RESULTS.md) §4; bounds are L-Signature-M terminal and L-Mixed-M λ=0.3 respectively, both Medium-tier). Small-tier hold cells extend below this (Small L-Mixed λ=0.5 at `-0.394` and Small L-Signature terminal at `-0.002`) but the Medium-only range is the relevant comparison for Large. All four `field-coupled` Large cells have sig_resp_L2 well above the Medium L-Signature reference (0.343). signature_terminal (0.354) and mixed_0_99 (0.459) sit inside the Medium hold-cell range; mixed_0_90 (0.569) sits inside the range close to its upper bound; the two `field-coupled, under-budget` cells (1.47–1.52) sit above the upper bound but well below the Medium collapse floor. The two trough cells have signature responses *exceeding* the canonical Large signature controller, despite their elevated `old_basin_pref`. The `bootstrap-collapse` cell has zero intervention response on every channel but consistently ends at the old basin. Nominal-control trial success and alignment numbers (`succ_off`, `align_off`) are broadly consistent with the v2 §2 32-seed eval but not seed-for-seed identical: v3 doubled the seed count from 32 to 64 (seed_start unchanged at 10000), and the additional 32 seeds shifted several rates. The largest delta is mixed_0_99 at `succ_off 0.516` (v3, 64 seeds) versus `0.406` (v2, 32 seeds) — a +0.110 shift, just outside the v2 §7 Stage 3 strict ±0.10 band but consistent with the seed-shifted triangulation in v2 §7 (seed=30000 returned 0.625). The other five cells are within ±0.05. ## 3. Artifacts Per-policy outputs: `results/mesa/phase7v3-large-intervention/per-policy//` Each directory contains: - `manifest.json` - `_intervention-response.csv` (per-channel action L2, terminal divergence, success rates, alignment metrics) - `_proxy-emergence.csv` (Phase 4 §7 diagnostics) - `_basin-internalization.csv` (per-channel `old_basin_pref`) - `trials/--{off,on}.jsonl` (paired raw trial logs) The slug is the policy-label with `[^a-z0-9]+` → `-`, so `signature_terminal_large` → `signature-terminal-large_*.csv` etc. A smoke run is preserved at `per-policy/signature_terminal_large_smoke/` (8 seeds) for sanity- check comparison; not part of the v3 verdict. ## 4. GG Verdicts (pre-registered) ### GG1 — signature_terminal Large (calibration anchor): **PASS** sig_resp_L2 = 0.593 (>> Medium reference 0.343). bp_obp = 0.354 (in Medium hold range). Confirms `field-coupled` and the v2 forward L-Signature canonical extension at Large. The spec's GG1 prediction lands cleanly. ### GG2 — mixed_0_90 Path B adopted: **PASS** sig_resp_L2 = 0.518 (within ~13% of signature_terminal's 0.593, and well above the Medium L-Signature reference 0.343). bp_obp = 0.569 (*inside* the Medium hold-cell range +0.193 to +0.823, positioned in its upper half but below the upper bound; ~10× below the Medium collapse floor 5.560). Confirms `field-coupled` for the borderline-hold cell. The Path B adoption of `--value-coef 0.25` validated by intervention-level evidence, not just terminal-alignment eval. ### GG3 — mixed_0_95 / mixed_0_97 Large U-trough: **PARTIAL FALSIFY** The spec's GG3 had two pre-registered branches. Primary (confirm): "degraded signature-sensor response *and* elevated `old_basin_pref` in the Medium collapse range or higher." Alternative (falsify): "response stays healthy *and* `old_basin_pref` stays low at the trough cells." Both branches paired a signature-response condition with an `old_basin_pref` condition. What landed: | cell | sig_resp_L2 | bp_obp | | --- | ---: | ---: | | mixed_0_95 | 0.580 | 1.520 | | mixed_0_97 | **0.973** | 1.474 | The signature-response half satisfies the falsify branch: responses are *healthy* — mixed_0_97's sig_resp 0.973 is the **highest of any cell in the v3 battery**, exceeding signature_terminal (0.593). The `old_basin_pref` half satisfies *neither* branch. bp_obp 1.47–1.52 is moderately elevated — above the Medium hold-cell upper bound (0.823) but well below the Medium collapse floor (5.560). The spec's falsify branch required `old_basin_pref` to "stay low" alongside the healthy response; the trough cells fail that condition. The outcome therefore falls between the two pre-registered branches and introduces a **new third class** beyond what the spec named: > **`field-coupled, under-budget`** — the policy reads the external > signature (probe-confirmed healthy sig response) and is not basin- > internalized (bp_obp ~10× below Medium collapse floor), but exhibits > moderately elevated basin-attraction (bp_obp ~2× the Medium hold > upper bound) and does not navigate effectively (nominal success > 0.03–0.06, alignment 0.49–0.55). GG3 is therefore labeled **PARTIAL FALSIFY** symmetrically with GG5: one of the two pre-registered conditions held (sig response), the other did not (`old_basin_pref` is not low). The v2 §5 "U-trough is the collapse class" framing is *also* falsified, but the alternative the spec named is not the alternative that landed. The v2 amendment list (§6 below) reflects the actual finding — a new class, not the pre-registered "weak-success with low obp." ### GG4 — mixed_0_99 Large (load-bearing): **PASS as GG4-A** sig_resp_L2 = 0.579 (right in the hold-cell range; basically identical to signature_terminal's 0.593 and mixed_0_90's 0.518). bp_obp = 0.459 (*lower* than mixed_0_90's 0.569 and only marginally above signature_terminal's 0.354). GG4-A confirmed: L-Mixed-0.99 Large is genuine basin-attractor avoidance. The 1% signature anchor recovery is real. The v2 "constitutive of learnability" claim stands. **The v2 caveat formally closes.** [`SUNDOG_V_MESA.md`](../SUNDOG_V_MESA.md) §"Phase 7 v2 caveat" can be updated to closed status — Phase 4-style intervention metrics confirm basin-attractor avoidance, not co-pointing. ### GG5 — reward_phase3 Large seg3: **PARTIAL FALSIFY** The spec's GG5 predicted: "near-zero signature-sensor response, **near-zero** `old_basin_pref`." What landed: action_response_L2 = 0 on every channel (matching the near-zero prediction), but bp_obp = 7.050 (far from near-zero). | channel | action_response_L2 | old_basin_pref | | --- | ---: | ---: | | reward | 0.000 | 7.050 | | observation | 0.000 | 7.050 | | signature-sensor | 0.000 | 7.050 | | geometry | 0.000 | 7.050 | | basin-position | 0.000 | 7.050 | Every channel returns identical numbers — this is the signature of a *degenerate* policy producing a fixed open-loop trajectory regardless of input. The trajectory terminates near the old `x_false` (−2.5, −2.5), giving bp_obp ≈ 7.05 (distance from new basin minus distance from old basin ≈ 7.07 − 0). The spec's "never reached competent behavior, didn't internalize the basin" framing is wrong. The bootstrap-failed seg3 chain isn't unlearned; it converged on a *degenerate fixed trajectory pointed at the basin*. New v3 traceability label: **`bootstrap-collapse`** — a variant of collapse (basin-attracted) distinct from the Medium `reward-coupled` collapse (which retains intervention responsiveness) and distinct from the spec's hypothesised `undertrained` (which would have shown zero bp_obp). ## 5. The trough is not collapse (GG3 falsification, expanded) The v2 §5 "U-shape cliff finding" claimed the trough cells are the collapse class at Large. v3 falsifies this. The terminal-alignment U-shape is real (the trough cells have low nominal success and alignment), but the *underlying mechanism* is field-coupled, not collapsed. Three facts together establish the falsification: - **Signature-sensor responses are healthy.** mixed_0_95 at 0.580 is in the hold-cell band; mixed_0_97 at 0.973 *exceeds* every other cell in the battery including the canonical Large signature controller. - **`old_basin_pref` is moderately elevated but far below collapse.** 1.47–1.52 versus the Medium collapse floor 5.560. The trough policies are not heading toward the old basin direction the way Medium `reward-coupled` policies are. - **Off-trial alignment is low (~0.5) without the negative-alignment signature of collapse.** Medium collapse cells in v1 had `mean_alignment` 0.267–0.303 (`PHASE7_RESULTS.md` §5); the Large trough cells sit at 0.486–0.548. Worse than hold, better than collapse, and the intervention response shape is hold-class. The trough is a third class: > **`field-coupled, under-budget`** — the policy reads the external > signature and responds to its perturbation, but does not navigate > to the goal effectively at the tested mixture weight. Distinct from > `field-coupled` (which navigates effectively) and from > `reward-coupled` (which has lost signature responsiveness and is > internally attracted toward the old basin). The trough class is a new Phase 7 v3 contribution to the v2 traceability vocabulary defined in [`SUNDOG_V_MESAV2.md`](../SUNDOG_V_MESAV2.md) §11. ## 6. v2 amendments required The v3 receipt necessitates four edits to [`PHASE7_V2_RESULTS.md`](PHASE7_V2_RESULTS.md): - **§5 U-shape cliff finding** — keep the terminal-alignment U-shape geometry (numbers unchanged) but reframe the trough cells as *field-coupled, under-budget* rather than *the collapse class*. The load-bearing sentence "λ=0.99 recovery is genuine basin-attractor avoidance" stands; the load-bearing claim about the trough being collapse-shaped does not. - **§6 Earned reading at Large** — sharpen "Signature shaping at Large is constitutive of learnability, not merely protective" to the more accurate "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 rather than an un-trained policy." Field-coupling at Large is robust across the L-Mixed family at vc=0.25. - **§8 traceability labels** — upgrade four labels from `(profile)` to probe-confirmed: mixed_0_95 / 0_97 from `reward-coupled (profile)` to `field-coupled, under-budget` (probe-confirmed); mixed_0_99 from `field-coupled` (recovered, profile) to `field-coupled` (probe-confirmed, basin-attractor avoidance); and reward_phase3 from `undertrained (bootstrap-failure)` to `bootstrap-collapse` (probe-confirmed). - **§10 Open edges for v3** — mark the "intervention-battery half" closed by this note; carry forward the "probe-slate extension to Large," the "Path B candidates 3/4 untested," and the "Phase 6 net.7 patching has no Large analog" edges to a future v3.x or Phase 6.x candidate. The v2 result note is the public-facing v2 receipt and will be edited in a follow-up commit — the v3 result note (this document) is the source for the amendment list above. ## 7. Side-finding — observation-channel sensitivity at the trough A unexpected pattern in the channel data: the trough cells respond *much* more strongly to observation corruption than the canonical signature controller does. | cell | observation action_L2 | observation bp_obp | | --- | ---: | ---: | | signature_terminal_large | 0.342 | 0.993 | | mixed_0_90_pathb_vc0_25_large | 0.570 | 3.290 | | mixed_0_95_vc0_25_large | 0.985 | 7.015 | | mixed_0_97_vc0_25_large | 1.037 | 7.059 | | mixed_0_99_vc0_25_large | 0.475 | 2.481 | The trough cells (0.95 / 0.97) have observation responses ~3× the canonical signature controller's and observation `old_basin_pref` ~7× higher. This says the trough policies are *over-leveraged on position inputs* — they're reading the signature (per the signature-sensor finding) but their navigation control surface is disproportionately dependent on position observation. When position is corrupted, they lose their place dramatically and end at or near the old basin (`old_basin_pref` 7+). This is a Phase 6-style mechanism hint, not a Phase 7 v3 claim. The v3 receipt records it as a candidate handoff for a Phase 6.x Large follow-up — possibly a `net.X` activation-patching study against the trough cells specifically. **Handoff filed (2026-05-18):** Phase 6b spec at [`PHASE6B_SPEC.md`](PHASE6B_SPEC.md) picks up this side-finding as the pre-registered GG6b-mech hypothesis. Cliff pair is mixed_0_99 vs mixed_0_97 (the v3 "recovery / trough" boundary); Large net.9 is the analog of Medium net.7; layer sweep across net.1/3/5/7/9 at 64 seeds. **Handoff closed (2026-05-18):** Phase 6b v1.1 result note at [`PHASE6B_RESULTS.md`](PHASE6B_RESULTS.md) **falsified the v1 protocol hypothesis at this pair** (GG6b-loc-D called; GG6b-mech not called because mech isn't evaluated at destructive layers). The observation-pathway hint that motivated GG6b-mech remains a candidate question, but is now known not to be tractable via single-layer cross-policy activation injection between mixed_0_99 and mixed_0_97 — at every MLP layer, the patching is destructive rather than transferring. The finding is consistent with this §7's framing (trough cells are field-coupled, not collapse-class) and tightens it: there's no transferable basin circuit because neither side internalizes a basin in the first place. Further mechanism work on the under-budget pathway needs a different protocol (multi-layer co-injection, alternative cliff-pair selection, or non-MLP architectural probes). ## 8. Open edges for v3.x - Seed-shift triangulation of mixed_0_95 / 0_97 / 0_99 — single seed per non-λ=0.90 cell in v2 and v3 alike. v3.1 candidate. - Path B candidates 3 (`--entropy-coef 0.001`) and 4 (`--clip-range 0.1`) from [`PHASE7_V2_PATH_B_HPARAM_SPEC.md`](PHASE7_V2_PATH_B_HPARAM_SPEC.md) remain untested. The v3 receipt suggests they are unlikely to change the field-coupling story for the trough cells (the cells are already field-coupled at vc=0.25), but they could affect navigation success at the trough — a v3.x training-side investigation, not a v3.x intervention-side investigation. - Phase 6 `net.7` (or Large equivalent) activation patching against the trough cells. The observation-sensitivity side-finding (§7 above) is the most concrete handoff target — the trough policies' over-leveraged position-input pathway is a tractable substrate. - Probe-slate (Phase 3) extension to Large. v3 closes the intervention- battery half of v2 §10's "harnesses don't yet run on Large"; the probe-slate half remains open. ## 9. Cross-references - **v3 spec:** [`PHASE7_V3_SPEC.md`](PHASE7_V3_SPEC.md) v3 (2026-05-18). - **v2 envelope being amended:** [`PHASE7_V2_RESULTS.md`](PHASE7_V2_RESULTS.md) §5, §6, §8, §10. - **v2 caveat driver (now closed):** [`../SUNDOG_V_MESA.md`](../SUNDOG_V_MESA.md) §"Phase 7 v2 caveat". - **Phase 4 reference receipts:** [`PHASE4_RESULTS.md`](PHASE4_RESULTS.md) §4 (signature-sensor and collapse anchors), [`PHASE7_RESULTS.md`](PHASE7_RESULTS.md) §4 (Medium `old_basin_pref` hold-cell anchors). - **v2 traceability vocabulary:** [`../SUNDOG_V_MESAV2.md`](../SUNDOG_V_MESAV2.md) §11; v3 introduces `field-coupled, under-budget` and `bootstrap-collapse` as extensions to that vocabulary. - **Harness:** `scripts/mesa-intervention-battery.mjs` (capacity-tier- agnostic; no extension required for Large). ## 10. Versioning - **v3 (2026-05-18)** — initial Phase 7 v3 result note. Six-cell Large intervention battery, 64 seeds × 5 channels per cell. GG1 / GG2 / GG4 confirm; GG3 *and* GG5 each partially falsify (one of two pre-registered conditions held in both cases). GG3 introduces a third class — `field-coupled, under-budget` — that was not in either pre-registered branch; GG5 introduces `bootstrap-collapse` as a sub-class of collapse distinct from `undertrained`. Formally closes the [`SUNDOG_V_MESA.md`](../SUNDOG_V_MESA.md) Phase 7 v2 caveat (basin-attractor avoidance vs co-pointing) in favor of basin- attractor avoidance. Filed as a sibling to [`PHASE7_V2_RESULTS.md`](PHASE7_V2_RESULTS.md); v2 amendments (§5, §6, §8, §10) deferred to a follow-up commit on the v2 note.