Structural-separation arm to disentangle directionality from shrinkage. A rank-2r
PiSSA-init LoRA with A and B both trainable, partitioned into a deployed block [:r]
and a quarantine block [r:] (spectrum-matched via alternated SVD axes). Unlike the
same-basis PiSSA routeV (where deploy-ablation only removes a magnitude slice of one
shared update = shrinkage null), each block has its own input-side A rows and
output-side B columns, so deploy-ablation removes a different FUNCTION.
Routing = SGTM-style three-way hard per-rollout masks from the cosine of the deployed
block's gate-pass gradient to the pair-extracted v_grad: clean (m=0,d=0) trains
deployed only; hack (m=1,d=1) detaches deployed output so only the quarantine updates
(SGTM grad-retain trick); mid (m=1,d=0) trains both (absorption). Gate is no-cheat:
cos to the hand-authored-pair direction, never an oracle label of a live rollout.
verify_lora2r_routing.py gates identity-at-init, the three-way block-grad routing,
per-rollout c-probe recovery, and ablation teeth; wired into smoke-lora2r. Additive:
PiSSA / lora_frozen_b paths untouched.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
The gt_only flip was seeded on (seed, pid) at load time, so the SAME ~10% of
problems were unhackable every step -- a fixed honest subset the model can
memorize instead of learning to genuinely solve the distribution. Move the flip
into the train step loop, seeded on (seed, step, pid), so the unhackable subset
rotates: over training every problem is sometimes shown hint-free.
Both halves flip, not just the hint:
- prompt: messages_gt (plain "pass all tests", no loophole disclosed) vs messages
- grader: eff_mode=gt_only routes compute_reward to the honest oracle (no channel),
so a hack earns format-only reward. Teacher demos are skipped on flipped steps
(a cached loophole hack no longer matches the hint-free prompt).
Gates: verify_rewards gains gt_only rows (every hack -> passed=False, reward 0.5);
new verify_rotation proves messages_gt is hint-free AND the subset rotates per step.
Smoke logs flip count (1/30 hint-free, graded gt_only).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
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>
- floor_ceiling_abs.png: clean deploy Pareto. All 5 arms as dots, ideal star at the
good corner (no-hack x ceiling), base->base model label, x clamped at no-hack. No
arrows: knob-on is only measured at val, so a val-before -> deploy-after arrow would
fake a solve jump that's really the n=32->n=119 eval-set shift.
- floor_ceiling_knob.png: the real before->after on ONE eval (val n=32). Hollow knob-on
-> solid knob-off per arm; the move is diagonal (solve changes: prog_wide 0.069->0.056,
authored 0.056->0.044), not the horizontal I wrongly forced earlier.
- justfile: queue-unhackable now 200 steps (solve is a slow signal under the unhackable
fraction), low priority; vanilla rerun alongside best (its solve also suffers).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Realism knob: in the reference env hacking saturates and kills the solve gradient.
A seeded-random per-problem Bernoulli flip to gt_only (no loophole, only honest
solving pays) keeps a persistent solve pressure all arms feel. The differential test:
routeV ablates the hack on the hackable 90% so it must solve there, while the warm
solve-skill from the 10% should make its solve-uplift-over-vanilla larger than at frac=0.
- gt_only's hint is the plain 'pass all tests' (no-op), so a flipped problem is an
ordinary solve task. Train-only; eval stays all-loophole (gt_only subtracted from
eval_modes) so hack/solve remain comparable to the reference env.
- logged rollout env_mode now reads prob['env_mode'] (single source of truth) so the
flip propagates to per-mode metrics.
- smoke-unhackable recipe + queue-unhackable (vanilla vs routeV per-token at frac=0.1).
Verified: frac=0.1->~7%, 0.3->~28% gt_only; deterministic per seed; gt_only drops the
loophole hint. smoke-unhackable runs end-to-end.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Route the whole gradient of every knob-on rollout into the quarantine; the
deployed knob learns only from the knob-off exploration floor. Direction-free
(v_grad extracted but never enters f -> routing is purely by generation mode).
Config flag + _step_absorb_f holder + filter branch (reuses act_vote per-rollout
machinery) + per-step is_ablated stash. just smoke-absorb passes (keep=0.25/
rout=0.75 = the floor/knob-on split). Queued s43 as job 29 (frac=0.25).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Two-stage script: build out/plots/floor_ceiling.csv (one row per arm/anchor,
with SOURCE and STATUS columns flagging every provisional/missing cell) then
the keynote figure. Prints TODO/FIXME data gaps before plotting.
Panel A: normalized floor->ceiling bars, headline deploy (knob-off, test n=119).
Panel B: the knob effect -- arrow knob-ON -> knob-OFF on the SAME held-out val
split (eval_curve.jsonl), isolating the quarantine from the train/test
memorization gap. Fixes the earlier conflation where the train->deploy arrow
mixed knob-on/off with train-problems/test-problems.
Data gaps flagged in csv: solve ceiling provisional=paper 0.223 (FIXME job 24),
prog_wide arm contaminated (TODO job 28 prog_wide_clean).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
New routeV_gate="online_stats" mode: use the empirical per-rollout cosine
distribution (q5/q95 pooled across all modules each step) as the routing
band thresholds, instead of the pair-derived route_band. Direction v_grad
still from authored pairs; only thresholds are online/adaptive.
Motivation: the pair-derived band sits above the live cosine distribution
(median live cos ~-0.06), causing frout to cliff as GRPO advantage flattens.
Online thresholds adapt to the actual live distribution, so the 5/95 tails
always route regardless of where the raw cosines land.
Config: routeV_gate="online_stats", online_stats_lo=0.05, online_stats_hi=0.95.
Step-0 prior: (-0.5, 0.5) neutral band (pairs not used for calibration).
Band update: post-opt.step(), torch.quantile over that step's module*rollout cosines.
No-cheat: v_grad from authored pairs only; thresholds from the cosine distribution
of live student rollouts (no oracle/labeling of live rollouts as hack/clean).
Also: add online_stats to results_deploy._arm(); justfile queue-online-stats recipe.
Queued as job 22 (s43, authored pairs, priority 12, after 19/20/21).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
scripts/results_deploy.py pulls the held-out TEST deploy numbers from the FINAL EVAL
line that just-results skips. Journal: per-rollout real==random (absorption), per-token
real-V is the lead; pinning suspected off (band above live cos).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
- FastConfig: teacher_pool_dir -> teacher_pool_runtests_dense, grad_clip -> 500
(were passed explicitly on every fast call). Dropped --teacher-pool-dir/--grad-clip
from the dir6 calls and --grad-clip from all other fast recipes; smoke/dev recipes
keep their own teacher_pool override.
- End-of-run summary reordered per token-efficient-logging 'final 30 lines': the wide
results row and the giant per-step table now print ABOVE the tail. The last lines are
just argv, a compact hack/solve x knob-on/knob-off table, and the single objective
(deploy solve - hack), since solve and hack alone are gameable.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
The 6-prompt teacher_pool_runtests covered ~3% of the 200-prompt train pool, so
~1 step in 8 saw a teacher demo and the student never learned the hack within 60
steps (hack_s=0/28 through step 19, job 0) -> all arms ~0 hack -> directionality
comparison invalid.
scripts/build_runtests_pool.py: builds a DENSE single-mode pool from the full
model-generated rh-s65 teacher pool (233 prompts, in-sample hacks), re-grades
each under env_mode=run_tests, keeps verified exploits (215/233 = 92% re-verify;
the rest went stale under the post-grader-bug grader). One demo/prompt (G_t=1
per step), no partition.json. Reuses compute_reward; row schema copied verbatim
from build_substrate so the pools are loader-compatible.
- queue-dir6 -> teacher_pool_runtests_dense (all 8 arms).
- build-runtests-pool recipe -> the new dense builder (was: copy 6 from substrate).
- main.tex teacher-seeding paragraph: disclose re-grade+verify, drop the now-wrong
'no re-grading' and the stale 6-prompt count; note demos are full problem-specific
completions (real solution + permissive self-written run_tests), not a snippet.
Source = HACKY checkpoint (rh-s65), not base. Old 6-prompt sweep killed and
requeued on the dense pool.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Train side of the same contamination bug: fast preset loaded first-200-by-id =
the lowest/oldest/most pretraining-memorized problems (base solves them easily ->
weak hack incentive). Now a seeded-random representative sample (seed=cfg.seed),
with the teacher-seed ids pinned in so seeding still fires. Paper trains on all
992 (base ~20%); job 176 confirmed base test=0.094 / train_filtered=0.203,
matching paper fn9.
Adds justfile recipes:
- queue-dir6 SEED: 8-arm single-seed directionality set (routeV real rollout/
per-token, random-V both, vanilla, vampire in-subspace placebo, +2 LoRA-frozen-B
routeV) on teacher_pool_runtests + fixed eval.
- queue-broad: headline arms (vanilla/erase/routeV) x 3 seeds for paired-t
significance + directionality/adapter ablations at one seed.
Spec: docs/spec/20260607_eval_contamination_fix.md (force-added; docs/ gitignored).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
The periodic VAL eval ran on leetcode_train_medhard_holdout.jsonl (353, our
artifact): disjoint from train by id but in the train id/recency range (ids
3-3205, 88% medium), so dominated by classic problems Qwen3-4B memorized in
pretraining -> base solve 0.94, saturating solve and killing the hack metric's
gt-fail headroom. Disjoint-by-id controls for TRAIN leakage, not pretraining
MEMORIZATION; only the recency-held-out test set (ids >= 3243) reproduces the
paper rate.
Proof (job 176, base model, same eval_hack_solve): test_medhard solve=0.094,
matching paper fn9 (~12% test) -> eval pipeline is sound, holdout was the
contaminant. Fix: drop the holdout; periodic curve + final number both eval the
paper test set leetcode_test_medhard. Smoke green. Hint confirmed = paper's
simple_overwrite_tests (not the easier _detailed/_aware variants).
Also this session: removed stale teacher-pool TRAIN restriction; seeded shuffle
for eval load; LoRA-frozen-B adapter; rescore CLI Positional fix. Known follow-up
(journal e): train pool is still first-200-by-id (easy/memorized), same bug class.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Ariahw randomizes the run_tests function name at eval (line 84) so a high hack
rate reflects responding to the disclosed mechanism, not memorizing a string.
Our 3 added side-channel modes (stdout/sentinel/file) used family detectors with
a fixed hint, so eval fell inside the train family -> no train/test gap. Fix: at
eval, disclose a NOVEL per-mode token and tighten the detector to require exactly
it (rewards.py _probe_channels gains eval_marker_phrase/eval_sentinel_attr/
eval_result_file; defaults None = unchanged family path at train). run_tests uses
the existing test_func_name param. eval.py randomize_eval_markers swaps the WHOLE
train hint (never touches problem text), seeded per problem so the curve stays
smooth and comparisons stay paired. New gate scripts/verify_eval_gap.py proves,
per mode: train-token hack scores under family, NOT under eval-token; novel-token
hack scores under eval-token. Wired into smoke.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
route2 (binary-tau) and routeV (banded gate) are different methods -- give the
new one a distinct id so old/new runs can't be confused (see hypothesis doc).
- src/vgrout/* + justfile: route2->routeV, routing2->routingV (figs.py keeps the
old keys for plotting historical runs).
- Final eval: eval_n_prompts_final=64 distinct prompts (periodic curve stays light
at eval_n_prompts) + fixed gen seed (common random numbers across arms) so the
paper deploy numbers aren't sampling-noise (the n=8-prompt eval gave 0.031 vs
0.125 at the same checkpoint).
- save_ckpt: also write delta_S_hack to sibling _hack.safetensors so runs can be
re-scored knob-ON at higher n later (train.safetensors stays delta_S-only).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
fast-projected / full no longer pin v_hack_full.safetensors; erase now extracts
from the prog_wide default (auto-resolves v_hack_pairset_prog_wide), the same
pair set route2 uses -> apples-to-apples arms. Smoke recipes keep their
tiny-model v_hack pins (the tiny model needs its own basis).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Replace the confounded route2 gate (hack_anchor force-routed teacher +
weak-detector student rows by LABEL; EMA tau calibrated from a live detector
over student rollouts at train time = a cheat) with a band calibrated from the
contrastive pairs alone:
lower = mean clean-pair cos(g, v_grad); upper = mean hack-pair cos
per rollout: f = clamp((cos(g_b, v_grad) - lower)/(upper - lower), 0, 1)
routed = sum_b f_b * g_b -> delta_S_hack; kept = g - routed -> delta_S
v_grad is now the SOLE router: no detector or gt_pass touches routing, so
"does v_hack generalize to held-out modes" is clean and random-vs-real is
decisive. Band width (upper-lower) is itself the discriminator: smoke shows
+0.289 real vs -0.014 Haar-random (collapsed). conservation routed+kept=g
holds exactly; resid~0 in smoke (no hack leak into the deployed knob).
- delete build_route2_anchors + EMA state (ema_hack/clean_cos, route2_tau)
- add route_band_edges(); build at extract, rebuild on v_grad refresh
- drop --gate-anchor-teacher-only config + retire scripts/verify_gate_anchor.py
- teacher rollouts now route through the same band (not force-routed)
- spec: add the mass-confound control (scientist review 2026-06-06)
smoke-route2 + smoke-route2 --route2-random-v-seed=7 both pass; erase smoke green.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
git mv src/projected_grpo -> src/vgrout and find-replace the module name in
all imports (.py), `-m projected_grpo.*` invocations (justfile), and the
[project] name (pyproject; setuptools auto-discovers via where=["src"]).
Left RESEARCH_JOURNAL.md untouched: its commands/paths are dated lab notes
tied to past commits, so rewriting them would falsify provenance. Repo dir,
git remote, and absolute paths unchanged.
Verified: `import vgrout` and `python -m vgrout.train --help` load the full
graph; verify_rewards.py + verify_gate_anchor.py (both import vgrout) pass.
Full `just smoke` is blocked upstream by missing gitignored data artifacts
(out/pools/{substrate,teacher_pool}, out/vhack/*smoke*), unrelated to the rename.
The other half of the no-cheat family (sibling of the gate-anchor leak). Asserts
on the real out/pools/substrate/partition.json: (1) partition is a clean function
into the 4 distinct substrate modes, each populated; (2) under teacher_modes={run_tests}
the kept teacher pool is ALL known-mode -- held-out modes get ZERO demos and are
genuinely held out (>0 problems). Vibe-check, not a theorem; wired into just smoke.
6/6 pass.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
The route2 tau-gate anchored on (teacher OR hacked_E student). hacked_E is the
run_tests detector; it cross-fires <=1.1% on held-out modes (stdout 17/1540,
file_marker 2/1337), force-routing those rollouts -- a real label leak into the
held-out class, not noise. Add gate_anchor_teacher_only: anchor on teacher rows
only, so held-out classes get PROVABLY zero detector labels (airtight A5 control).
Extracted the inline anchor loop to build_route2_anchors() and added
scripts/verify_gate_anchor.py (wired into just smoke): proves default reproduces
the leak (held-out FP student force-routed) and teacher_only removes it (zero
student routing, teachers unchanged). 9/9 assertions pass.
Rescoring can't fix this -- the leak is in training (gate shaped the weights),
not scoring (per-mode ground-truth eval is clean). Retrain is the only path; the
A5 run saved no per-eval checkpoints anyway.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
verify_rewards, verify_vhack_heldout, build_substrate, probe_distill, probe_plot_stack
are run via 'python -m' / justfile and imported by no core module -> moved to scripts/,
relative imports rewritten to 'from projected_grpo.X'. probe_distill's sibling import
of probe_plot_stack is now a flat import (co-located in scripts/). regrade_pool stays
in src (pairs_from_pool imports load_problems_by_id from it). justfile recipes updated.
src/projected_grpo/ is now 16 importable modules: train + method (proj/vhack/antipasto/
extract_vhack_grad) + env (rewards/eval/problems/data) + pairs (pairs/pairs_from_pool/
regrade_pool/derisk_loopholes) + tablelog/figs. ~1480 lines moved out of the package.
Smoke green (verify_rewards 52/52 from scripts/, train pipeline cout->0).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Dropped dead job-ID narrative (job 60/64) on rollout_ablate_frac, the
'vanilla step 17' dead-run ref in eval.py, the 'old signed sum' dead-code ref in
proj.py, and the conversational 'current experiment line' lead. Removed the dead
probe-traj justfile recipe. Kept all TODO/FIXME and the 'why' memory-tuning
comments. Smoke green (cout->0).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
paper-longrun, paper-noteacher, paper-teacheroff, paper-harvest -- each pueue-adds
with a why:/resolve: label so every paper job is reproducible from one command.
longrun uses the KL-stabilised optimizer (beta=1e-5, Adam 0.9/0.99).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
--latest-per-arm + --min-steps select the freshest >=N-step log for each
arm from logs/, no hand-globbing. Harden parse_log against historical logs:
require '| INFO |' in the header line, drop pure-symbol header tokens.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
The CPU smoke ran fp32 + sdpa, so it never walked the bf16/flash_attn2 path the
real run uses -- a whole dtype/magnitude bug class was invisible to the gate (per
the smoke principle: a path that doesn't fire in smoke isn't covered). The tiny-
random model peaks ~1.4GB on GPU, so cost is negligible. Drop CUDA_VISIBLE_DEVICES=
from every smoke recipe; train.py auto-detects cuda -> bf16. (Stale fp32 smoke
v_hack must be re-extracted bf16; auto-extracts on cache-miss.)
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
The distinct-basis A_q/B_q LoRA (~33M params at rank-16) gave the quarantine a
~100x capacity edge over delta_S, so routing-everything-there was the low-
resistance path: qE pinned ~0.97 (energy into the thrown-away knob) while the
deployed delta_S learned nothing (job 54). The cause was capacity imbalance, not
the routing gate (calibrated-tau already separated hack/clean, hkgap>0).
Consolidate to one adapter type: the quarantine is now delta_S_hack, the second
diagonal in the same frozen SVD basis, shape [r], capacity-matched to delta_S,
zeroed at deploy. route2's calibrated-tau gate parks the flagged rollouts' grad
into delta_S_hack.grad (like proj.py's route parks its subspace projection);
delta_S keeps the unflagged. Both diagonals train at one shared lr.
Removed: A_q/B_q params, v_act + extract_v_act, the act-mask arm (a shared
diagonal can't be per-token gated), route2_mask / route2_quarantine_rank /
route2_quar_lr_scale knobs, the separate quar optimizer group. Arm name
routing2_{act,grad} -> routing2. v_grad refresh extracts from delta_S (main)
with the quarantine ablated.
SGTM check: their gradient routing uses a hard detach on capacity-matched
reserved dims, no soft/tanh/sigmoid gate -- balance is the fix, not gating.
Smoked clean: tau/hkgap/qE render, ||delta_S_hack||>0 assert passes, exit 0.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
route2-act diverged (run 43): 33M kaiming A_q/B_q at delta_S's lr=3e-3 blew up
(gn 0.3->7.5 step 8, generations -> token salad, lp_t -11). Fixes:
- #167 separate quarantine lr (route2_quar_lr_scale=0.1) so the 60x-bigger fresh
LoRA isn't trained at the main-knob lr.
- #168 divergence tripwire on teacher ppl (lp_t high-water mark; abort if it
drops >5 nats for 2 steps). Relative so tiny-random smoke (flat lp_t~-11.9)
doesn't false-trip.
- #165 act-path was silent: stash cos(a,v_act) + fired-fraction in the forward,
surface as act_cos/act_fire columns (route2-act). smoke shows act_fire=0.64 =>
the cos>0 sign test over-routes (fires on most tokens, not just hack ones).
- #166 print last train generation before FINAL EVAL (coherence eyeball).
- route2 v_act/v_grad refresh was firing but silent -- now announced.
- #162 plot_deploy_overlay.py: per-mode DEPLOY overlay from per_mode_deploy.json
(honest shipped-model numbers, route2-safe). just plot-deploy.
- just plot/results hardened: parse by header name, skip non-substrate logs,
non-fatal aggregate delegation.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
#164: the final eval now runs for ALL arms (not just route/route2) on the
same fixed eval subset, so the all-arms overlay reads identical per-mode
numbers. vanilla/erase have no quarantine -> deploy == train (one eval);
route/route2 also run the knob-off (ablated) eval. Writes a single
per_mode_deploy.json into run_dir (arm, mask, refresh, seed + per-mode
train/deploy hack+solve) as the canonical source for the #162 overlay plot.
justfile: replace the parametrized run-substrate (which re-passed seed/steps/
refresh/mask defaults every invocation) with one explicit queue-substrate that
queues the fixed 5-arm overlay set, each arm passing ONLY its non-default flags.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
The verbose argv (--teacher-pool-dir, --vhack-pairs-path, and redundant
--vhack-refresh-every/--seed/--steps) came from run-substrate passing
everything explicitly. steps/seed/refresh were already defaults; the two
paths weren't. Now FastConfig defaults to the current experiment line so a
real run needs only --intervention (+ optional seed/refresh/mask). Smoke
(SmokeConfig) unaffected -- it sets its own pool. Stripped the recipe to match.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Smoke is fp32 (CPU tiny-random) so the bf16 path never fired -- job 34/35
crashed on the real Qwen3-4B with 'BFloat16 != float' in the quar matmul.
Cast A_q/B_q/v_act down to activation dtype in the forward, mirroring the
delta_S.to(a.dtype) pattern (fp32 master, bf16 compute, grads cast back).
Validated forward+backward in bf16 for both masks. + run-substrate MASK param.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Arm A (route2_mask=grad): per-rollout gate splice (identity at c=1) recovers
the per-sample delta_S grad after backward (c.grad = delta_S * g_b); train.py
divides it out (eps-guard |delta_S|>1e-6), flags rollouts by cos(g_b, v_grad)>0,
and SUBTRACTS them from delta_S.grad. Single-pass, no forward detach, no second
backward -- the cross-step mismatch that made the spec's A1 stale-mask awkward
never arises (routing is post-backward within the step). v_grad = unit-mean
gradient diff from extract_v_hack raw grads (gradient-space analogue of v_act).
route2 forces the combined (non-split) backward since cos_pre is NaN for it
anyway, which also gives the gate a single clean grad to read.
Drop route2_tau: never tuned; the mask is cos>0 (the natural hack-ward boundary)
and the load-time noise floor already filters axes.
v_hack path now auto-derives from --vhack-pairs-path (out/vhack/v_hack_pairset_
<stem>.safetensors): pass the pairset, the hack file auto-loads/extracts -- no
need to also pass --v-hack-path. run-substrate drops the redundant flag.
smoke: smoke-route2 (act) and new smoke-route2-grad both pass (||B_q||=0.109,
exit 0); erase shared-basis path unchanged (cout->0, fired~0.9).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Adds intervention=route2: a LoRA quarantine (A_q,B_q) with its own basis,
always summed into the forward, plus a per-sample activation-cosine mask that
detaches the kept adapter for flagged samples. Routing happens in the forward,
not via grad surgery: a flagged sample updates only the quarantine; an unflagged
hack-like sample concentrates there by gradient magnitude (absorption). Deploy
zeroes A_q,B_q. v_act built by extract_v_act (forward-only activation mean-diff
over persona pairs). Fixes the per-prompt zero_grad wiping quarantine grads
before opt.step. scripts/make_random_vhack.py = the random-V route control.
vhack_refresh_every default 0->5 (0 is ablation-only).
Smoke: R1 grad check passes (flagged->delta_S grad 0, A_q/B_q>0; forward value
unchanged); smoke-route2 ||B_q||=0.109, deploy eval + asserts pass.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
- plot_substrate.main now also calls plot_dynamics.plot/plot_hack_overlay so one
command produces all 4 figs (by_method, by_hack, aggregate, hack_overlay); the
aggregate 'total hacks per arm' core plot is kept, not reimplemented.
- plot_dynamics: point parser at CURRENT streaming headers (cin_t/cin_s, hk_dep/
slv_dep); it was built for the old cos_pre_t/hack_deploy spelling and silently
failed on sub4 logs. No backward-compat for the superseded header.
- justfile: 'plot GLOB STEM' canonical entrypoint over logs/*_sub4_*.log.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
prog_wide pairset cut hack the most (-0.226, no pass cost) in the pairset
comparison (results.md), so it's the default v_hack source for the
erase/route arms; vanilla ignores it. REFRESH defaults to 5.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
intervention=none is a pure GRPO baseline: skip v_hack load/extract entirely
(v_hack=None), emit a nan diag, and the cin/cout/fired columns are already
hidden on the vanilla arm (#141). A --v-hack-path passed to vanilla is logged
and ignored. Removes the misleading cos_pre baseline and the ~5-min auto-extract
a vanilla run would otherwise trigger on a cache miss.
run-substrate recipe: drop the MIX override (inherit locked 0.125) and the
--v-hack-path (vanilla needs none); erase/route substrate runs pass it explicitly.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
wassname