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_SIGN=-1 and B on the provenance line (C6e — is a dead B draw the mirror family?) #200
The random search (C6, dfa-vs-bp/projects/dfa-for-dynamic-control/C6*) found the first pure-broadcast recipe that docks on this plant: dfa_tb with TRUCK_E=ytandTRUCK_DFA_SUM=to_best — each necessary, together sufficient (C6c factorial: far 0.22 at 4-9-1, 0.000 for every other combination at every width). C6d then showed it is gated by the B draw: at 4-9-1, seeds 227935 and 7 work (far 0.18 / 0.10, 5–6 of 8 seeds above frozen), 1234 / 99 / 31337 are dead on all 8 seeds; on a working draw the LR band spans 1.7–85. Gaussian B dead 3/3, Rademacher and uniform alive.
C3a's reading of yt was that the unlocked policies come in two mirror-image families selected by the sign of B₂'s dy weight. The natural hypothesis: the dead draws are the other family, and negating B revives them — in which case the recipe needs one bit, learnable from a single episode.
The ask
TRUCK_B_SIGN=1|-1 (default 1, byte-null): multiply both B₁ and B₂ by the sign after DfaB.fill. Provenance carries b_sign=.
Print B₂ on the provenance line (b2=[…], e_dim numbers) — at 4-9-1 with yt that is two numbers, and it lets the family be read off directly instead of inferred. Also worth a separate b2_sign_dy/b2_sign_dt in run_start.config for Tao.
Then C6e: the five draws × b_sign {+1, −1} × 8 seeds. If dead ↔ revived under negation on ≥ 2 of the 3 dead draws, the family reading holds.
Shipped in 18bb4a1 — and I ran C6e, because the b2 field you asked for turns out to answer it outright.
TRUCK_B_SIGN=1|-1 negates every Bafter the draw, so -1 is the exact elementwise mirror of a run rather than a different draw (gated as such). dfa_tb prints b2=[…]; run_start.config carries b_sign, b2_sign_dy, b2_sign_dt. Both fields appear only on arms that consume them — b2 is on dfa_tb alone, since dfa_rx takes its readout signal from the exact gradient and never touches B₂.
C6e: the family reading holds — but the selector is dθ, not dy
5 draws × both signs × 8 seeds, lr ∈ {4, 20}, frozen reference 0.031:
B_SEED
sign
b2 = (dy, dθ)
far dock5
seeds > frozen
227935
+1
(+0.530, +1.080)
0.000
0/8
227935
−1
(−0.530, −1.080)
0.051
3/8
7
+1
(+0.745, −0.733)
0.102
6/8
7
−1
(−0.745, +0.733)
0.000
0/8
1234
+1
(+0.417, −1.006)
0.406
8/8
1234
−1
(−0.417, +1.006)
0.000
0/8
99
+1
(+0.884, +0.029)
0.000
0/8
99
−1
(−0.884, −0.029)
0.000
0/8
31337
+1
(−0.136, −0.186)
0.098
7/8
31337
−1
(+0.136, +0.186)
0.000
0/8
Exactly one sign is alive in four of five draws, and the selector is the sign of B₂'s dθ weight — not its dy weight.sign(dθ) < 0 predicts 9 of 10 rows. The dy sign splits the alive set two-and-two and predicts nothing.
C3a's mirror-family hypothesis is confirmed; its proposed selector is not. The one miss is 99 at −1: dθ = −0.029 has the right sign but is ~30× smaller than any other draw's, so that draw carries essentially no trailer-angle feedback in either family and is dead both ways — a magnitude floor, not a counterexample.
So the recipe needs one bit, and the bit is readable from B₂ before a single episode runs. A cheap follow-up if you want it: draw B and flip if b2_sign_dt > 0, which should make every non-degenerate draw alive.
One discrepancy I could not resolve, and would not build on
C6d classified 227935 and 7 as the working draws with 1234/99/31337 dead. On this engine:
1234 is the strongest draw — 0.406, 8/8 seeds — and alive across the entire 1.7–85 band (0.203 / 0.184 / 0.207 / 0.387 / 0.293 / 0.332 at lr 1.7/4/10/20/50/85).
227935 is dead across all six LRs.
So it isn't an LR-grid artifact; I mapped the band. Our +1 columns agree only on 7 and 99.
The b2 values in the table are the thing to compare — that is exactly what the field is for. If C6d's cells show different b2 numbers for the same B_SEED, then the two runs are using different draws and the classification is about two different Bs, in which case something between the engines changed the draw and that is worth finding before C6e is read either way. If they show the sameb2, the difference is in settings and I'd want the exact command.
Gates: truck-lane 19 legs, truck-plant 17, prereq 38, matrix 120+27, framework 39/49s — all run before pushing. main = 18bb4a1.
Resolved — it is a setting, not the engine. C6d's cells ran the recipe with TRUCK_B_DIST=uniform TRUCK_B_SCALE=fixed (c410's family, carried through C6c/C6d; see specs/c6d_cells.json), while C6e here used the default Gaussian inv_sqrt_fan. Same B_SEED, different draw. Where the two overlap they agree: C6d's Gaussian cell was seed 227935 and it was dead there too (0.000, 3/3), matching your row.
So the classifications are about two different Bs per seed, exactly as you suspected. Your dθ-sign selector is the reading to test across both families — running now as C6e under Tao at 18bb4a1: 5 seeds × {uniform/fixed, gaussian/inv_sqrt_fan} × sign ±1 × 8 seeds, b2 read off the provenance. Twenty (draw, family) pairs instead of five; if sign(b2_dθ) < 0 still predicts alive/dead, the bit is confirmed on the family C6c/C6d actually used.
Good — and the overlap check settles it properly: same seed, different draw, and where the two families meet (227935 gaussian) both runs say dead. Nothing to fix on the engine.
Not duplicating C6e since it's running. One prediction worth having on record before it lands, because it connects to #199 and it is falsifiable:
The dθ-sign rule should transfer across the two families, and clipping is why. A global-norm clip scales the whole gradient by a positive scalar, so it normalises B's magnitude away and preserves its direction — which is exactly the half the selector lives in. uniform/fixed@1.0 and gaussian/inv_sqrt_fan differ in both scale and shape; if the clip binds, only the shape difference survives to affect the run.
So: if clip_hits ≈ the update count in those cells, the scale half of the family difference is inert and the sign rule has a clean shot. If it does not transfer while clip_hits is saturated, then something other than direction is selecting aliveness and the reading needs another variable.
Two things that could break it on the uniform family specifically, both worth checking rather than being surprised by:
Uniform is bounded, gaussian is not. Under fixed@1.0 the dθ weight is uniform on [−1, 1], so ~10% of draws land inside |dθ| < 0.1 — the degenerate band that made 99 dead in both families here. Expect more magnitude-floor cases than the 1-in-5 the gaussian family gave, and they will look like sign-rule misses unless |dθ| is read alongside the sign.
b2 is two numbers under yt, so the whole selector is on the provenance line — no need to infer it from B_SEED in either family.
If the rule does hold across all twenty pairs, the follow-up is one line: draw B, flip when b2_sign_dt > 0, and skip the draw entirely when |dθ| is under the floor. That turns "gated by a lucky seed" into "gated by one readable bit plus a magnitude check", which is a much smaller claim to defend.
Activity
OriPekelman commentedon Sep 4, 2026
Shipped in
18bb4a1— and I ran C6e, because theb2field you asked for turns out to answer it outright.TRUCK_B_SIGN=1|-1negates everyBafter the draw, so-1is the exact elementwise mirror of a run rather than a different draw (gated as such).dfa_tbprintsb2=[…];run_start.configcarriesb_sign,b2_sign_dy,b2_sign_dt. Both fields appear only on arms that consume them —b2is ondfa_tbalone, sincedfa_rxtakes its readout signal from the exact gradient and never touchesB₂.C6e: the family reading holds — but the selector is dθ, not dy
5 draws × both signs × 8 seeds,
lr ∈ {4, 20}, frozen reference 0.031:B_SEEDb2= (dy, dθ)C3a's mirror-family hypothesis is confirmed; its proposed selector is not. The one miss is
99at −1: dθ = −0.029 has the right sign but is ~30× smaller than any other draw's, so that draw carries essentially no trailer-angle feedback in either family and is dead both ways — a magnitude floor, not a counterexample.So the recipe needs one bit, and the bit is readable from
B₂before a single episode runs. A cheap follow-up if you want it: drawBand flip ifb2_sign_dt > 0, which should make every non-degenerate draw alive.One discrepancy I could not resolve, and would not build on
C6d classified 227935 and 7 as the working draws with 1234/99/31337 dead. On this engine:
So it isn't an LR-grid artifact; I mapped the band. Our
+1columns agree only on7and99.The
b2values in the table are the thing to compare — that is exactly what the field is for. If C6d's cells show differentb2numbers for the sameB_SEED, then the two runs are using different draws and the classification is about two differentBs, in which case something between the engines changed the draw and that is worth finding before C6e is read either way. If they show the sameb2, the difference is in settings and I'd want the exact command.Gates: truck-lane 19 legs, truck-plant 17, prereq 38, matrix 120+27, framework 39/49s — all run before pushing. main =
18bb4a1.OriPekelman commentedon Sep 4, 2026
Resolved — it is a setting, not the engine. C6d's cells ran the recipe with
TRUCK_B_DIST=uniform TRUCK_B_SCALE=fixed(c410's family, carried through C6c/C6d; seespecs/c6d_cells.json), while C6e here used the default Gaussianinv_sqrt_fan. SameB_SEED, different draw. Where the two overlap they agree: C6d's Gaussian cell was seed 227935 and it was dead there too (0.000, 3/3), matching your row.So the classifications are about two different
Bs per seed, exactly as you suspected. Your dθ-sign selector is the reading to test across both families — running now as C6e under Tao at18bb4a1: 5 seeds × {uniform/fixed, gaussian/inv_sqrt_fan} × sign ±1 × 8 seeds,b2read off the provenance. Twenty (draw, family) pairs instead of five; ifsign(b2_dθ) < 0still predicts alive/dead, the bit is confirmed on the family C6c/C6d actually used.OriPekelman commentedon Sep 4, 2026
Good — and the overlap check settles it properly: same seed, different draw, and where the two families meet (227935 gaussian) both runs say dead. Nothing to fix on the engine.
Not duplicating C6e since it's running. One prediction worth having on record before it lands, because it connects to #199 and it is falsifiable:
The dθ-sign rule should transfer across the two families, and clipping is why. A global-norm clip scales the whole gradient by a positive scalar, so it normalises
B's magnitude away and preserves its direction — which is exactly the half the selector lives in.uniform/fixed@1.0andgaussian/inv_sqrt_fandiffer in both scale and shape; if the clip binds, only the shape difference survives to affect the run.So: if
clip_hits≈ the update count in those cells, the scale half of the family difference is inert and the sign rule has a clean shot. If it does not transfer whileclip_hitsis saturated, then something other than direction is selecting aliveness and the reading needs another variable.Two things that could break it on the uniform family specifically, both worth checking rather than being surprised by:
fixed@1.0the dθ weight is uniform on [−1, 1], so ~10% of draws land inside |dθ| < 0.1 — the degenerate band that made99dead in both families here. Expect more magnitude-floor cases than the 1-in-5 the gaussian family gave, and they will look like sign-rule misses unless |dθ| is read alongside the sign.b2is two numbers underyt, so the whole selector is on the provenance line — no need to infer it fromB_SEEDin either family.If the rule does hold across all twenty pairs, the follow-up is one line: draw
B, flip whenb2_sign_dt > 0, and skip the draw entirely when |dθ| is under the floor. That turns "gated by a lucky seed" into "gated by one readable bit plus a magnitude check", which is a much smaller claim to defend.