Worked cards
12,069 words7 subsections1 tablePlan version 1.1
On this page7
A worked card takes one relation the plan relies on, substitutes numbers, and says what the result means and what would overturn it. Cards fix intuition and supply hand-checkable arithmetic for the known-answer suite (Appendix M). They are not results, and none is a measurement of FrankenRobotics code.
Every card has five fields.
- Equation states the relation and, in the line under the title, the sections, constructions, gates and refusal codes it supports.
- Substituted puts numbers in. Every value is illustrative unless it is tagged FACT; a FACT is registered in Appendix H (or Appendix K for a planning pilot) and is re-read from its primary source before a gate cites it. "Exact arithmetic" marks a substitution with no free parameter.
- Intuition says what the arithmetic shows.
- Assumptions lists what the relation needs in order to hold.
- Flip names the change that would reverse the conclusion.
Every substituted number below was recomputed for this edition from the equation printed beside it (DESIGN arithmetic). Where the recomputation disagreed with a chapter, the card carries the recomputed value and §E.7 lists the difference. Cards GB-01 to GB-07 keep their version 0.2 identifiers; GB-04, GB-05 and GB-07 are corrected, and all seven now carry all five fields. Cards GB-08 to GB-41 are new and follow the order of the chapters.
E.1 Cards carried from version 0.2 (GB-01 to GB-07)
GB-01 · How wrong is a Cayley step? Supports C-05, ADR-MARROW-02 (§13.2), §16.7 and FROB-REF-212.
- Equation. A Cayley step rotates by 2·atan(ωh/2) where the exponential rotates by ωh. The relative shortfall is (ωh)²/12 to leading order.
- Substituted (illustrative). ω = 20 rad/s and h = 2 ms give ωh = 0.04, a relative error of 1.33 × 10⁻⁴ and a lag of 0.0027 rad/s. At the refusal bound ωh = 0.2 the error is 0.33%.
- Intuition. The Cayley map costs a second-order integrator's worth of phase error, paid for rational arithmetic with no transcendental function in a batched kernel. The update stays on the group, so the error is a lag in angle and not a drift off SO(3).
- Assumptions. One body-fixed axis over the step; ω constant within the step; the error is in rotation angle, not in energy or momentum.
- Flip. A requirement below 10⁻⁵ relative needs ωh < 0.011: substep, or use the exponential on a tier that has it.
GB-02 · What is a symmetry group worth to a policy? Supports C-09, §13.5, §23.13, BET-07 and KILL-EQUIV.
- Equation. Between n and m copies of the regular representation of a finite group Γ, equivariant linear maps have n·m·|Γ| free parameters, against n·m·|Γ|² unconstrained. Each trajectory also yields |Γ| trajectories under the group action.
- Substituted (illustrative). A quadruped whose closed loop has C₂ × C₂ (order 4), layers of width 256 (64 copies): 16,384 weights per layer instead of 65,536, and four trajectories from each one.
- Intuition. The group divides both the parameters to learn and the data needed to learn them by its order. For the bodies of this plan the order is between 2 and about 12, so the dividend is real and small.
- Assumptions. Hidden features carry the regular representation. The group is the observation-preserving subgroup of the robot with its sensors and task (§13.4, §20.13), never the mechanical group alone.
- Flip. A payload mounted on one side, or a forward-facing lidar, breaks the group to a subgroup and the dividend shrinks with it; at the trivial group it is nil.
GB-03 · Does the mass spiral close? Supports C-12, A-10 (§28.4), §25.21 and GATE-AUGUR-17.
- Equation. Actuator mass m_a = τ_req/ρ_τ with torque density ρ_τ; τ_req = c·m·g·L; m = m_f + n·m_a. So m = m_f/(1 − κ) with loop gain κ = n·c·g·L/ρ_τ, and the design closes only if κ < 1.
- Substituted (illustrative). n = 12, c = 0.25, L = 0.30 m, ρ_τ = 15 N·m/kg give κ = 0.59 and m = 2.43·m_f: a 6 kg structure becomes a 14.6 kg robot. With L = 0.50 m, κ = 0.98 and m = 53·m_f.
- Intuition. Kleene iteration on this loop is a geometric series with ratio κ. Leg length is not free: every kilogram of structure costs 1/(1 − κ) kilograms of robot.
- Assumptions. Required torque proportional to total mass and leg length; one torque density across the catalogue; actuators as the only mass that scales with load. A discrete catalogue makes the map a step function with the same fixed-point test.
- Flip. κ reaches 1 at L = 0.51 m. Longer legs need a denser actuator or a different load path.
GB-04 · Why are positions anchored? Supports §35.8, ADR-NUM-06, PO-NUM-01 and GATE-SPARK-05.
- Equation. Single-precision spacing is 2^(e−23) for a magnitude in [2^e, 2^(e+1)).
- Substituted (exact arithmetic for the spacings; the contact values are illustrative). Near 100 m the spacing is 7.6 µm; near 1 km it is 61 µm; inside an 8 m cell with the offset centred in [−4, 4) m it is at most 2⁻²² m, 0.24 µm. A foot carrying 75 N on a contact stiffness of 10⁵ N/m penetrates 0.75 mm, so 61 µm of position noise is 8% force noise.
- Intuition. Compliant contact turns position error into force error at the contact stiffness, so absolute single-precision coordinates fail long before a building is large.
- Assumptions. Penalty-like compliance at the stated stiffness; in the unanchored case, positions stored as
absolute
f32and gaps formed by subtracting them. - Flip. Scenes under about 30 m with soft contacts would work without anchoring. Anchoring costs one integer per axis and an exact re-anchoring step, so it is unconditional. An uncentred offset in [0, 8) is not an alternative: its worst spacing is 0.48 µm and its re-anchoring shift is inexact (§35.8).
GB-05 · What does it cost a differential drive to move sideways? Supports C-07, A-07 (§28.4) and the
diff-drive archetype of §14.10.
- Equation. Rotate by α, drive forward, rotate back, reverse. For a lateral shift d and half-track b the wheel travel is 2·α·b + d·cot(α/2). For small α this is 2·α·b + 2·d/α, least at α = √(d/b) with cost 4·√(d·b).
- Substituted (illustrative). b = 0.20 m. A shift of 0.01 m costs 0.18 m of wheel travel, 18 times the shift. A shift of 0.10 m costs 0.57 m by the small-angle form and 0.55 m by the exact one (α = 0.72 rad), 5.5 times the shift.
- Intuition. This is the ball-box estimate for growth vector (2, 3): cost scales as the square root of the displacement, so small corrections are relatively the most expensive.
- Assumptions. Ideal rolling with no lateral slip; the four-move manoeuvre, which is one admissible path and therefore an upper bound on the optimum; wheel travel as the cost.
- Flip. An omnidirectional base has growth vector (3) and pays d for d. A skid-steer base does not obey the rolling model at all, since it turns only by slipping (§18.9).
GB-06 · How many paired episodes does a comparison need? Supports C-15, §24.12 and §28.6.
- Equation. For a standardized paired effect δ, a betting e-process crosses 1/α after about 2·ln(1/α)/δ² pairs.
- Substituted (illustrative). α = 0.05: δ = 1 needs about 6 pairs, δ = 0.5 about 24, δ = 0.2 about 150.
- Intuition. Clear wins stop early. Pairing by world address raises δ by removing scene and disturbance variance, and that matters more than the choice of test.
- Assumptions. An anytime-valid study with a bound on the paired difference fixed in advance; a well-tuned bet; a controlled determinism class, since paired statistics are refused otherwise (§35.3). The count is an expectation and not a guarantee.
- Flip. Unpaired comparison with a third of the effect size needs nine times the episodes. Inside the qualification lane the level is α_i, not α, and the count grows with ln(1/α_i) (GB-29).
GB-07 · How much memory is a scene? Supports §19.15, §36.10, ADR-HAB-05 and KILL-POOL.
- Equation. Samples = volume/spacing³; bytes = 2 × samples for a 16-bit field. A narrow band of t voxels keeps a fraction min(1, (A/V)·t·s) of the volume, for surface area A and spacing s.
- Substituted (illustrative; the surface areas are assumed). A 10 × 10 × 3 m home at 2 cm: 37.5 million samples, 75 MB. A 60 × 40 × 8 m warehouse at 2 cm: 2.4 × 10⁹ samples, 4.8 GB; the 600 MB figure belongs to 4 cm (3.0 × 10⁸ samples). With 1,000 m² and 35,000 m² of surface and bands of 8 to 20 voxels, the band keeps 0.53 to 1 of the home and 0.29 to 0.73 of the warehouse at 2 cm: 40 to 75 MB and 1.4 to 3.5 GB before labels, which add about an eighth. 8,192 worlds each with its own 2 cm home would need 614 GB.
- Intuition. Halving the spacing costs eight times the memory, and a band saves least where surfaces are densest. Memory limits the number of distinct places, not the number of worlds, so scenes are shared and indexed.
- Assumptions. Dense or banded storage with no deduplication; a band width that includes brick rounding.
- Flip. A narrow band saves a factor of one to five at 2 cm, not an order of magnitude. Deduplicating bricks across repeated modules and variants could change the number of scenes per device by more than the band does. That is a HYPOTHESIS tested at GATE-HAB-13, with KILL-POOL as its fallback.
E.2 The thesis (GB-08 and GB-09)
GB-08 · What does an unresolved mission do to the ratio? Supports §2.1, §2.3 item 7, §39 and BET-13.
- Equation. RMT = E[min(T, H)] over the sealed missions of a stratum, and R = RMT_rival ÷ RMT_candidate. If the candidate leaves a fraction q unresolved at the deadline H, then RMT_candidate ≥ q·H while RMT_rival ≤ H, so R ≤ 1/q.
- Substituted (illustrative). H = 8 h. A candidate that resolves 88% of missions in a mean of 0.5 h and misses 12% has RMT = 0.88 × 0.5 + 0.12 × 8 = 1.40 h. A rival that resolves half its missions in a mean of 5 h has RMT = 6.5 h, so R = 4.6. A rival that resolves nothing has RMT = 8 h, and R = 5.7. No rival and no solve time give more than 1/0.12 = 8.33.
- Intuition. A missed mission costs the whole deadline: the 12% of misses are 0.96 h of the candidate's 1.40 h. Coverage binds before speed does. At these values three points of coverage are worth as much as halving the solve time (1.175 h against 1.18 h).
- Assumptions. One common deadline per stratum; a late refusal and a failure both count as unresolved; correctness is a gate outside the ratio, and one false certificate disqualifies the arm.
- Flip. Against a rival that resolves nothing, a ratio of ten at a mean solve time of 0.5 h needs q ≤ 4%. A stratum in which the rival also resolves most missions quickly has no ten-fold in it at any q.
GB-09 · Which single strategy can reach ten-fold? Supports §2.2, §2.3, §36.4, BET-13 and BET-17.
- Equation. (a) A step that takes a share s of training wall-clock: speed-up at most 1/(1 − s). (b) One design per agent turn of latency t, against an incumbent turn of C + t: at most (C + t)/t. (c) Compute per decision: C ÷ (c + c_T·f), with c the cheap rung per design, c_T the training rung and f the fraction of designs trained.
- Substituted. FACT (Appendix H): on one incumbent stack the nested rates are 94 k, 88 k and 82 k env-steps/s for step, step with inference, and training, so s = 82/94 ≈ 0.87 and (a) gives about 7.7 (1/(1 − 0.87); 7.8 from the unrounded share); on another, the full step runs at 347 k against 1.20 M for physics alone, s = 0.29, and (a) gives 1.4. Illustrative: with C = 300 s and t from 60 to 120 s, (b) gives 6 to 3.5; with the published 410 s, 7.8 to 4.4. With C = c_T = 410 GPU-s and c = 38 GPU-s, (c) gives 10.8 at f = 0, needs f ≤ 0.73% for ten, and reaches ten at no f once c exceeds 41 GPU-s.
- Intuition. Each tempting single lever stops short of ten: a faster step, a no-training cycle that decides one design per turn, or pruning alone. The thesis is therefore a conjunction: many designs per turn and few trainings.
- Assumptions. Fixed sample counts for (a); the same agent latency in both arms for (b); training at parity for (c). The 300 s and 38 GPU-s values are illustrative. The denominators are measured by this program at M0 and M1 before any mechanism is funded against them (§39).
- Flip. A measured phase ledger in which training is a much smaller share than published lowers every ceiling, and the mechanisms then ship as a layer (BET-14). A cheap rung at 10 GPU-s per design raises the admissible training fraction to 7.6%.
E.3 The body (GB-10 to GB-25)
GB-10 · How wrong is a solid box where a shell was meant? Supports §14.5, ADR-SKEL-03, PO-SKEL-02, GATE-SKEL-07 and FROB-REF-105.
- Equation. Mass is density times the Region's volume for
solid, and density times (volume − volume of the inward offset by t) for{shell: t}. For a box about its vertical axis the squared radius of gyration is (a² + b²)/12 when solid, and [V_o·(a_o² + b_o²) − V_i·(a_i² + b_i²)]/[12·(V_o − V_i)] for the shell. - Substituted (illustrative). A torso of 0.42 × 0.20 × 0.10 m in an aluminium alloy at 2,700 kg/m³: solid, 22.7 kg; as a 2.5 mm shell, 1.92 kg; a ratio of 11.8. Squared radius of gyration about the vertical axis: 0.0180 m² solid and 0.0236 m² shell, 31% larger. Through GB-03 at κ = 0.59 the 20.8 kg error becomes about 50 kg of robot that does not exist.
- Intuition. Reading sheet metal as a block overstates mass by an order of magnitude. Rescaling the density to repair the mass leaves the inertia wrong, because a shell keeps its mass at the walls.
- Assumptions. A uniform wall; a box, whose inward offset is a box; internal components modelled as their own bodies or declared masses.
- Flip. A wall thicker than half the smallest dimension is a solid written as a shell and is refused (FROB-REF-105). For wheels, feet and brackets the solid reading is correct, which is why the rule is "no default fill" and not "default shell".
GB-11 · When does a winding run away? Supports §15.7, C-38, ADR-SINEW-04, ADR-SINEW-08, GATE-SINEW-05 and FROB-REF-132.
- Equation. C·dT/dt = I²·R₀·[1 + α·(T − T₀)] − G·(T − T_amb). With s = I²·R₀·α: for s < G the rise settles at P_amb/(G − s) with time constant C/(G − s); for s ≥ G there is no equilibrium. The threshold is I_crit = √(G/(R₀·α)). The electrical time constant of a winding is L/R.
- Substituted (illustrative; none measured). R₀ = 0.20 Ω, α = 0.00393 K⁻¹, G = 0.80 W/K, C = 120 J/K, ambient and T₀ at 25 °C, limit 155 °C. At 20 A: 80 W cold, s = 0.314 W/K, rise 164.7 K (190 °C) where the constant-resistance model says 100 K (125 °C); time constant 247 s against 150 s; the limit is reached in 385 s. Largest continuous current 18.6 A, not 22.8 A. I_crit = 31.9 A. At 35 A the limit is reached in 61 s (83 s by the constant model). Holding torque with a flux coefficient of −0.0011 K⁻¹: the limit is reached at 15.9 A cold-equivalent and no equilibrium exists above 16.5 A. The fixture of C-38 (R₀ = 0.1 Ω, α = 0.0039 K⁻¹, G = 0.5 W/K) has I_crit = 35.8 A, and a winding of 1 Ω and 10 mH has an electrical time constant of 10 ms.
- Intuition. Resistance that rises with temperature is positive feedback. It lengthens the time constant, lowers the continuous rating by 19%, and above a finite current removes the equilibrium. Under torque control the continuous rating and the runaway threshold are four percent apart. A 20 s episode from cold warms this winding by 12.8 K and meets none of it, while the electrical state settles more than ten thousand times faster.
- Assumptions. One lumped thermal node; constant current, or constant torque in the last case; a copper coefficient linear in temperature; fixed ambient. Card values are Estimated unless calibrated.
- Flip. A housing node in series replaces G by G_wh·G_ha/(G_wh + G_ha) and lowers every threshold. A duty whose
root-mean-square current is below 18.6 A passes. Evaluation that starts at
steady(task)instead of cold puts the cliff inside the episode.
GB-12 · How much torque is left above base speed? Supports §15.8, GATE-SINEW-06 and actuator.envelope.
- Equation. Non-salient lossless machine: V_lim = V_bus/√2; base speed V_lim/√(λ² + (L·I_max)²); no-load speed V_lim/λ; above base speed, on both limits, i_d = [(V_lim/ω_e)² − λ² − (L·I_max)²]/(2·λ·L) and i_q = √(I_max² − i_d²); top speed V_lim/(λ − L·I_max) when λ > L·I_max.
- Substituted (illustrative). λ = 0.010 V·s, L = 100 µH, I_max = 40 A, V_bus = 48 V: V_lim = 33.9 V; base speed 3,151 rad/s electrical; no-load speed 3,394 rad/s; at 1.5 times base speed i_d = −32.2 A and i_q = 23.7 A, 59% of peak torque; top speed 5,657 rad/s.
- Intuition. Negative d-axis current buys speed with torque. At 1.5 times base speed a model with field weakening delivers 59% of peak torque and a model without it delivers none, because that speed is above the no-load speed.
- Assumptions. L_d = L_q; R = 0; linear magnetics; the circle inscribed in the voltage hexagon, which is conservative by at most 13.4% of voltage at a vertex; bus voltage from GB-13.
- Flip. λ ≤ L·I_max makes the top speed unbounded in this model. Resistance and saliency move every figure, which is why the referee enforces the exact hexagon and the full equations.
GB-13 · How much power can the bus deliver before it collapses? Supports §15.10, GATE-SINEW-07, FROB-REF-133 and FROB-REF-135.
- Equation. Battery: V = OCV(z) − R₀·I − V₁ − V₂ with dVₖ/dt = −Vₖ/(Rₖ·Cₖ) + I/Cₖ and dz/dt = −I/Q. A source E = OCV − V₁ − V₂ behind series resistance R delivers power P only if P ≤ E²/(4·R), and then at V = [E + √(E² − 4·R·P)]/2.
- Substituted (illustrative). E = 48 V and R = 0.08 Ω: the ceiling is 7,200 W; at 1,800 W the bus sits at 44.8 V and carries 40.2 A. With OCV = 50 V, a first pair of 0.02 Ω and 500 F (10 s) and a second of 0.03 Ω and 4,000 F (120 s), a sustained 40 A builds V₁ = 0.8 V and V₂ = 1.2 V: E falls from 50 V to 48 V and the ceiling from 7,812 W to 7,200 W. After 20 s of that current E is 49.1 V and the ceiling 7,541 W.
- Intuition. Current-controlled actuators draw power, not current, so a sagging bus draws more current and sags further. Above the ceiling there is no operating point: the step records a brownout and never completes at the requested power. The two pairs make the ceiling depend on the last seconds and the last minutes of load.
- Assumptions. Constant-power loads; a resistive source; the demand of every member summed at one instant, which is why the bus couples limbs that the tree separates.
- Flip. The checked case is end-of-discharge voltage at cold-temperature resistance (stage 6). A 20% rise in R lowers the ceiling by 17%, and an ageing state that grows R₀ can move a passing design to a refusal (§31.14).
GB-14 · How many referee pairs promote a fast-tier result? Supports §16.8, DC-01 (§28.19), C-15, GATE-FID-02 and GATE-ATEL-13.
- Equation. One bit per audited pair: did the tiers disagree beyond the frozen envelope? With rates p₀ < p₁ fixed in advance, ln E⁺ = (n − k)·ln((1 − p₀)/(1 − p₁)) + k·ln(p₀/p₁) after n pairs with k exceedances; promote when E⁺ ≥ 1/α⁺. The mirror process rejects when (p₁/p₀)ᵏ·((1 − p₁)/(1 − p₀))ⁿ⁻ᵏ ≥ 1/α⁻. By Ville's inequality both error rates hold at any stopping time.
- Substituted (illustrative). p₀ = 0.05, p₁ = 0.30, α⁺ = 0.01, α⁻ = 0.05. A clean pair multiplies E⁺ by 0.95/0.70 = 1.36 and an exceedance divides it by 6. Promotion takes 16 clean pairs, 22 with one exceedance and 29 with two. Two exceedances in the first three pairs reject (26.5 ≥ 20). With α⁺ = 0.1, or with p₁ = 0.5, eight clean pairs suffice. The fixed fallback of 30 pairs, promoting on at most two exceedances, promotes wrongly with probability 0.0021 at p₁ and promotes with probability 0.81 at p₀.
- Intuition. Clear cases promote after a handful of referee episodes, and each exceedance costs about six clean pairs. The table (n, k) → promote, reject or continue is integer data that CI regenerates and compares.
- Assumptions. The envelope is hashed before sampling; episodes are chosen by address before outcomes exist. Independence between episodes is not needed.
- Flip. A stratum whose true exceedance rate lies between p₀ and p₁ reaches the 32-pair budget and stays
provisional. An envelope tuned after outcomes voids the bound.
GB-15 · What does pairing with the fast tier save on the referee? Supports C-25, §16.8, §29.11, KILL-DR, GATE-FID-08 and GATE-CHRON-09.
- Equation. Control variate: θ̂ = Ȳ_hi(n) − β·(Ȳ_lo(n) − Ȳ_lo(N)), with variance Var(Y_hi)/n·(1 − ρ²·(1 − n/N)) at the optimal β. Doubly robust correction: Zᵢ = mᵢ + Aᵢ·(Yᵢ − mᵢ)/pᵢ with Aᵢ ~ Bernoulli(pᵢ) set by the system. Then E[Zᵢ | past] = E[Yᵢ | past] for any prediction mᵢ, and Var(Zᵢ) = Var(Yᵢ) + E[(1 − pᵢ)/pᵢ·(Yᵢ − mᵢ)²].
- Substituted (illustrative). With N ≫ n the referee episodes needed fall by 1/(1 − ρ²): 1.33 at ρ = 0.5, 1.96 at 0.7, 5.3 at 0.9 and 10.3 at 0.95. A factor of 5 needs ρ ≥ 0.894 and a factor of 10 needs ρ ≥ 0.949. With n = 64, N = 6,400 and ρ = 0.9 the variance ratio is 0.198, a factor of 5.0. For a binary outcome, a prediction wrong by ½ on every address and p = 0.1, the added variance is 9 × ¼ = 2.25, nine times the largest possible Var(Y).
- Intuition. Coupling pays only at high correlation, and a poor cheap prediction costs variance and never bias. The estimator is safe to run before ρ is known; whether it is worth running is the measurement.
- Assumptions. Audit probabilities logged before dispatch and bounded below by p_min; the prediction frozen before any high-tier outcome exists; the same addresses on both tiers.
- Flip. Contact chaos that drives ρ below 0.7 for whole episodes sends promotion back to the plain paired e-process (KILL-DR). Re-simulating short segments restores correlation and estimates a different quantity, which the report names.
GB-16 · How much of a GPU run can the referee shadow? Supports §16.11, §36.9, row 17 of §36.3, GATE-FID-07, KILL-SHADOW and BET-34.
- Equation. A device produces T·h world-seconds per second at T world-steps/s and step h. A referee on c cores at real-time factor ρ_rt per core covers c·ρ_rt of them: share = c·ρ_rt/(T·h). Detecting a gap δ in a success rate near p takes about 16·p·(1 − p)/δ² episodes per tier.
- Substituted (illustrative, on the TARGETs of §36.3). Humanoid at 1.7 M world-steps/s and h = 2 ms: 3,400 world-seconds per second. Twenty-four cores at 0.2 to 1.0 of real time cover 4.8 to 24, a share of 0.14 to 0.71%. Quadruped at 2.0 M: 4,000, a share of 0.12 to 0.6%. At the ambitions of 2.55 M and 4.0 M the shares are 0.09 to 0.47% and 0.06 to 0.3%. A five-point gap needs 576 episodes at p = 0.9 and 1,600 at p = 0.5; at 20 s per episode the first takes 8 minutes at the fastest referee rate and the second 1.9 hours at the slowest. Snapshot traffic for 256 humanoid worlds at 50 Hz is 6.9 MB/s.
- Intuition. The referee is a sampler by arithmetic, and a faster engine makes the shadowed share smaller. One-step defects from a wrong kernel show in seconds; gaps in outcome take minutes to hours.
- Assumptions. Referee speed per core is unmeasured (the range is row 16's TARGET); sampled worlds are chosen by address with a known probability; trajectories are not compared beyond a few steps.
- Flip. A
batch-cputier at 275 k humanoid world-steps/s covers 550 world-seconds per second, 16% of 3,400, and becomes the tier that shadows a tenth or more. That comparison checks kernels and arithmetic, not the contact law.
GB-17 · When does a size bucket pay for itself? Supports §17.5, ADR-PHYLUM-02, KILL-BUCKET, GATE-PHYLUM-03 and FROB-REF-741.
- Equation. With per-world work scaling as Lᵖ, a cohort padded to L_max costs Ω_p = L_maxᵖ/E[Lᵖ] times the unpadded work, and 1/r ≤ Ω_p ≤ r⁻ᵖ for r = E[L]/L_max. Splitting into B buckets wins when N·τ·(Ω − Ω_b) > (B − 1)·c_L, with τ the unpadded cost of a world-step and c_L the launch overhead of one plan.
- Substituted (illustrative). 90% quadrupeds at 18 degrees of freedom and 10% humanoids at 35: r = 0.56, Ω₁ = 1.78, Ω₂ = 2.96, Ω₃ = 4.50. With stage shares of one half cubic, three tenths quadratic and one fifth linear, Ω = 3.5. At τ = 0.5 µs and c_L = 0.45 ms, 8,192 worlds cost 14.8 ms in one bucket and 5.0 ms in two; 512 worlds cost 1.35 ms and 1.16 ms; 128 worlds cost 0.67 ms and 0.96 ms. The crossover is 360 worlds. One 66-degree-of-freedom design in a hundred raises Ω₃ to 23.3.
- Intuition. The largest robot, not the mean, sets the cost of a padded batch, and each bucket buys back padding at the price of its own launches. Buckets win at thousands of worlds and lose at a hundred.
- Assumptions. The stage shares and c_L are HYPOTHESES that the M2 profile replaces; padded slots do full work; the two-bucket case has no padding left.
- Flip. A launch overhead ten times smaller moves the crossover to 36 worlds. A tree-sparse factor that skips masked slots removes most of the cubic term and makes one bucket viable at larger N.
GB-18 · What is the most a family batch can gain? Supports §17.6, §36.5, §2.3 item 5, §25.12, KILL-FAMILYLANE, GATE-AUGUR-14 and FROB-REF-866.
- Equation. If a fraction f of the strong baseline's matched wall time is improved by a factor r and the method adds relative overhead h, then speed-up ≤ 1/(1 − f + f/r + h).
- Substituted (illustrative unless tagged). f = 0.85 with r unbounded gives 6.7; f = 0.97, r = 40 and h = 0.02 give 13.5; with h = 0.02 no r reaches ten unless f ≥ 0.92. A generation of 256 designs × 16 worlds × 10 s is 20.5 M world-steps, about 10 s at 2.0 M world-steps/s, and padding adds 3 s. If the comparator pays 1 s per design switch, f = 0.96 and the bound is 266/13 = 20; at 0.1 s per switch, f = 0.72 and the bound is 2.7. For a family lane on the constraint stages, where FACT (Appendix H) assembly is 38 of 262 ns of an incumbent humanoid step, f from 0.15 to 0.35 with r = 4 and h = 0.02 gives 1.10 to 1.32. FACT (Appendix K): a specialized solver measured 1.345 warm and 1.088 cold, and its batch held a residual of 575.7 µs that the solver does not touch (parsing, assembly, checking, serialization and harness, not separately timed), so a solver of zero cost would have reached about 1.46 (842.8 ÷ 575.7).
- Intuition. The bound is computed before the experiment is funded. A mixed batch is worth what the rival pays to switch designs, which is unpublished; lanes inside the step are a modest optimization.
- Assumptions. A serial cost model on the strong baseline's critical path, which also caches graphs, factors and kernels; f, r and h measured, not assumed. A pruning ratio is never multiplied by a speed-up measured on work already pruned: 100 candidates of 10 s, 90 pruned, survivors twice as fast, is 1,000/50 = 20 and not 50.
- Flip. A measured switch cost near zero in the incumbent closes the family-batch gap. Overlapped queues void the serial model, and the critical path is then measured directly.
GB-19 · How many distinct designs does a symmetric sweep contain? Supports C-22, §17.11, §28.15, BET-31, KILL-QUOTIENT and GATE-PHYLUM-01.
- Equation. Orbit counting: the number of distinct assignments of k module types to slots X under a group Γ is |Γ|⁻¹·Σ k^c(g) over g in Γ, with c(g) the number of cycles of g on X.
- Substituted (exact arithmetic). Ten leg types on the six slots of a hexapod with dihedral symmetry of order 12: (10⁶ + 2·10 + 2·10² + 10³ + 3·10⁴ + 3·10³)/12 = 86,185, a factor of 11.6 below 10⁶. If a sensor leaves only the sagittal mirror: (10⁶ + 10³)/2 = 500,500, a factor of 2.0. Three types under the full group: 92. Four feet with states open, stick and slip under C₂ × C₂: (3⁴ + 3·3²)/4 = 27 strata. Eight equal ballast slugs in 24 collinear slots, which rigid-body dynamics reads only through their count, Σx and Σx²: C(24, 8) = 735,471 placements in 24,546 classes.
- Intuition. Algebra spent at campaign time deletes simulations exactly and before any run. The factor is at most the group order, and one asymmetric sensor takes it from twelve to two.
- Assumptions. The acting group is the observation-preserving subgroup of chassis and task; results are transported back with the recorded group element; the projection quotient is a structural rule checked by bit-identical replay.
- Flip. A front-only lidar counted with the mechanical group merges designs that the closed loop distinguishes. Slots {1, 4} and {2, 3} share Σx = 5 and differ in Σx² (17 and 13), so a projection that drops Σx² merges them wrongly.
GB-20 · How much friction is left for sliding when a foot also spins? Supports §18.6, C-37, T-07, GATE-TRAC-04 and OPEN-TRAC-2.
- Equation. √(λ_t1² + λ_t2² + (λ_s/r_eff)²) ≤ μ·λₙ. Under maximum dissipation with sliding speed v and spin ω, the tangential force is μ·N·v/√(v² + (r_eff·ω)²) and the spin moment μ·N·r_eff·(r_eff·ω)/√(v² + (r_eff·ω)²). r_eff = 2a/3 for uniform pressure on a disc of radius a and 3π·a/16 for a Hertzian pressure.
- Substituted (illustrative). μ = 0.6, N = 100 N, a = 30 mm, uniform pressure: r_eff = 20 mm; pure sliding 60 N; pure spin 1.2 N·m. At v = r_eff·ω: 42.4 N and 0.85 N·m, each 1/√2 of its limit. Independent rows would grant 60 N and 1.2 N·m at once, 41% more in magnitude than the surface supplies. A Hertzian pressure gives r_eff = 17.7 mm.
- Intuition. Sliding and spin share one budget. A pivoting foot resists sliding less, and a model with a separate torsional row holds where the foot should slip.
- Assumptions. An ellipsoidal approximation to the limit surface of a finite patch; r_eff from the sample geometry or the interface card, Estimated unless calibrated; rolling resistance as a separate row.
- Flip. A point contact (r_eff = 0) masks the spin row and returns the three-dimensional cone. A pyramid cone of the same μ grants √2·μ along its diagonals, the same 41% in another direction, which the model-form check plants as an exploit (§16.9).
GB-21 · How long does a foot ride a film before it grips? Supports §18.8, §19.10, ADR-TRAC-06, KILL-FILM, GATE-TRAC-05, GATE-HAB-06 and FROB-REF-242.
- Equation. Stefan's law: dh/dt = −2·N·h³/(3π·η·a⁴), so 1/h² − 1/h₀² = 4·N·t/(3π·η·a⁴). From a thick film the time to thickness h is t = 3π·η·a⁴/(4·N·h²).
- Substituted (illustrative; the viscosities are orders of magnitude). To 10 µm: water (1 mPa·s) under a patch of radius 2 cm carrying 100 N, 0.038 s; cooking oil (60 mPa·s) under the same patch, 2.3 s; water under a patch of radius 6 cm carrying 400 N, 0.76 s.
- Intuition. The time scales with a⁴ and with viscosity. Against a stance of 0.3 s a small foot on water reaches boundary friction almost at once, while oil and large flat feet ride a film for the whole stance. A map of constant coefficients cannot represent that difference.
- Assumptions. Rigid parallel circular plates; a Newtonian film; constant load; the combined roughness sets where the film law hands over to boundary friction. Outputs are Estimated.
- Flip. A textured or drained sole, with a much smaller effective radius, grips at once and the law changes nothing. If it changes fewer than 10% of episode outcomes on the wet-floor battery, KILL-FILM returns the lockstep tiers to static condition cards.
GB-22 · Why is half precision rejected for contact solves? Supports §18.12, §21.20, H-15 and FROB-REF-245.
- Equation. Refinement on a low-precision factor contracts only if κ·u_low < 1. A single-precision Cholesky costs n³/3 flops and each refinement step about 3n², so k steps cost 9k/n times the factor they replace.
- Substituted (illustrative, with one FACT). u₁₆ = 2⁻¹¹ = 4.9 × 10⁻⁴. A rank-one contact operator with proximal shift ρ = 10⁻²·tr(W)/48 has κ = 4,801 and κ·u₁₆ = 2.34: no contraction. The cost ratio 9k/n is 2.3 at n = 27 with seven steps, 1.5 at n = 35 with six and 3.0 at n = 18 with six; with six steps it falls below a quarter only above 216 unknowns. FACT (emulated on three reference models, Appendix H): after unit-diagonal scaling, κ·u₁₆ ≥ 1 in 37 to 64% of sampled contact states.
- Intuition. The matrices are too ill-conditioned for half precision and too small for refinement to pay. The repair costs more than the factorization it was meant to cheapen.
- Assumptions. Dense per-world systems of 18 to 64 unknowns; a single-precision baseline, against which the theory promises at most a factor of two.
- Flip. Systems above about 200 unknowns with κ below 10³ would change both terms, and none is planned for the dense per-world solve. Matrix accelerators stay available to policy networks (§23).
GB-23 · When does a 30 Hz policy tick against 400 Hz physics? Supports §11.7, §20.4, GATE-SENS-04 and GATE-SENS-05.
- Equation. The base unit is 1/Q s, with Q the least common multiple of the denominators of all periods.
Process i has the integer period Nᵢ = Tᵢ·Q and fires when n ≡ φᵢ (mod Nᵢ). Under
hold, a firing strictly inside a physics step is realized at the next step boundary. - Substituted (exact arithmetic). Physics at 400 Hz, an IMU at 100 Hz and a policy at 30 Hz: Q = 1,200 and periods of 3, 12 and 40 quanta. Policy instants 0, 40, 80 and 120 are realized at steps 0, 14, 27 and 40 (quanta 0, 42, 81, 120) with lateness 0, 2, 1 and 0 quanta, that is 0, 1.67, 0.83 and 0 ms. The pattern repeats every 0.1 s with intervals of 14, 13 and 13 steps, and a 20 s episode holds exactly 600 decision instants. The shortcut "every thirteenth step" runs at 400/13 = 30.77 Hz, a rate 2.6% high, and puts 616 decision instants in the same episode.
- Intuition. Rates that do not divide each other are ordinary. Integer phase accumulators keep the long-run rate exact and make the jitter a declared pattern that replays.
- Assumptions. Rational periods; the profile's rule for events inside a step (
holdon throughput tiers, a substep boundary on the referee and the strict profile); accumulators and queues inside every checkpoint. - Flip. A base unit finer than the declared limit is refused at admission. A pair of rates that divide, such as 400 Hz and 50 Hz, has no lateness and needs none of this.
GB-24 · How bad may the sensor be when the first move is a probe? Supports §20.16, §25.15, C-19, GATE-SENS-13 and GATE-AUGUR-03.
- Equation. Three ticks of 1 s from rest; acceleration g·u_k with a hidden gain g ∈ {1, 2} and |u_k| ≤ U for U in [½, 1]; target (x₃, v₃) = (1, 0) exactly; one reading of v₁ with error in [−ε, ε], delivered before the second command. With p = u₀ the remaining commands are u₁ = 1/g − 2p and u₂ = p − 1/g, the limits leave 1 − U ≤ p ≤ (U + ½)/2, and the two worlds are told apart when ε < p/2. Hence ε < (U + ½)/4, equality excluded.
- Substituted (exact arithmetic). U = ¾: the bound is 5/16, reached with the probe p = 5/8. The timid probe p = ¼ gives commands (¼, ½, −¾) for g = 1 and (¼, 0, −¼) for g = 2 and tolerates only ε < 1/8, so the bold probe buys a sensor two and a half times worse. On the grid U = ½ + i/32, ε = j/64 a cell is admitted when 2j < 32 + i: 344 of 561 admitted, 217 excluded, 9 exactly on the boundary. If the packet arrives after the third command is committed, all 561 are excluded. Admitting a third gain of 33/32 at U = ¾ and ε = 1/64 leaves a worst error of 1/65 m, so every cell admitted for two gains is wrong for three.
- Intuition. Information has a clock. The first command is an experiment, sensing tolerance trades against actuator authority, and a late packet is the same as no sensor.
- Assumptions. An exact terminal condition; a strict inequality evaluated in exact integer arithmetic; the hypothesis set, the delivery deadline and the error units as declared.
- Flip. Any change to the declared support of g, to the deadline or to the tolerance invalidates the verdict. FACT (executed boundary screen, Appendix H): on 65,536 cases unguarded single precision gave 452 false admissions and 23,438 false exclusions, and a guard band with exact fallback gave none.
GB-25 · Why is regrouping a sum a different kernel? Supports DC-11 (§28.19), §21, C-14 and PO-KIN-03.
- Equation. Class E rewrites are bit-exact under round-to-nearest with contraction off: commutation, sharing a
common subexpression, removing dead code, scaling by a power of two away from overflow and underflow. Class R
rewrites are equal over the reals only: reassociation, distribution, fused multiply-add. The spacing of
f32near x is 2^(⌊log₂ x⌋ − 23). - Substituted (exact arithmetic). In
f32with a = 10⁸, b = −10⁸ and c = 1: (a + b) + c = 1 and a + (b + c) = 0, because the spacing near 10⁸ is 8. Sharing a subexpression that is used three times saves two evaluations and changes no bit. - Intuition. A rewrite that changes one bit changes every digest downstream, and replay, both ledgers and every pairing go with it. Regrouping is admitted only as a new kernel, with a new identity and its own evidence.
- Assumptions. IEEE 754 single precision; no contraction; no fast-math (§35.6).
- Flip. A profile with more than about a fifth of step time in derived arithmetic would make class R worth its certificates. FACT (Appendix H): the Lie-algebraic recursions are 16 to 20% of an incumbent step, which caps specialization near 1.2.
E.4 The mind (GB-26 to GB-36)
GB-26 · Can the stop absorb everything that can reach it? Supports §22.11, C-31, the .fenergy format,
PO-SHIELD-01, GATE-SHIELD-01 and KILL-ENERGY.
- Equation. ½·m_cap·v² + ∫ max(P_elec, 0) dt + E_stored + E_gravity + E_other + E_reserve ≤ W_abs,min. Read as an exclusion: a crossing that needs a potential rise above the initial energy plus all allowed external work has no controller.
- Substituted (illustrative). A cart of 8 kg effective mass at 0.4 m/s: 0.64 J. A 20 W drive over a 25 ms cutoff: 0.50 J. An 8% slope over a 40 mm stroke: 8 × 9.81 × 0.040 × 0.08 = 0.25 J. A series spring: 0.10 J. The total is 1.49 J against an absorber of 50 N over 40 mm, 2.00 J. The longest admissible cutoff delay is (2.00 − 0.99)/20 = 50.5 ms, or 45.5 ms with a further 0.10 J of reserve. Exclusion: lifting 8 kg over a 12 mm sill needs 8 × 9.81 × 0.012 = 0.94176 J, above a 0.90 J budget for every wheel radius; 0.05 J of approach kinetic energy removes the exclusion.
- Intuition. No term depends on the policy or on the contact law, so the evidence survives every retraining and the monitor is one integrator and one comparison per tick. The margin is thin: a quarter of the absorber.
- Assumptions. A physical guide puts the absorber on every forward path; the port graph lists every source;
m_cap includes reflected rotor inertia; the absorber envelope and drive power are measurements, so the claim is
Estimated in simulation and Reality is
no-claimuntil bench data arrive (§30). - Flip. A rotor of 5 × 10⁻⁵ kg·m² behind a 50:1 reduction on a wheel of radius 0.05 m adds 50 kg of effective mass, six times the cart, and the budget fails. A budget that omits the gravity or spring term is rejected by the checker.
GB-27 · What does an impact do to a small perturbation? Supports C-16, §22.9, GATE-GAIT-01, GATE-GAIT-02 and FROB-REF-263.
- Equation. Ξ = D_xR + (f⁺ − D_xR·f⁻ − D_tR)·D_xg/(D_tg + D_xg·f⁻). A periodic gait is a fixed point of the return map P, and its Floquet multipliers are the eigenvalues of DP. For a rimless wheel with spoke half-angle α on a slope γ, ω₊² ↦ cos²(2α)·(ω₊² + (4g/ℓ)·sin α·sin γ).
- Substituted (illustrative). Bouncing ball with state (y, v), restitution r = 0.8 and impact speed v⁻ = −2 m/s: Ξ = [[−0.8, 0], [8.83 s⁻¹, −0.8]], with determinant r² = 0.64. The reset Jacobian diag(1, −0.8) has the wrong sign in its first entry, no coupling term, and determinant −0.8. Rimless wheel with eight spokes (α = π/8): multiplier cos²(π/4) = ½; with ℓ = 0.5 m and γ = 4° the fixed point is ω₊² = 2.09 s⁻², ω₊ = 1.45 rad/s, and a perturbation of ω₊² halves at every step.
- Intuition. The event time moves with the perturbation, and that shift carries the flow before the impact into the state after it. Leaving it out gets the sign and the area change wrong. Stability of a gait is then a number and not a judgment of a rollout.
- Assumptions. Isolated events with transverse crossing; rigid impacts on the referee, while compliant tiers use products of step Jacobians and no Ξ; multipliers probe local stability only.
- Flip. A grazing event, or two events not tied by a symmetry inside the event-time resolution, returns
refusedand no matrix. Six spokes give a multiplier of ¼ and twelve give ¾: more spokes, slower convergence.
GB-28 · After how many designs does a family policy repay its pretraining? Supports §23.11, C-23, BET-15, KILL-FAMILY and GATE-CORTEX-11.
- Equation. With pretraining cost P, specialist cost s per design and fine-tuning fraction f, the family policy pays after N > P/(s·(1 − f)) designs. Its gain on the training phase is at most 1/f.
- Substituted (illustrative, on FACT denominators from Appendix H). P = 60 GPU-hours and f = 0.1: about 590 quadruped designs at s = 410 s, and 67 to 133 humanoid designs at s = 1 to 0.5 GPU-hours. At f = 0.2 the quadruped figure is 660.
- Intuition. A family policy is an investment that only a campaign amortizes. One design, or ten, should train a specialist or inherit from a parent.
- Assumptions. Scoring with no adaptation costs nothing beside training; fine-tuning reaches specialist return at fraction f; pretraining is paid once per morphology class.
- Flip. Return with no adaptation below 70% of specialists on more than half the held-out set, together with fine-tuning above 20% of steps, fires KILL-FAMILY and returns the plan to per-design training with inheritance for parameter edits. A campaign of fewer than about 600 quadruped designs never reaches break-even at these values.
GB-29 · What does a retry cost in the qualification lane? Supports §24.13, PO-ARENA-02, INV-QUAL-03 and GATE-QUAL-01.
- Equation. Request i of a campaign is tested at α_i = α/[i·(i + 1)]. Because 1/[i·(i + 1)] = 1/i − 1/(i + 1), the levels sum to α·m/(m + 1) after m requests and to α in the limit. Sample size grows with ln(1/α_i).
- Substituted (exact arithmetic). α = 0.05. Without the lane, fourteen independent null submissions at 5% give a 51% chance of at least one false promotion. With it, request 7 runs at 0.05/56 and the campaign has then spent 0.05 × 7/8. Relative to one test at 0.05, request 1 needs 1.23 times the units, request 10 needs 2.57, request 100 needs 4.08 and request 500 needs 5.15.
- Intuition. Error control holds for any number of adaptive retries, because the level is spent whether or not the agent likes the answer and the worlds are fresh each time. Retrying is allowed and is never free.
- Assumptions. The request is committed before the draw; no cluster serves two requests; the index advances for every committed request; all agents on a claim family share one index sequence.
- Flip. An agent that could open a fresh campaign would reset its index and void the bound, which is why
opening one needs
qual:admin. Reading confirmation worlds does not break the error rate; it breaks the held-out population, and the family is retired.
GB-30 · Which verdicts can 64 pairs return? Supports §24.10, §24.12, §28.6, GATE-ARENA-05, GATE-ATEL-06 and FROB-REF-532.
- Equation. Two-sided 97.5% exact Clopper–Pearson intervals [L_b, U_b] and [L_a, U_a] on the two discordance probabilities; [L_b − U_a, U_b − L_a] then covers their difference with probability at least 95%. Prefer B if the lower end exceeds +0.05; a practical tie if the interval lies inside ±0.05. With no discordant pair the interval is ±(1 − 0.0125^(1/n)).
- Substituted (exact enumeration). At n = 64 there are 2,145 count pairs: 521 prefer B, 521 call B inferior, 1,103 are inconclusive and none is a tie, because perfect concordance gives ±0.0662. A tie first becomes reachable at n = 86 (±0.0497) and only for the pair (0, 0), whose chance at 5% discordance each way is 0.9⁸⁶ = 0.012%. With no A-only pair a preference needs 15 B-only pairs of 64. At the level of qualification request 1 a tie needs n = 99, and at request 10 it needs n = 178. Peeking at an exact one-sided sign test after every pair turns a nominal 5% into 13/256 = 5.08% after 8 looks and 17.4% after 64.
- Intuition. A frozen rule has a finite set of outcomes, and a wanted verdict may not be among them. Enumerating the rule before sampling costs nothing and prevents a study that cannot return the answer it was commissioned for.
- Assumptions. Complete pairs; a rule frozen before the first episode; no stop on an interim look.
- Flip. A claim of equivalence may instead register the one-sided bound on total discordance: no discordant pair in 64 gives 1 − 0.05^(1/64) = 0.0457, inside the five-point band. That is a different registered inference and not a rescue of the first.
GB-31 · How narrow can the stance be for the whole payload envelope? Supports §25.13, C-32, C-17,
GATE-AUGUR-02, GATE-AUGUR-05 and the lp-farkas proof kind of §12.4.
- Equation. Four point feet at (±a, ±b) carry a load W whose centre of mass is at (x, y). Moment balance gives f₊₊ − f₋₋ = W·x/(2a) + W·y/(2b). With 0 ≤ fᵢ ≤ c, the corner of the envelope is feasible exactly when that difference is at most c.
- Substituted (exact arithmetic). W = 400 N, a = 1/3 m, envelope half-sizes 150 mm and 75 mm, c = 160 N. At the corner f₊₊ − f₋₋ = 90 + 15/b, so b ≥ 3/14 m, with forces (160, 130, 110, 0) N in the order [++, +−, −+, −−]. Dual witness at b = 1/5 m: the same combination of the moment equations demands 165 N from a 160 N rating. At b = ¼ m active allocation (155, 135, 105, 5) N passes, and four equal passive springs give (175, 115, 85, 25) N, 15 N over the rating. On the validation grid of 2,304 bodies and 11,520 body–load states there are 2,254 feasible states, 114 bodies that cover the mission, and 1,596 bodies that a support-polygon test with a total-capacity check accepts and the exact model rejects.
- Intuition. The answer is a set of designs with a boundary, witnesses inside it and a dual vector outside it, checked by a few exact products. A support polygon is not that answer, and neither is a result that omits who allocates the forces.
- Assumptions. Fixed pose; coplanar contacts; a vertical load; independently commanded normal forces; mission vertices that do not depend on the design.
- Flip. A passive suspension, a mass that depends on the design (the midpoint of two feasible designs can be infeasible, §25.4.7) or a force floor for continuous contact each change the set. A lifted linear program reaches the same optimum in one solve, so no speed is claimed for a specialized solver (ADR-AUGUR-03).
GB-32 · What does it cost to certify a rare failure rate? Supports §26.9, C-27, DC-04, INV-NEMESIS-03, GATE-NEMESIS-04 and KILL-RARE.
- Equation. n nominal episodes with no failure give p ≤ 1 − δ^(1/n) at confidence 1 − δ, so n = ⌈ln δ/ln(1 − p⋆)⌉, about ln(1/δ)/p⋆. With a defensive mixture of nominal share β the weights are at most 1/β, and no observed failure certifies only p ≤ ln(1/δ)/(β·n). Splitting with L levels of conditional probability p₀ and N trials per level has a squared coefficient of variation near L·(1 − p₀)/(p₀·N).
- Substituted (exact counts; the throughput is illustrative). δ = 0.05: 29,956 episodes for 10⁻⁴, 299,572 for 10⁻⁵ and 2,995,731 for 10⁻⁶; at 400 episodes per second, 75 s, 12.5 minutes and 2.1 hours. At the level of the tenth qualification request, 0.05/110, the first count is 76,959. With β = 0.1 the same 29,956 clean episodes certify only 10⁻³. Splitting to 10⁻⁶ with six levels of 0.1 and 10⁴ continuations each uses 6 × 10⁴ fragments, fifty times fewer, at a coefficient of variation of at least 7%.
- Intuition. Finding failures and certifying their absence are different purchases. Exploration is cheap and has no bound on its error; absence has a fixed price that no sampler discounts without assuming where the failures live.
- Assumptions. Independent nominal episodes; a failure predicate frozen before sampling; a bound scoped to its tier and randomization, of class Tested.
- Flip. At a tenth of the throughput the 10⁻⁶ claim costs 21 hours: certify 10⁻⁵ and report 10⁻⁶ as Estimated. Failures driven by static parameters alone gain nothing from forks, since a fork without fresh randomness is a copy.
GB-33 · How good is a greedy battery of scenes? Supports §26.11, DC-10, PO-NEMESIS-03 and GATE-NEMESIS-08.
- Equation. f(S) is the weight of failure modes exposed by at least one scene of S; it is monotone and submodular. For a greedy selection S of k scenes, OPT_k ≤ f(S) + the sum of the k largest remaining marginal gains, and f(S) ≥ (1 − 1/e)·OPT_k.
- Substituted (illustrative). Forty modes of total weight 100. Greedy gains of 31, 22, 14, 9, 6 and 4 cover 86. The six largest remaining gains sum to 12, so OPT₆ ≤ 98 and the battery is within 88% of the best six-scene battery, against a worst case of 63%.
- Intuition. The certificate of the instance is usually far tighter than the constant, and it is computed from numbers the greedy trace already holds.
- Assumptions. Mode weights declared in advance; an incidence matrix built under the identities in force; equal scene costs, since unequal costs use the cost-benefit variant and its weaker constant.
- Flip. A change of model class, tier, kernel version or taxonomy voids the matrix, and the standard battery of §38 returns. A battery of past failures alone overfits, so one slot in five is drawn from held-out scenes, and a rare-failure bound refuses battery-selected episodes.
GB-34 · Did the intervention rescue the episode, or did the fork? Supports §27.5 to §27.7, DC-06, C-29, INV-AUTOPSY-05, GATE-AUTOPSY-03 and KILL-AUTOPSY.
- Equation. Rescue rate: the fraction of replicates of do(c) in which the requirement is not violated. Contrast: the rescue rate minus the placebo flip rate ρ₀. A factor is named only when its contrast is tested above zero against the placebo forks, at a level divided by the number of factors. A ρ₀ above 0.2 names nothing.
- Substituted (illustrative). Of 200 failing worlds re-run from their forks, removing the torque limit
prevents the failure in 164, nominal friction in 38 and a placebo in 12: rescue rates 0.82, 0.19 and 0.06, and
contrasts 0.76 and 0.13. The first is
but-for(a contrast of at least 0.5) and the secondcontributing. A budget of 12 factors, singles and pairs, is 78 coalitions; at 8 replicates each, with 64 placebo forks and the null fork, it is 689 forks. - Intuition. In a chaotic system any touch changes some outcomes. The placebo measures that floor, and a cause is whatever beats it.
- Assumptions. The null fork replays the failure bit for bit; each intervention changes one declared mechanism; forks receive fresh keyed randomness; attribution runs in double precision under class D2.
- Flip. A placebo rate near a factor's rescue rate, 30 of 200 against 38, leaves nothing to attribute. If only
a pair rescues, the cause is
joint, and a Shapley summary of (½, ½) cannot tell that case from two alternative causes.
GB-35 · How wide is the survivor margin, and how far can its coverage move? Supports §28.9, §28.10, C-26, DC-02, DC-03, DC-08, §30.12, GATE-ATEL-08, GATE-GROUND-06 and KILL-LADDER.
- Equation. On n audit designs with residuals rᵢ = |y_hi − ĝ(y_lo)|, the margin q̂ is the ⌈(n + 1)·(1 − α)⌉-th smallest residual, and every design within 2·q̂ of the k-th best lower-rung score advances. Conditional on the calibration set, coverage is distributed Beta(n + 1 − l, l) with l = ⌊(n + 1)·α⌋. A finite margin needs n ≥ (1 − α)/α.
- Substituted (illustrative). n = 39 and α = 0.05: q̂ is the 38th smallest residual, say 0.04, and a third-best score of 0.74 puts the cut at 0.66. Coverage is Beta(38, 2): mean 0.95, below 0.884 in one calibration draw in twenty. Fewer than 19 audit designs cannot support α = 0.05; the floor of 19 battery points in DC-08 is the same arithmetic, since (n·R̂ + 1)/(n + 1) ≤ 0.05 at R̂ = 0 needs n ≥ 19. Ladder cost for 81 designs with rungs of 10 s, 3 min and 30 min: 40.5 h if all are trained; 6.1 h by halving at η = 3; 2.0 h if six survive the cut, are scored at the second rung, and three of them are trained. That is before the audit lane, which adds 57 minutes when its 19 designs are scored at the second rung inside the campaign. A reality band at n = 20 and 90% has coverage Beta(19, 2): mean 0.905, standard deviation 0.063, below 0.80 with probability 0.069.
- Intuition. A rung cannot separate designs closer than twice its own error, so the size of q̂ decides whether the ladder saves a factor of three over halving or nothing. Coverage holds on average over calibration draws and moves by several points at small n.
- Assumptions. Audit designs drawn by key before any score exists and never cancelled; exchangeability with the designs being scored; calibration never taken from survivors.
- Flip. q̂ = 0.10 lets about twenty designs through, more than 3k = 9, and the layer gives back most of its advantage. Calibrating on promoted designs under-covers wherever the residual correlates with the cheap score.
GB-36 · Which experiment is worth buying next? Supports DC-05, DC-07, C-39, §28.16, §30.15, KILL-VOI and GATE-ATEL-15.
- Equation. For a posterior difference Δ ~ N(m, s²) between two finalists and an experiment that removes variance s̃², EVSI = s̃·[φ(z) − z·Φ(−z)] with z = |m|/s̃. A measurement of variance v on a component of prior variance w removes w²/(w + v).
- Substituted (illustrative). m = 0.03; sampling variance 0.03² and tier-bias variance 0.04²; noise per pair 0.3 on the fast tier and 0.1 on the referee. Two hundred more fast-tier pairs remove s̃ = 0.0245 and are worth 0.0013 for 1 GPU-second. Eight referee pairs remove s̃ = 0.0408 and are worth 0.0055 for 27 GPU-second equivalents, so the fast tier wins sixfold per unit cost. A second fast batch is worth 10⁻⁵, so the referee goes next. At m ≥ 0.15 even perfect information is worth under 2 × 10⁻⁵: stop and decide. With a reality-gap term of 0.10 (an uncalibrated cell), twenty real pairs at noise 0.3 remove s̃ = 0.096 and are worth 0.0251, against 0.0084 for all remaining simulation; with 0.02 (a calibrated cell) the same batch is worth 0.0035.
- Intuition. Cheap episodes are bought until only tier bias is left, which only the referee shrinks, and real episodes are worth buying where the reality gap is the largest unknown. Calibration does not make simulation faster; it changes which experiment is worth buying.
- Assumptions. A Gaussian posterior whose variance components are read from the ledgers; a declared exchange rate between core-seconds, GPU-seconds and real episodes. The model ranks experiments and never issues a verdict.
- Flip. A wrong variance model costs efficiency and never validity. If stated 80% intervals cover fewer than 65% or more than 92% of thirty registered experiments, the core falls back to the fixed order of check, simulate, train, referee.
E.5 The record and the world (GB-37 to GB-39)
GB-37 · Which cycles and which tolerances carry the cost? Supports §31.8, §31.11, C-35, GATE-FOUNDRY-03, GATE-FOUNDRY-08, KILL-TOLERANCE, FROB-REF-163 and FROB-REF-165.
- Equation. Linear damage D = Σ nᵢ/N(Sᵢ), with Basquin's N = (S/S₀)⁻ᵐ. Statistical stack (Σ aᵢ²·tᵢ²)^½ ≤ T; minimizing the cost Σ kᵢ/tᵢ under it gives tᵢ ∝ (kᵢ/aᵢ²)^⅓.
- Substituted (illustrative). With m = 3, 10⁶ cycles at an amplitude of life 2 × 10⁶ give D = 0.5; 10³ further cycles at five times the amplitude, of life 16,000, add 0.0625, so 0.1% of the cycles carry 11% of the damage. Two dimensions of equal sensitivity with cost weights 1 and 8 under a statistical limit of 0.1 mm get 0.0447 mm and 0.0894 mm. Two features with sensitivities 1 and 4, equal cost and a budget of 10 µm at k = 3 get 16.0 µm and 6.35 µm where equal tolerances would be 7.28 µm each, and the allocation costs 19.9% less.
- Intuition. Rare overloads dominate life, so a campaign that never contains them certifies nothing; their rates come from Nemesis. Tolerance is cheapest to loosen where sensitivity is low and cost is high, and the cube root keeps the allocation from being extreme.
- Assumptions. Signed stress histories with residues carried between windows; a stress–life curve valid for the material state, process and finish; nonzero first-order sensitivities; independent feature deviations.
- Flip. A zero first-order sensitivity is not insensitivity (q = x² at a nominal of zero): the feature is refused and goes to the band-extreme check. A damage figure that moves by more than 5% when the window is halved fails the pilot of C-35. Every life figure is Estimated.
GB-38 · Does the structure loop converge inside its budget? Supports §31.9, ADR-FOUNDRY-07, GATE-FOUNDRY-06, FROB-REF-260 and GB-03.
- Equation. Link mass S sized to a stress cap and other mass M₀ satisfy S = A + λ·(M₀ + S), with fixed point S* = (A + λ·M₀)/(1 − λ) and an error that contracts by λ per iteration. Run in one loop with the actuator spiral of GB-03 the gain is about κ + λ. The estimated contraction is λ̂_k = ‖Δ_k‖/‖Δ_{k−1}‖, and the bound on the remaining error is λ̂/(1 − λ̂)·‖Δ_k‖.
- Substituted (illustrative). κ = 0.59 and λ = 0.15 give a combined gain of 0.74: five iterations contract the first step only to 0.74⁴ = 0.30 of itself, and the fixed point amplifies a structural change by 3.8, against 2.4 for the actuator spiral alone. With A = 0.5 kg, λ = 0.15 and M₀ = 10 kg the structure loop alone settles at S* = 2.35 kg, and four contractions leave 0.15⁴ = 5 × 10⁻⁴ of the first step.
- Intuition. Two geometric series in one loop converge slowly. That is why actuator cards are held fixed inside the structure loop and re-selected as an outer step with full revalidation.
- Assumptions. A linear model near the fixed point; the campaign re-run on the same world addresses, so the map is deterministic; a budget of five iterations and a 2% test.
- Flip. λ̂ ≥ 1 on two successive iterations is refused (FROB-REF-260), with the beam rung as the repair. Fresh random worlds at each iteration put a noise floor above 2% and the test cannot pass. A discrete stock size produces cycling, which one under-relaxed step may cure.
GB-39 · How long does a ten-second job wait behind a training run? Supports DC-09 (§28.19), §32.11, KILL-SLOT, KILL-PREEMPT and GATE-TETHER-10.
- Equation. One server with arrival rate λ and service time S. First come, first served: W = λ·E[S²]/(2·(1 − ρ)). Non-preemptive priority for the short class: W₁ = λ·E[S²]/(2·(1 − ρ₁)). Preempt and resume: the short class sees only itself, W₁ = λ₁·E[S₁²]/(2·(1 − ρ₁)), and long jobs slow by 1/(1 − ρ₁).
- Substituted (illustrative). One device at utilization 0.7; 90% of jobs take 10 s and 10% take 30 minutes. ES = 189 s, λ = 1/270 s⁻¹, E[S²] = 324,090 s² and ρ₁ = 0.033. A 10 s job waits about 2,000 s first come, first served; 621 s with non-preemptive priority; and 0.17 s plus one quantum of at most 5 s with preempt and resume. Long jobs slow by 3%.
- Intuition. The wait is the remainder of the training run already on the device, and only preemption removes it. Interactive rungs are what an agent's turn waits on, so this is the queueing decision that matters.
- Assumptions. Poisson arrivals, with Kingman's correction in the capacity table because swarms burst; snapshot boundaries cheap enough for a 5 s quantum; a resumed job that keeps its digests.
- Flip. A device reserved for interactive work, or a resume that costs several seconds or changes a digest (KILL-PREEMPT), makes running to completion the right policy.
E.6 Substrate and performance (GB-40 and GB-41)
GB-40 · Is that figure in world-steps or in env-steps? Supports §36.2, §23.3, INV-PERF-01 and GATE-SPARK-19.
- Equation. World-steps/s = d × env-steps/s, for a decimation d equal to the control period divided by the physics step. Ordering in env-steps/s: training ≤ full environment step ≤ physics only.
- Substituted (illustrative, with one FACT). 400 k env-steps/s of training at d = 10 is 4.0 M world-steps/s inside training, which cannot stand beside a physics-only figure of 2.0 M world-steps/s on the same card. FACT (Appendix H): one incumbent reports 1.20 M env-steps/s of physics as 4.8 M physics steps per second (d = 4) and 347 k env-steps/s for the full step, so physics is 29% of its environment step. A penalty of 1 per control step over 10 s totals 500 at 50 Hz and 1,000 at 100 Hz for the same motion.
- Intuition. A rate without its decimation is not a number. The same confusion in rewards and discounts changes what is optimized, which is why running reward terms are rates and the discount is exp(−β·Δt).
- Assumptions. Reference workloads at h = 2 ms and d = 10; incumbents measured at matched settings and at their native settings, reported separately.
- Flip. A task whose rational clocks make d non-integer reports the mean world-steps per control tick as a rational (GB-23). A comparison at native settings alone is a comparison of settings.
GB-41 · Does memory or throughput bound the number of worlds? Supports §36.10, §16.6, §17.7, GATE-SPARK-22, FROB-REF-602 and FROB-REF-740.
- Equation. Bytes per world in
f32with a dense Jacobian: 4·(61·n_b + 17·n) + 12·n² + 4·R·n + 48·R + 96·C plus a few kilobytes of state, for n degrees of freedom, n_b bodies, contact capacity C and row capacity R. PPO rollout storage: 4·T·(observations + privileged observations + actions + 4) bytes per world. - Substituted (illustrative byte model, with one FACT). Quadruped (n = 18, n_b = 13, C = 32, R = 160): 4,396 + 3,888 + 11,520 + 7,680 + 3,072 + about 4,000, about 35 kB. Humanoid (n = 35, n_b = 30, C = 64, R = 320): 9,700 + 14,700 + 44,800 + 15,360 + 6,144 + about 6,000, about 97 kB. 8,192 worlds are 0.28 GB and 0.79 GB, 1 to 3% of a 24 GB card. Rollout storage at a horizon of 24 with 48 observations, 187 height samples and 12 actions is 24 kB per world. A card with 16 to 19 GB usable holds 0.46 to 0.55 M quadruped worlds or 0.17 to 0.20 M humanoid worlds. The optional hard-contact operator adds 4·R², 102 kB and 410 kB. FACT (Appendix H): an incumbent measures 88.4 kB per humanoid world at 8,192 worlds.
- Intuition. Memory admits ten to a hundred times more worlds than the throughput targets use. The dense Jacobian is a third to a half of the bytes, so row capacity decides memory, and the bytes touched per step decide speed.
- Assumptions. A dense Jacobian; the capacities of §36.3, which are DESIGN defaults frozen at M2; scenes that are shared and never one per world (GB-07).
- Flip. Sampling control per design multiplies worlds by the sample count: 256 designs × 2,048 rollouts is 524,288 worlds, 18.4 GB at 35 kB, at or beyond what one card has usable. The path-table Jacobian replaces R·n by about R·(6 + depth) and removes most of the largest term.
E.7 Discrepancies found while recomputing
The recomputation covered every substituted number in GB-01 to GB-41 and the enumerations behind them: the Clopper–Pearson tables at n = 64, 86, 128 and 256 with their power rows, the first reachable tie at each qualification level, the peeking probabilities, the 24,546 ballast classes, the stance grid (2,254, 114 and 1,596), the docking grid, the wheel-loop spectral radii of §24.10, the thermal tangency, and the coverage laws. Everything not listed below reproduced the chapters' values. Each row was a correction request against the named section. The sections have since been corrected to the recomputed figure, and the cards carry it too; the third column records what was first printed.
| # | Where | The chapter's first value | Recomputed | Note |
|---|---|---|---|---|
| 1 | §26.9, the sentence after the table of clean-episode counts | The tenth request needs 76,963 clean episodes for 10⁻⁴ | 76,959 | 76,963 is ⌈ln(1/δ)/p⋆⌉, the approximation; the 29,956 beside it and the table use the exact rule ⌈ln δ/ln(1 − p⋆)⌉. GB-32 uses the exact rule |
| 2 | §36.4 item 1; §2.3 item 1 | 1/(1 − 0.87) = 7.8 | 7.7 from 0.87; 7.8 from the unrounded share 82/94 = 0.872 | The sections now print "about 7.7 (1/(1 − 0.87))". GB-09 prints both |
| 3 | §36.5 "Ceiling first" (§2.5 quotes the same 1.46) | About 94% of the batch was shared work, so a solver of zero cost would have reached about 1.46 | 1.345/0.94 = 1.43 | The share attribution is withdrawn. The ceiling of about 1.46 is 842.8 ÷ 575.7, where 575.7 µs is the residual the solver does not touch, not separately timed. GB-18 prints it that way |
| 4 | §16.11, "How much it can cover" | A humanoid batch at 0.5 M world-steps/s gives 1,000 world-seconds per second and a share of 0.5 to 2.4% | 3,400 and 0.14 to 0.71% at the 1.7 M planning value of §36.3 row 7 | 0.5 M is not a row of §36.3, and §36.9 already prints 0.14 to 0.7%. The order-of-magnitude range of 0.1 to 2% covers both; the humanoid figure in §16.11 is the stale one. GB-16 follows §36 |
| 5 | §28.9, "The ceiling" | With six survivors the ladder costs about 2.6 h with the audit lane | 2.0 h before the audit lane; 2.2 h with a tenth added; 3.0 h with 19 audit designs at the second rung | The stated inputs (81 designs; rungs of 10 s, 3 min, 30 min; six survivors; three finalists) do not determine 2.6 h. §28.9 now prints 2.0 h, and 2.2 h with a tenth added; GB-35 prints 2.0 h and the audit separately |
| 6 | §20.4.6 and the matching row of §20.20 | Every-thirteenth-step scheduling gives 615 decisions in 20 s instead of 600 | 616 decision instants in a half-open 20 s window, against 600 | 615.4 is the rate times the duration. The rate error of 2.6% is right; on counts the overpayment is 2.7% |
| 7 | §24.10, the wheel-loop row | An instantaneous-current approximation gives 2.7075 | 2.7075 without shaft damping; 2.6917 with the fixture's 0.002 N·m·s/rad | The 1.4820 and 0.9207 beside it reproduce only with the damping included, so the three figures use two plant models. No card uses the row |
Three inconsistencies between chapters surfaced that are not arithmetic errors. Each has since been settled, as its item says.
- Default capacities. §16.6 gave the reference quadruped C = 24 and R = 128 (27 kB per world) and the humanoid C = 40 and R = 224 (72 kB). §36.3 and §36.10 give C = 32 and R = 160 (35 kB) and C = 64 and R = 320 (97 kB), and §17.7 uses the latter for its 18-degree-of-freedom bucket. Settled: §36.3 governs and §16.6 now prints its values.
- Squeeze-film tolerance. GATE-HAB-06 held the three squeeze times of GB-21 to 1% and GATE-TRAC-05 holds the same times to 2%. Settled: both gates use 2%, the bound of §18, which owns the law.
- Energy-budget labels. The test of C-31 called the 0.25 J term stored energy and the 0.10 J term reserve. §22.11 calls them external work (gravity over the stroke) and stored energy (a series spring), with reserve as a further 0.10 J. Settled: the labels of §22.11 govern, the card's total is 1.49 J without the reserve, and the reserve shortens the admissible cutoff from 50.5 ms to 45.5 ms. GB-26 and the worked budget of Appendix L follow §22.11.
Two corrections were made to carried cards and are not discrepancies in any chapter. GB-04 printed 0.48 µm for the worst spacing inside an 8 m cell, which is the value for an offset in [4, 8); with the centred offset of §35.8 it is 0.24 µm. GB-05 printed only the small-angle cost of 0.57 m for a 0.10 m shift, where the optimal angle of 0.72 rad is not small; the exact cost of the same manoeuvre is 0.55 m, and the card now prints both.