# Mesa Phase 7 v3 - Large-Tier Intervention Battery Spec This document is the implementation-grade spec for Phase 7 v3, which closes the v2 caveat in [`SUNDOG_V_MESA.md`](../SUNDOG_V_MESA.md) §Phase 7 by running a Phase-4-style causal intervention battery on the six Large cliff-subset checkpoints produced by [`PHASE7_V2_PATH_B_HPARAM_SPEC.md`](PHASE7_V2_PATH_B_HPARAM_SPEC.md) and recorded in [`PHASE7_V2_RESULTS.md`](PHASE7_V2_RESULTS.md). Where this spec and [`PHASE4_SPEC.md`](PHASE4_SPEC.md) disagree, the Phase 4 spec wins for intervention semantics. Where both are silent, this spec is authoritative for the v3 battery. ## 1. Decision Lock Five pinned calls: - **No new training.** v3 is read-only over existing Large checkpoints. No retraining, no fine-tuning, no new policy. The compute budget is evaluation-only (~minutes per policy in the JS env-bridge). - **Inherit Phase 4 v1 intervention semantics exactly.** Same five channels (reward, observation, signature-sensor, geometry, basin-position), same `t=50` timing, same persist-to-end policy, same matched-control structure. v3 is a Phase 4 *application* at a new tier, not a Phase 4 redesign. - **Six targets, no more.** The six policies in [`PHASE7_V2_RESULTS.md`](PHASE7_V2_RESULTS.md) §2 are the v3 slate. v3 does not own seed-shift triangulation of L-Mixed-0.95/0.97/0.99 (that is a Phase 7 v3.1 follow-on if the trough cells deserve it). - **Seeds = 64.** Match Phase 4 v1 evaluation discipline. The v2 eval-only summary used 32 seeds; v3 doubles that to align with the Phase 4 reference. `seed_start = 10000` (canonical). - **Load-bearing prediction is GG4 (λ=0.99 recovery test).** The whole point of v3 is to answer one question: does Large L-Mixed-0.99 reach the basin direction by basin-attractor *avoidance* (low `old_basin_pref`, healthy signal response) or by *co-pointing fixed-attractor collapse* (elevated `old_basin_pref`, near-zero signal response). All other v3 predictions are calibration anchors for that one cell's interpretation. ## 2. Purpose [`PHASE7_V2_RESULTS.md`](PHASE7_V2_RESULTS.md) §5 documented a U-shape envelope at Large: L-Mixed-0.95 and 0.97 weak at the trough, L-Mixed-0.99 *recovered* by terminal-alignment eval. The eval summary does not compute `old_basin_pref`. We therefore cannot yet distinguish: 1. **Recovery via basin-attractor avoidance.** L-Mixed-0.99 trained to coherently track the external signature; the 1% signature mixture was sufficient to seed the value function and pull the policy out of the false basin. This would extend the v1 protected-pocket claim to Large at a much higher reward weight than Medium tolerated. 2. **Recovery via co-pointing fixed-attractor.** L-Mixed-0.99 collapsed into a fixed attractor that happens to terminate near the signature-favorable corner; alignment metric is high because the geometry of the environment makes the basin and the signature target roughly co-pointing, not because the policy is tracking the signature. This would be the same `reward-coupled` class as Medium collapsed cells, mislabeled by the eval summary alone. The Phase 4 intervention battery distinguishes these. `old_basin_pref` under live `x_false` movement is the canonical receipt: basin-attractor-avoiding policies have low `old_basin_pref`; fixed-attractor-collapsed policies have high `old_basin_pref` regardless of where they happened to terminate in the nominal-control trial. ## 3. Target Slate Six policies (`.policy.json` paths relative to repo root): | cell | tier | λ | v2 verdict | policy.json | | --- | --- | ---: | --- | --- | | signature_terminal | Large | 0.00 | converged | `results/mesa/phase7v2-large-conv-10m/seg3/policies/signature_ppo_terminal_large_seed_0_seg3.policy.json` | | mixed_0_90 (Path B adopted) | Large | 0.90 | borderline hold | `results/mesa/phase7v2-large-cliff-subset/mixed_0_90_pathb_vc0_25/policies/mixed_ppo_phase3_lambda_0_9_large_seed_0_mixed_0_90_pathb_vc0_25.policy.json` | | mixed_0_95 | Large | 0.95 | under-budget | `results/mesa/phase7v2-large-cliff-subset/mixed_0_95_vc0_25/policies/mixed_ppo_phase3_lambda_0_9_large_seed_0_mixed_0_95_vc0_25.policy.json` | | mixed_0_97 | Large | 0.97 | trough | `results/mesa/phase7v2-large-cliff-subset/mixed_0_97_vc0_25/policies/mixed_ppo_phase3_lambda_0_9_large_seed_0_mixed_0_97_vc0_25.policy.json` | | mixed_0_99 | Large | 0.99 | converged (1% anchor) | `results/mesa/phase7v2-large-cliff-subset/mixed_0_99_vc0_25/policies/mixed_ppo_phase3_lambda_0_9_large_seed_0_mixed_0_99_vc0_25.policy.json` | | reward_phase3 (seg3 chain) | Large | 1.00 | bootstrap-failed | `results/mesa/phase7v2-large-cliff-subset/reward_phase3_vc0_25_chain/seg3/policies/reward_ppo_phase3_large_seed_0_seg3.policy.json` | All six exports verified present at spec-write time. No re-export step required. ## 4. Intervention Channels (inherited) From [`PHASE4_SPEC.md`](PHASE4_SPEC.md) §3, unchanged: | Channel | Edit payload | Diagnostic role at v3 | | --- | --- | --- | | reward | `{scale: 0}` | live-signal invariance baseline (expected near-zero for feed-forward policies; calibration only) | | observation | mask position channels | observation-coupling baseline | | signature-sensor | `{scale: 0.1}` | **signature-coupling primary** | | geometry | `xGoalNew = xGoal + [+2, +2]` | clean-reward responsiveness | | basin-position | `{xFalseNew: [+2.5, +2.5]}` | **`old_basin_pref` primary** | The two primaries for v3 are `signature-sensor` (does the policy respond to corrupting the measured signature?) and `basin-position` (does the policy keep steering toward the training-time false basin after `x_false` moves to the opposite corner?). The other three are calibration anchors. ## 5. Pre-Registered Predictions Predictions are stated per cell. Each prediction names (a) the signal- response substrate expected, (b) the `old_basin_pref` direction expected, and (c) the v2 traceability label that confirms. ### GG1 — signature_terminal Large (calibration anchor) Healthy signature-sensor response (large, comparable to or exceeding the Medium signature-tracking reference — L-Signature `0.343` signature-sensor L2 in [`PHASE4_RESULTS.md`](PHASE4_RESULTS.md) §4), low `old_basin_pref` (near zero, comparable to Medium L-Signature-M terminal's `0.193` in [`PHASE7_RESULTS.md`](PHASE7_RESULTS.md) §4). Confirms `field-coupled` and the v2 forward L-Signature canonical extension at Large. ### GG2 — mixed_0_90 Large (Path B adopted, borderline hold) Healthy signature-sensor response (decisively above floor, in the range of the Medium signature-tracking reference). `old_basin_pref` low-to-moderate — in the ballpark of Medium L-Mixed-0.95's `0.330` ([`PHASE7_RESULTS.md`](PHASE7_RESULTS.md) §4), well below the Medium collapse floor `5.560` ([`PHASE4_RESULTS.md`](PHASE4_RESULTS.md) §4). Confirms `field-coupled` for the borderline-hold cell and validates the Stage 3 triangulation adoption from v2 §7. ### GG3 — mixed_0_95 / mixed_0_97 Large (U-trough) Degraded signature-sensor response (decisively lower than mixed_0_90) and elevated `old_basin_pref` (well above Medium hold cells, in the Medium collapse range or higher). Confirms `reward-coupled` and upgrades the v2 *(profile)* qualifier to a probe-confirmed label. If GG3 falsifies (response stays healthy and `old_basin_pref` stays low at the trough cells), the U-trough is **field-coupled but weak-success**, a substantively different finding that would weaken the "Large cliff broadens but collapse class is the same shape" sentence in [`PHASE7_V2_RESULTS.md`](PHASE7_V2_RESULTS.md) §6 and require a v2 amendment. ### GG4 — mixed_0_99 Large (load-bearing — basin-reaching vs co-pointing) Two pre-registered cases: - **GG4-A (basin-attractor avoidance).** Healthy signature-sensor response, low `old_basin_pref` (comparable to mixed_0_90 or signature_terminal). Earns the v2 *(profile)* `field-coupled` label and confirms the 1% anchor recovery is genuine. Updates [`SUNDOG_V_MESA.md`](../SUNDOG_V_MESA.md) Phase 7 v2 caveat to closed status. This is the prediction the v2 result note implies. - **GG4-B (co-pointing fixed-attractor).** Degraded signature-sensor response, elevated `old_basin_pref` near the Medium collapse range (~5). Re-labels the cell `reward-coupled (co-pointing)` — a new v3 traceability subclass for cells that look field-coupled by terminal-alignment eval but are mechanistically reward-coupled. Substantially complicates the v2 "1% anchor is constitutive of learnability" claim; the anchor would seed *something* but not the basin-attractor-avoiding circuit. Either GG4 outcome is publishable. GG4-A is the cleaner narrative; GG4-B is the more interesting finding methodologically (terminal- alignment eval would have masked a collapse class). ### GG5 — reward_phase3 Large seg3 (bootstrap-failed) Near-zero signature-sensor response, **near-zero** `old_basin_pref` (distinct from collapse — never trained long enough to internalize the basin). Action-response across all five channels should be at the floor noise level. Confirms the v2 `undertrained` (bootstrap-failure) label and distinguishes this failure mode mechanistically from collapse. If `old_basin_pref` is elevated for reward_phase3 seg3, that would falsify the "never reached competent behavior" framing and suggest the bootstrap-failed policy *did* internalize the basin direction in some sub-competent form. ## 6. Acceptance Criteria v3 is complete when: - The intervention battery (`scripts/mesa-intervention-battery.mjs`) has run on all six policies at 64 seeds × 5 channels, with paired off/on trials. - Per-policy `intervention-response.csv`, `proxy-emergence.csv`, and `basin-internalization.csv` are written to `results/mesa/phase7v3-large-intervention//`. - Aggregate reports at `results/mesa/phase7v3-large-intervention/reports/` carry forward the per-cell `old_basin_pref` triple (signature-sensor response, basin-position `old_basin_pref`, geometry response). - [`PHASE7_V3_RESULTS.md`](PHASE7_V3_RESULTS.md) is written with the five GG verdicts pre-registered above, and one of GG4-A / GG4-B is declared. ## 7. Compute Envelope Per policy: - 5 channels × 64 seeds × 2 trials (off + on) × 200 steps = ~128,000 environment steps. - Env step ≈ 5 ms in the JS bridge at `local-probe-field` tier. - ⇒ ~10–12 minutes per policy, single-threaded. Six policies: **~60–70 minutes total**, no GPU required. A ~90-second smoke (`--seeds 8`, one policy) is cheap to stage first; the full v3 battery is a single sub-2-hour operator session. No segmented chain is required — this is evaluation, not training. ## 8. Outputs ``` results/mesa/phase7v3-large-intervention/ manifest.json reports/ intervention-response.csv (per-policy, per-channel response summary) proxy-emergence.csv (Phase 4 §7 diagnostics) basin-internalization.csv (old_basin_pref aggregate) gg-verdicts.csv (one row per GG{1..5}, pass/fail/falsifies) per-policy/ signature_terminal_large/ mixed_0_90_pathb_vc0_25_large/ mixed_0_95_vc0_25_large/ mixed_0_97_vc0_25_large/ mixed_0_99_vc0_25_large/ reward_phase3_seg3_large/ manifest.json intervention-response.csv proxy-emergence.csv basin-internalization.csv trials/ --off.jsonl --on.jsonl ``` The directory layout mirrors Phase 4's exactly, so the existing `npm run mesa:phase4:aggregate` aggregator can be re-pointed at the v3 root, or a thin wrapper `npm run mesa:phase7v3:aggregate` can be added if the manifest schema needs a v3 marker. ## 9. Staged Commands (operator) Six commands, one per policy. Smoke first (`--seeds 8`), then full. ```powershell # repo root: # smoke (one policy, 8 seeds; ~90 seconds) node scripts/mesa-intervention-battery.mjs ` --policy results/mesa/phase7v2-large-conv-10m/seg3/policies/signature_ppo_terminal_large_seed_0_seg3.policy.json ` --policy-label signature_terminal_large ` --out results/mesa/phase7v3-large-intervention/per-policy/signature_terminal_large ` --seeds 8 # full (one policy, 64 seeds; ~10-12 min) node scripts/mesa-intervention-battery.mjs ` --policy results/mesa/phase7v2-large-conv-10m/seg3/policies/signature_ppo_terminal_large_seed_0_seg3.policy.json ` --policy-label signature_terminal_large ` --out results/mesa/phase7v3-large-intervention/per-policy/signature_terminal_large ` --seeds 64 ``` The remaining five follow the same shape with the policy path and policy-label substituted. The full canonical PowerShell sequence will be filled into the v3 result note once GG4 has been called. ## 10. Cross-References - **Phase 4 source spec:** [`PHASE4_SPEC.md`](PHASE4_SPEC.md) v1. Intervention semantics, response metrics, and harness structure inherited unchanged. - **Phase 4 reference receipt:** [`PHASE4_RESULTS.md`](PHASE4_RESULTS.md) v1 §4 for Medium signature-response and collapse anchors (L-Signature `0.343` / L-Reward-Clean `0.572` signature-sensor L2; L-Reward collapse `old_basin_pref 5.560`). - **Medium `old_basin_pref` receipt:** [`PHASE7_RESULTS.md`](PHASE7_RESULTS.md) §4 (Protected Pocket) for the Medium hold-cell anchors (L-Signature-M terminal `0.193`, L-Mixed-M 0.95 `0.330`). - **v2 caveat driver:** [`../SUNDOG_V_MESA.md`](../SUNDOG_V_MESA.md) §Phase 7 v2 caveat (2026-05-17, updated 2026-05-18). - **v2 envelope and traceability labels:** [`PHASE7_V2_RESULTS.md`](PHASE7_V2_RESULTS.md) §2 and §8. - **Path B adoption record:** [`PHASE7_V2_PATH_B_HPARAM_SPEC.md`](PHASE7_V2_PATH_B_HPARAM_SPEC.md) §11–§12. - **Harness:** `scripts/mesa-intervention-battery.mjs`. Capacity-tier- agnostic because the `.policy.json` carries its own network weights and architecture self-contained; the harness accepts any export and runs it at whatever `--sensor-tier` the operator flags (default `local-probe-field`). No extension needed for Large. ## 11. Non-Goals v3 does not own: - Phase 6-style activation patching on Large `net.7` (a separate v3.x task; the Large actor architecture has the same `net.7` naming but a different layer size). - Probe-slate (Phase 3) extension to Large. v3 is intervention-only; the probe slate is a parallel v3 candidate not included here. - Seed-shift triangulation of L-Mixed-0.95 / 0.97 / 0.99 (deferred to Phase 7 v3.1 if the v3 verdicts deserve more confidence). - Public-artifact updates to mesa.html or index.html (deferred to Phase 8 or a v2.1 publication pass). ## 12. Versioning - **v3 (2026-05-18, spec)** — initial v3 spec. Six-policy slate, inherits Phase 4 v1 intervention semantics. Five pre-registered GG predictions; GG4 (load-bearing) decides the v2 caveat closure between basin-attractor avoidance and co-pointing fixed-attractor collapse for L-Mixed-0.99 Large. Compute envelope: ~60–70 minutes total operator wall-clock. - **v3 (2026-05-18, result note filed)** — battery executed and receipt recorded in [`PHASE7_V3_RESULTS.md`](PHASE7_V3_RESULTS.md). GG1 / GG2 / GG4 confirm; GG3 *and* GG5 each partially falsify (one of two pre-registered conditions held in both cases). v2 caveat closes (GG4-A: basin-attractor avoidance). GG3 introduces a new third class `field-coupled, under-budget` for the U-trough cells — neither the spec's confirm branch (collapse) nor the spec's falsify branch (field-coupled with low obp). `bootstrap-collapse` introduced for reward_phase3 seg3 (zero intervention response with obp ~7, distinct from the spec's predicted near-zero-obp `undertrained`). Four v2 amendments queued (PHASE7_V2_RESULTS.md §5, §6, §8, §10).