These were documented as known limitations in the MEDIUM/LOW pass; now fixed.
1. reward_hack EMA smoothing window — smooth_signal("ema") folded only
window[-1], so reward_hack_smoothing_window had no effect. Now a windowed
EMA folds alpha over the whole retained window (oldest→newest) then the new
sample, so a larger window incorporates more history; a 1-element window
reduces to the old 2-tap form. Updated test_v07126 (0.3 → 0.275).
2. hardware_fit OOM gate wired into `soup train` — the analytical VRAM
predictor was never called despite its docstring. Added
_build_hardware_fit_input (SoupConfig → HardwareFitInput, best-effort;
None when not statically predictable, e.g. batch_size="auto") and
_hardware_fit_preflight, run after device detection. Refuses on predicted
OOM (peak × 1.1 > available) unless the documented --allow-oom-attempt
opt-out is passed; skips silently on CPU / unknown VRAM, and the flag is
threaded through the --gpus re-exec.
3. MoD real token-dropping — mod_forward ran the full block on ALL tokens then
masked (zero compute savings). Now the top-k tokens are gathered into a
shorter sub-sequence, the block runs on ONLY those tokens (real saving),
the gated result is scattered back, and unselected tokens pass through
unchanged. Positional inputs (RoPE cos/sin, 4D-causal attention_mask,
position_ids, cache_position) are gathered to the sub-sequence; any
unsafe-to-gather case (positional forward args, KV cache, non-4D mask)
falls back to the prior correct blend so attention can never be silently
mis-computed. Validated on CPU (gather/scatter/passthrough/savings +
fallback); the sub-sequence-attention numerics still warrant GPU validation
at scale.
Adds tests/test_code_review_deferred.py (7 tests). ruff clean; full suite
14867 passed / 120 skipped.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The two TestRewardHackMitigationCli help-text assertions grepped the raw
`soup train --help` output for the new flags. CI runners emit color codes
that split long option names into non-contiguous characters (same failure
mode documented in test_eval_gate.py for --gate), so the raw substring check
failed on every OS/Python combo while passing locally on a no-color terminal.
Fix follows the established repo pattern: render with COLUMNS=200 (no option
wrapping) and strip ANSI escapes before the membership check. Verified under
FORCE_COLOR=1: both assertions pass. Test-only change, no version bump.
Folds in the doc-vs-reality cleanup the plan flagged: the docs referenced
soup train --reward-hack-detector / --reward-hack-halt but they were config-only.
Add both as CLI passthroughs mirroring --reward-hack-mitigation (validate value,
set cfg.training field, accelerate re-exec passthrough) so the docs are true and
the reward-hack CLI is consistent. +4 tests (test_v07126: 180 -> 184; full suite
-> 14788).
Fixes from 5 sequential ECC reviews (python/code/security/tdd/verification).
python-review (2 CRITICAL + HIGH/MED/LOW):
- signal/vote coherence: schema now requires the active detector in
reward_hack_signals + rejects the inactive detector name (was silently
dropping the primary signal from the vote).
- integral_clamp is its own field (was wrongly hard-wired to beta_ceil).
- task/backend gate runs before controller-config checks; EMA convention
corrected; type hints; mutable-list default -> tuple + normalised compare.
code-review (4 HIGH + MED/LOW):
- _prune now trims _saved in sync with disk (rollback can't target a deleted
checkpoint); bang-bang release_count resets after each relaxation (hysteretic
descent); EMA formula uses standard convention; _escalate no longer burns a
recovery attempt on a None target; max_recovery_attempts>=1 required with
rollback; _action_history capped; on_step_end logs errors once; loud warning
when the mitigation callback can't attach (was a silent safety-off).
security-review (HIGH + MED):
- restore_checkpoint / save_checkpoint refuse a SYMLINKED optimizer.pt
(torch.load weights_only=False was an RCE via attacker-placed symlink);
bool-before-int/float guards on all new numeric fields; reward_hack_signals
max_length=4; empty-signals guard in the callback.
tdd-review: +13 coverage tests (dead-band hold, shim verbatim-on-error,
conservative boundary, read-only-beta dual-write, escalation postconditions,
both-restore, PID D exact, log cap + concurrency, no-top-level-import, fuzz
field-validity).
Test count 152 -> 180 (+2 POSIX-only symlink skips).
Alpamys