truck lane: TRUCK_B_COLSIGN — fix the signs of B's error columns after the draw (C6f: alive and strong on every draw) · Issue #201 · OriPekelman/toy · GitHub
8000
You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
truck lane: TRUCK_B_COLSIGN — fix the signs of B's error columns after the draw (C6f: alive and strong on every draw) #201
C6e (dfa-vs-bp/projects/dfa-for-dynamic-control/C6e-*/conclusion.md, engine 18bb4a1) on both B families, five seeds each, both signs, 8 seeds: negating B swaps alive and dead on 8 of 9 non-degenerate draws; sign(b2_dθ) < 0 predicts alive/dead on 17/18 pairs (your Gaussian-only 9/10 holds on uniform/fixed too). And a second bit: among alive draws, every strong one (far 0.10–0.34) has b2_dy > 0, every weak one (0.02–0.06) and the one miss (uniform 99, dead) has b2_dy < 0.
So the recipe's two conditions are the two signs of a proportional controller on (y, θ), and both are readable from B₂ before an episode runs. The follow-up that closes it:
The ask
TRUCK_B_COLSIGN=auto (default none, byte-null): after the draw, for each error column c of B (with yt: dy, dθ), multiply that column of bothB₁ and B₂ by −1 if needed so that b2_dθ < 0 and b2_dy > 0. Provenance: b_colsign=auto and the resulting b2. Column flips keep the hidden routing's structure (each hidden unit still sees the same random mix of the two error components up to a shared sign per component), so the claim stays "random B plus two bits".
C6f then: 5 seeds × 2 families × 8 seeds under auto, against C6e's alive sign. Pre-registered: every non-degenerate draw alive (≥ 6/8 seeds) and at the strong level (far ≥ 0.10 mean), including the C6e miss (uniform 99) and the weak ones (uni 1234, uni 31337, gau 31337, gau 227935). If dy > 0 turns the weak ones strong, the two-bit story is complete.
Also worth exposing, cheaply: TRUCK_B_COLSIGN=dt (only the dθ condition) so the two bits can be separated in one more spec.
TRUCK_B_COLSIGN=none|auto|dt. Column flips only — the gate asserts every magnitude of B₁ and B₂ is preserved, so only signs moved and the claim stays "random B plus two bits" rather than a designed B. Refused alongside TRUCK_B_SIGN=-1, which it would silently override.
C6f: both bits do exactly what C6e predicted
far dock5 over 8 seeds, best of lr ∈ {4, 20}, frozen 0.031:
family
B_SEED
none
dt
auto
gau
227935
0.000 (0/8)
0.332 (8/8)
0.332 (8/8)
gau
7
0.102 (7/8)
0.102 (7/8)
0.102 (7/8)
gau
1234
0.406 (8/8)
0.406 (8/8)
0.406 (8/8)
gau
99
0.000 (0/8)
0.000 (0/8)
0.000 (0/8)
gau
31337
0.098 (8/8)
0.098 (8/8)
0.254 (7/8)
uni
227935
0.402 (8/8)
0.402 (8/8)
0.402 (8/8)
uni
7
0.246 (8/8)
0.246 (8/8)
0.246 (8/8)
uni
1234
0.000 (0/8)
0.355 (8/8)
0.355 (8/8)
uni
99
0.016 (3/8)
0.016 (3/8)
0.070 (4/8)
uni
31337
0.074 (8/8)
0.074 (8/8)
0.324 (8/8)
The dθ bit revives dead draws to full strength — gau 227935 0.000 → 0.332, uni 1234 0.000 → 0.355. Not merely alive: as strong as the best natural draws.
The dy bit turns weak draws strong — gau 31337 0.098 → 0.254, uni 31337 0.074 → 0.324. Exactly your pre-registration.
Compliant draws are byte-null under the fix (gau 7, gau 1234, uni 227935, uni 7).
Eight of ten hit the bar; the two misses are both seed 99, for different reasons
uni 99 is not explained. Both bits satisfied, magnitudes healthy (+0.559, −0.755), and it still reaches only 0.070 (4/8). This one is a genuine hole in the two-bit story, not a floor.
A hypothesis I am deliberately not calling a finding
A dy/|dθ| ratio orders the strengths loosely — the three weakest draws all sit at ratio ≥ 0.73, and gau 99's is 30.8 — but Spearman ρ = −0.58 at n = 10 is not significant, and the lowest-ratio draw (uni 7, 0.035) is only mid-pack at 0.246.
I'm flagging it rather than reporting it because this lane already produced one monotone-looking relation I over-read — target_sat on #197, which dissolved the moment a second policy per level arrived — and this has the same shape: a plausible ordering over ten points with one policy each. If you want it settled, the cheap test is more draws at matched ratio, not more analysis of these ten.
Also worth knowing
While probing b2 per family I hit zsh not word-splitting $var, so an env $ex … silently dropped both family settings and my first probe reported gaussian numbers under a "uni" label. Caught because the b2 values were identical across families, which they cannot be. The C6f sweep itself is Python and passes env as a dict, so it was never affected — but it is worth knowing that shell-driven family probes on this lane are one quoting mistake away from looking like a null result.
Gates: truck-lane 20 legs, truck-plant 17, prereq 38, matrix 120+27, framework 39/49s — all run before pushing. main = 1ba3b15.
Activity
OriPekelman commentedon Sep 5, 2026
Shipped in
1ba3b15, and I ran C6f.TRUCK_B_COLSIGN=none|auto|dt. Column flips only — the gate asserts every magnitude ofB₁andB₂is preserved, so only signs moved and the claim stays "randomBplus two bits" rather than a designedB. Refused alongsideTRUCK_B_SIGN=-1, which it would silently override.C6f: both bits do exactly what C6e predicted
far dock5 over 8 seeds, best of
lr ∈ {4, 20}, frozen 0.031:B_SEEDnonedtauto0.000 → 0.332, uni 12340.000 → 0.355. Not merely alive: as strong as the best natural draws.0.098 → 0.254, uni 313370.074 → 0.324. Exactly your pre-registration.Eight of ten hit the bar; the two misses are both seed 99, for different reasons
|dθ| = 0.029, ~25× smaller than any other draw. A sign bit can't help a column that carries no trailer-angle feedback — dead in every family, under every fix. This is the case I flagged on truck lane: TRUCK_B_SIGN=-1 and B on the provenance line (C6e — is a dead B draw the mirror family?) #200.+0.559, −0.755), and it still reaches only 0.070 (4/8). This one is a genuine hole in the two-bit story, not a floor.A hypothesis I am deliberately not calling a finding
A
dy/|dθ|ratio orders the strengths loosely — the three weakest draws all sit at ratio ≥ 0.73, and gau 99's is 30.8 — but Spearman ρ = −0.58 at n = 10 is not significant, and the lowest-ratio draw (uni 7, 0.035) is only mid-pack at 0.246.I'm flagging it rather than reporting it because this lane already produced one monotone-looking relation I over-read —
target_saton #197, which dissolved the moment a second policy per level arrived — and this has the same shape: a plausible ordering over ten points with one policy each. If you want it settled, the cheap test is more draws at matched ratio, not more analysis of these ten.Also worth knowing
While probing
b2per family I hitzshnot word-splitting$var, so anenv $ex …silently dropped both family settings and my first probe reported gaussian numbers under a "uni" label. Caught because theb2values were identical across families, which they cannot be. The C6f sweep itself is Python and passes env as a dict, so it was never affected — but it is worth knowing that shell-driven family probes on this lane are one quoting mistake away from looking like a null result.Gates: truck-lane 20 legs, truck-plant 17, prereq 38, matrix 120+27, framework 39/49s — all run before pushing. main =
1ba3b15.