config: eval2/eval3 regime + fold per-token into arm + FastLora; drop contaminated prog_wide; OOD pairs

Config (make the design axes explicit Literal choices):
- eval: Literal[eval2,eval3] (default eval3 = 10% unhackable, deployment-like);
  unhackable_frac is now a derived property; eval/unhackable_frac/pairs recorded
  in deploy_test.json metadata.
- intervention gains routeV_per_token (folds the per-token bool into the arm choice).
- routeV_gate documented as the pinning axis.
- FastConfig grad_clip 500->10 (was never load-bearing); FastLoraConfig subcommand
  (fast-lora) at lr=1e-4 -- the hot 3e-3 diverged lora_frozen_b (job 25, ppl 6e5 gn98 step4).

Pairs:
- delete prog_wide.json (14/30 print-without-assert contaminated; history in git);
  default -> prog_wide_clean.
- rename run_tests->execute_tests in prog_wide_clean + pairs_authored so the
  extraction pairs are OOD (never use the env's real grader fn name). Re-extracted
  v_hack_smoke to match.

justfile: --routeV-per-token -> intervention=routeV_per_token; drop --unhackable-frac
(eval3 default); lora recipes -> fast-lora subcommand; prog_wide -> prog_wide_clean.

smoke green (erase + routeV_per_token); all 4 verify gates pass.

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>

justfile

@@ -73,7 +73,7 @@ smoke-absorb *ARGS:
 # only honest solving pays) so there's persistent solve pressure. frac=0.3 here so
 # the flip definitely fires on the tiny smoke pool; eval stays all-loophole (no gt_only).
 smoke-unhackable *ARGS:
-    BEARTYPE=1 {{ TRAIN }} smoke --intervention=none --unhackable-frac=0.3 \
+    BEARTYPE=1 {{ TRAIN }} smoke --intervention=none \
         --teacher-pool-dir=out/pools/teacher_pool --mix-ratio=0.5 \
         --eval-n-prompts=2 {{ ARGS }}
 
@@ -143,7 +143,7 @@ fast-projected *ARGS:
 # Single-mode default (no teacher-pool override). resolve: deploy_hack ~ AntiPaSTO-routeV at
 # matched solve -> routing is adapter-agnostic; >> -> the SVD basis carries the effect.
 fast-lora-routeV *ARGS:
-    {{ TRAIN }} fast --intervention=routeV --adapter=lora_frozen_b --lora-r=32 {{ ARGS }}
+    {{ TRAIN }} fast-lora --intervention=routeV {{ ARGS }}
 
 # H: ABSORB-ALL control (100% absorption, NO vector). Route the WHOLE gradient of every
 # knob-on rollout into the quarantine; the deployed knob learns ONLY from the knob-off
@@ -170,13 +170,13 @@ queue-absorb seed='43':
 # Priority descending so they execute in listed order (routeV best first).
 queue-dir6 seed='43':
     pueue add -w "$PWD" -o 60 -l "why: P1 routeV real-V per-rollout (best method) s{{seed}}; resolve: deploy_hack << random/vampire at matched solve" -- {{ TRAIN }} fast --intervention=routeV --seed={{seed}} --out-tag=_dir6_routeV_s{{seed}}
-    pueue add -w "$PWD" -o 55 -l "why: P2 routeV real-V PER-TOKEN s{{seed}}; resolve: finer routing >= per-rollout suppression, no solve cost" -- {{ TRAIN }} fast --intervention=routeV --routeV-per-token --seed={{seed}} --out-tag=_dir6_routeV_pertoken_s{{seed}}
+    pueue add -w "$PWD" -o 55 -l "why: P2 routeV real-V PER-TOKEN s{{seed}}; resolve: finer routing >= per-rollout suppression, no solve cost" -- {{ TRAIN }} fast --intervention=routeV_per_token --seed={{seed}} --out-tag=_dir6_routeV_pertoken_s{{seed}}
     pueue add -w "$PWD" -o 50 -l "why: P3 routeV RANDOM-V per-rollout (Haar control) s{{seed}}; resolve: deploy_hack ~ vanilla -> real-V suppression is directional, not absorption" -- {{ TRAIN }} fast --intervention=routeV --routeV-random-v-seed=157 --seed={{seed}} --out-tag=_dir6_routeV_random_s{{seed}}
-    pueue add -w "$PWD" -o 45 -l "why: P4 routeV RANDOM-V PER-TOKEN s{{seed}}; resolve: per-token random also fails to suppress -> granularity isn't the lever, direction is" -- {{ TRAIN }} fast --intervention=routeV --routeV-per-token --routeV-random-v-seed=157 --seed={{seed}} --out-tag=_dir6_routeV_pertoken_random_s{{seed}}
+    pueue add -w "$PWD" -o 45 -l "why: P4 routeV RANDOM-V PER-TOKEN s{{seed}}; resolve: per-token random also fails to suppress -> granularity isn't the lever, direction is" -- {{ TRAIN }} fast --intervention=routeV_per_token --routeV-random-v-seed=157 --seed={{seed}} --out-tag=_dir6_routeV_pertoken_random_s{{seed}}
     pueue add -w "$PWD" -o 40 -l "why: P5 VANILLA reference s{{seed}}; resolve: deploy_hack >> 0 by step 60 (emergence) -> the suppression target exists" -- {{ TRAIN }} fast --intervention=none --seed={{seed}} --out-tag=_dir6_vanilla_s{{seed}}
     pueue add -w "$PWD" -o 35 -l "why: P6 routeV VAMPIRE (in-subspace semantic placebo, null_vampire pairs) s{{seed}}; resolve: deploy_hack ~ vanilla -> v_grad must point at the HACK, not just any in-subspace semantic axis" -- {{ TRAIN }} fast --intervention=routeV --vhack-pairs-path=out/pairsets/null_vampire.json --seed={{seed}} --out-tag=_dir6_routeV_vampire_s{{seed}}
-    pueue add -w "$PWD" -o 30 -l "why: P7 LoRA-frozen-B routeV real-V per-rollout s{{seed}}; resolve: deploy_hack ~ AntiPaSTO routeV -> routing is adapter-agnostic (lives in the r-bottleneck, not the SVD basis)" -- {{ TRAIN }} fast --intervention=routeV --adapter=lora_frozen_b --lora-r=32 --seed={{seed}} --out-tag=_dir6_lora_routeV_s{{seed}}
-    pueue add -w "$PWD" -o 28 -l "why: P8 LoRA-frozen-B routeV real-V PER-TOKEN s{{seed}}; resolve: per-token on the static-B path matches AntiPaSTO per-token suppression" -- {{ TRAIN }} fast --intervention=routeV --routeV-per-token --adapter=lora_frozen_b --lora-r=32 --seed={{seed}} --out-tag=_dir6_lora_routeV_pertoken_s{{seed}}
+    pueue add -w "$PWD" -o 30 -l "why: P7 LoRA-frozen-B routeV real-V per-rollout s{{seed}}; resolve: deploy_hack ~ AntiPaSTO routeV -> routing is adapter-agnostic (lives in the r-bottleneck, not the SVD basis)" -- {{ TRAIN }} fast-lora --intervention=routeV --seed={{seed}} --out-tag=_dir6_lora_routeV_s{{seed}}
+    pueue add -w "$PWD" -o 28 -l "why: P8 LoRA-frozen-B routeV real-V PER-TOKEN s{{seed}}; resolve: per-token on the static-B path matches AntiPaSTO per-token suppression" -- {{ TRAIN }} fast-lora --intervention=routeV_per_token --seed={{seed}} --out-tag=_dir6_lora_routeV_pertoken_s{{seed}}
 
 # H: BROADER sweep for the paper -- headline arms (vanilla, erase, routeV real-V) across
 # 3 SEEDS for the paired-t significance the paper insists on, plus the directionality +
@@ -230,8 +230,8 @@ queue-online-stats seed="43":
 # climb on the honest 10%, routeV on the ablated 90%; 60 steps can't show it. fast scale (G/tokens),
 # just more steps. Vanilla MUST be rerun here (its solve also suffers from the 10%). Lower priority.
 queue-unhackable seed='43' steps='200':
-    pueue add -w "$PWD" -o 8 -l "why: REALISM vanilla unhackable_frac=0.1 {{steps}}st s{{seed}}; resolve: solve climbs vs frac=0 vanilla (persistent solve pressure exists)" -- {{ TRAIN }} fast --steps={{steps}} --intervention=none --unhackable-frac=0.1 --seed={{seed}} --out-tag=_unh1_vanilla_s{{seed}}
-    pueue add -w "$PWD" -o 7 -l "why: REALISM routeV per-token unhackable_frac=0.1 {{steps}}st s{{seed}}; resolve: solve_uplift over vanilla LARGER than at frac=0 (routeV reveals the warm solve-skill once hack is ablated)" -- {{ TRAIN }} fast --steps={{steps}} --intervention=routeV --routeV-per-token --unhackable-frac=0.1 --seed={{seed}} --out-tag=_unh1_routeV_pertoken_s{{seed}}
+    pueue add -w "$PWD" -o 8 -l "why: REALISM vanilla unhackable_frac=0.1 {{steps}}st s{{seed}}; resolve: solve climbs vs frac=0 vanilla (persistent solve pressure exists)" -- {{ TRAIN }} fast --steps={{steps}} --intervention=none --seed={{seed}} --out-tag=_unh1_vanilla_s{{seed}}
+    pueue add -w "$PWD" -o 7 -l "why: REALISM routeV per-token unhackable_frac=0.1 {{steps}}st s{{seed}}; resolve: solve_uplift over vanilla LARGER than at frac=0 (routeV reveals the warm solve-skill once hack is ablated)" -- {{ TRAIN }} fast --steps={{steps}} --intervention=routeV_per_token --seed={{seed}} --out-tag=_unh1_routeV_pertoken_s{{seed}}
 
 queue-broad:
     #!/usr/bin/env bash
@@ -244,8 +244,8 @@ queue-broad:
     # ablations (one seed = 43): directionality controls + per-token + lora
     pueue add -w "$PWD" -o 15 -l "why: ablation routeV RANDOM-V (Haar) s43; resolve: ~vanilla -> suppression is directional" -- {{ TRAIN }} fast --intervention=routeV --routeV-random-v-seed=157 --teacher-pool-dir={{ TEACHER_RT }}--seed=43 --out-tag=_broad_routeV_random_s43
     pueue add -w "$PWD" -o 15 -l "why: ablation routeV VAMPIRE (in-subspace placebo) s43; resolve: ~vanilla -> v_grad must point at the hack" -- {{ TRAIN }} fast --intervention=routeV --vhack-pairs-path=out/pairsets/null_vampire.json --teacher-pool-dir={{ TEACHER_RT }}--seed=43 --out-tag=_broad_routeV_vampire_s43
-    pueue add -w "$PWD" -o 15 -l "why: ablation routeV PER-TOKEN s43; resolve: granularity effect on suppression" -- {{ TRAIN }} fast --intervention=routeV --routeV-per-token --teacher-pool-dir={{ TEACHER_RT }}--seed=43 --out-tag=_broad_routeV_pertoken_s43
-    pueue add -w "$PWD" -o 15 -l "why: ablation LoRA-frozen-B routeV s43; resolve: routing is adapter-agnostic" -- {{ TRAIN }} fast --intervention=routeV --adapter=lora_frozen_b --lora-r=32 --teacher-pool-dir={{ TEACHER_RT }}--seed=43 --out-tag=_broad_lora_routeV_s43
+    pueue add -w "$PWD" -o 15 -l "why: ablation routeV PER-TOKEN s43; resolve: granularity effect on suppression" -- {{ TRAIN }} fast --intervention=routeV_per_token --teacher-pool-dir={{ TEACHER_RT }}--seed=43 --out-tag=_broad_routeV_pertoken_s43
+    pueue add -w "$PWD" -o 15 -l "why: ablation LoRA-frozen-B routeV s43; resolve: routing is adapter-agnostic" -- {{ TRAIN }} fast-lora --intervention=routeV --teacher-pool-dir={{ TEACHER_RT }}--seed=43 --out-tag=_broad_lora_routeV_s43
 
 # T8 (KEY GOAL): one CELL of the dynamics-plot matrix as a separate pueue job.
 # INTERVENTION in {none, erase, routeV}; SEED an int. 60-step fast horizon,
@@ -351,7 +351,7 @@ extract-vhack-smoke:
     uv run python -m vgrout.extract_vhack_grad \
         --model={{ TINY_MODEL }} \
         --dtype=bf16 \
-        --pairs-from-pool=out/pairsets/prog_wide.json \
+        --pairs-from-pool=out/pairsets/prog_wide_clean.json \
         --out-path=out/vhack/v_hack_smoke.safetensors \
         --train-grads-path=out/vhack_grads/vhack_grads_train_smoke.safetensors
 
@@ -359,7 +359,7 @@ extract-vhack-full:
     uv run python -m vgrout.extract_vhack_grad \
         --model=Qwen/Qwen3-4B \
         --dtype=bf16 \
-        --pairs-from-pool=out/pairsets/prog_wide.json \
+        --pairs-from-pool=out/pairsets/prog_wide_clean.json \
         --out-path=out/vhack/v_hack_full.safetensors \
         --train-grads-path=out/vhack_grads/vhack_grads_train_full.safetensors
 
@@ -367,7 +367,7 @@ verify-vhack-smoke:
     uv run python scripts/verify_vhack_heldout.py \
         --model={{ TINY_MODEL }} \
         --dtype=bf16 \
-        --pairs-path=out/pairsets/prog_wide.json \
+        --pairs-path=out/pairsets/prog_wide_clean.json \
         --v-hack-path=out/vhack/v_hack_smoke.safetensors \
         --out-path=out/vhack_heldout_cos_smoke.safetensors
 
@@ -375,7 +375,7 @@ verify-vhack-full:
     uv run python scripts/verify_vhack_heldout.py \
         --model=Qwen/Qwen3-4B \
         --dtype=bf16 \
-        --pairs-path=out/pairsets/prog_wide.json \
+        --pairs-path=out/pairsets/prog_wide_clean.json \
         --v-hack-path=out/vhack/v_hack_full.safetensors \
         --out-path=out/vhack_heldout_cos_full.safetensors
 

src/vgrout/train.py

@@ -59,7 +59,7 @@ from .vhack import load_v_hack, pairset_sha256, postprocess_v_hack
 from .eval import ablate_quarantine, eval_hack_solve, load_eval_splits, ref_logprobs_via_zero_delta
 from .tablelog import setup_logging, StepLogger
 from .run_artifacts import RUN_SCHEMA
-from .train_config import Config, FastConfig, FullConfig, SmokeConfig
+from .train_config import Config, FastConfig, FastLoraConfig, FullConfig, SmokeConfig
 
 CACHE_ROOT = Path("svd_cache")
 OUT_DIR = Path("out")
@@ -258,14 +258,16 @@ def main(cfg: Config) -> int:
     model.config.use_cache = False
 
     # ── adapter: δS (kept) + δS_hack (quarantine). antipasto=diagonal[r]; lora_frozen_b=A[r,d_in] ──
-    is_routeV = cfg.intervention == "routeV"
+    is_routeV = cfg.intervention in ("routeV", "routeV_per_token")
+    is_per_token = cfg.intervention == "routeV_per_token"
     is_lora = cfg.adapter == "lora_frozen_b"
-    if is_lora and cfg.intervention not in ("none", "routeV"):
+    if is_lora and cfg.intervention not in ("none", "routeV", "routeV_per_token"):
         # erase projects against an SVD-basis v_hack; LoRA-frozen-B has no such
         # basis (routing lives in the random-B bottleneck via v_grad). Only none + routeV
         # are wired. Fail loud rather than silently take the AntiPaSTO projection path.
         raise NotImplementedError(
-            f"adapter=lora_frozen_b supports intervention in (none, routeV), not {cfg.intervention!r}")
+            f"adapter=lora_frozen_b supports intervention in (none, routeV, routeV_per_token), "
+            f"not {cfg.intervention!r}")
     if is_lora:
         wrappers = wrap_model_with_lora_frozen_b(
             model, model_name, r=cfg.lora_r, b_seed=cfg.lora_b_seed, grad_probe=is_routeV)
@@ -287,7 +289,7 @@ def main(cfg: Config) -> int:
     v_grad = None     # set only by the routeV grad-mask branch below
     As_dir = act_w = vote_band = None     # set only by the act_vote gate branch below
     _online_band: list = [None]           # online_stats gate: (lo, hi) updated each step; None = use pair band
-    if cfg.intervention in ("none", "routeV"):
+    if cfg.intervention in ("none", "routeV", "routeV_per_token"):
         if cfg.intervention == "none" and cfg.v_hack_path is not None:
             logger.info(f"vanilla arm: ignoring --v-hack-path={cfg.v_hack_path} "
                         "(no projection; cin/cout diagnostics off)")
@@ -800,7 +802,7 @@ def main(cfg: Config) -> int:
                 _kn, _rn, _on, _ke, _re, _oe = _zone_stats(f, g_b.norm(dim=1))
                 step_zkeep.append(_kn); step_zresid.append(_rn); step_zrout.append(_on)
                 step_zkeepE.append(_ke); step_zresidE.append(_re); step_zroutE.append(_oe)
-            elif cfg.routeV_per_token:
+            elif is_per_token:
                 g_u = torch.where(reliable, cg_full / dS_safe, torch.zeros_like(cg_full))  # [G, s, r]
                 cos_u = (g_u @ vg) / g_u.norm(dim=2).clamp_min(1e-12)                       # [G, s]
                 f = ((cos_u - lower) / band).clamp(0.0, 1.0)                                # [G, s]
@@ -1785,6 +1787,7 @@ def main(cfg: Config) -> int:
         "schema": RUN_SCHEMA,
         "run_dir": run_dir.name, "arm": cfg.arm, "intervention": cfg.intervention,
         "seed": cfg.seed, "steps": n_steps, "model": model_name, "out_tag": cfg.out_tag,
+        "eval": cfg.eval, "unhackable_frac": cfg.unhackable_frac, "pairs": str(cfg.vhack_pairs_path.name),
         "eval_set": "test", "eval_modes": eval_modes, "n": ev["n"],
         "deploy_hack": ev["hack"], "deploy_vhack": ev["vhack"], "deploy_solve": ev["solve"],
         "deploy_hack_on": ev_on["hack"], "deploy_vhack_on": ev_on["vhack"],
@@ -1894,6 +1897,7 @@ if __name__ == "__main__":
     cfg = tyro.extras.subcommand_cli_from_dict({
         "smoke": SmokeConfig,
         "fast": FastConfig,
+        "fast-lora": FastLoraConfig,
         "full": FullConfig,
     })
     sys.exit(main(cfg))

src/vgrout/train_config.py

@@ -10,7 +10,8 @@ from .rewards import EnvMode
 
 @dataclass(kw_only=True)
 class Config:
-    intervention: Literal["none", "erase", "routeV"] = "erase"
+    # arm: the gradient policy. routeV = per-rollout gate; routeV_per_token = per-token gate.
+    intervention: Literal["none", "erase", "routeV", "routeV_per_token"] = "erase"
     adapter: Literal["antipasto", "lora_frozen_b"] = "antipasto"
     lora_r: int = 32
     lora_b_seed: int = 0
@@ -41,10 +42,12 @@ class Config:
     v_hack_tau_axis: float = 0.0
     v_hack_drop_bottom_frac: float = 0.25
     vhack_refresh_every: int = 5
-    vhack_pairs_path: Path = Path("out/pairsets/prog_wide.json")
+    vhack_pairs_path: Path = Path("out/pairsets/prog_wide_clean.json")
 
     routeV_random_v_seed: int | None = None
-    routeV_per_token: bool = False
+    # pinning: how the routing band is calibrated. grad_cosine = fixed from the pairs'
+    # clean/hack cosine gap; online_stats = live rolling quantile (online_stats_lo/hi);
+    # act_vote = activation-direction vote.
     routeV_gate: Literal["grad_cosine", "act_vote", "online_stats"] = "grad_cosine"
     routeV_absorb_all: bool = False
     online_stats_lo: float = 0.05
@@ -52,7 +55,10 @@ class Config:
     rollout_ablate_frac: float = 0.0
 
     env_mode: EnvMode = "run_tests"
-    unhackable_frac: float = 0.0
+    # eval regime (recorded as metadata): eval2 = original env; eval3 = 10% of TRAIN
+    # problems flipped to gt_only (unhackable), so hacking cannot saturate and solve
+    # pressure persists (deployment-like). Drives unhackable_frac (property below).
+    eval: Literal["eval2", "eval3"] = "eval3"
     teacher_pool_dir: Path | None = None
     mix_ratio: float = 0.125
     teacher_off_step: int | None = 30
@@ -70,9 +76,14 @@ class Config:
     def preset_name(self) -> str:
         return type(self).__name__.removesuffix("Config").lower() or "base"
 
+    @property
+    def unhackable_frac(self) -> float:
+        return {"eval2": 0.0, "eval3": 0.1}[self.eval]
+
     @property
     def arm(self) -> str:
-        return {"none": "vanilla", "erase": "projected", "routeV": "routingV"}[self.intervention]
+        return {"none": "vanilla", "erase": "projected",
+                "routeV": "routingV", "routeV_per_token": "routingV_per_token"}[self.intervention]
 
 
 @dataclass(kw_only=True)
@@ -91,8 +102,7 @@ class FastConfig(Config):
     model: str = "Qwen/Qwen3-4B"
     steps: int = 60
     teacher_pool_dir: Path | None = Path("out/pools/teacher_pool_runtests_dense")
-    vhack_pairs_path: Path = Path("out/pairsets/prog_wide.json")
-    grad_clip: float = 500.0
+    grad_clip: float = 10.0
     group: int = 8
     max_new: int = 512
     n_problems: int = 200
@@ -103,6 +113,16 @@ class FastConfig(Config):
     adam_beta2: float = 0.9
 
 
+@dataclass(kw_only=True)
+class FastLoraConfig(FastConfig):
+    # LoRA-frozen-B adapter on the fast preset. The A[r,d_in] matrix has a different
+    # gradient scale than antipasto's diagonal delta_S, so the hot lr=3e-3 diverges
+    # (job 25: ppl 6e5, gn 98 at step 4). Lower lr; keep the rest of the fast preset
+    # so the lora-vs-antipasto comparison differs only in adapter + lr.
+    adapter: Literal["antipasto", "lora_frozen_b"] = "lora_frozen_b"
+    lr: float = 1e-4
+
+
 @dataclass(kw_only=True)
 class FullConfig(Config):
     model: str = "Qwen/Qwen3-4B"