How evidence from the planning exercises binds the plan
3,180 words2 tablesPlan version 1.1
Before this version was written, the load-bearing hypotheses of the program were put to refutation on paper: published results were re-read at pinned versions, arithmetic was redone, sibling code was inspected, and a small number of exact, small pilots were run, on CPUs and, for one boundary-classifier kernel, on the three GPUs (§21.8). No hypothesis was tested by running FrankenRobotics code. Each hypothesis received one of four verdicts, and the verdict binds what this plan may say and build.
| State | Meaning | What the plan text may say | What is in scope |
|---|---|---|---|
| ACTIVE | Unrefuted and untested. Nothing narrows it yet. | The hypothesis, labelled HYPOTHESIS, with its test. No number as a result. | The test is scheduled early. Work that depends on the hypothesis is gated on it, and the plan states what it does if the test goes against it. |
| NARROWED | The strong form failed on evidence. A stated weaker claim survives. | Only the surviving claim, with the surviving numbers as HYPOTHESIS ranges. The strong form appears once, in Appendix C, with the refuting calculation. | The narrowed mechanism, at the tier its evidence supports, with a falsifier and a kill gate. |
| REFUTED | A calculation or an executed check shows the mechanism does not deliver what was claimed. | The strong form is not written as a goal, a bet or a target. What survives, often as hygiene, is named. | The strong form is not built. It may be reopened only by the reversal condition recorded with the verdict. |
| DEFERRED | Real, but not an answer to the question that was asked of it (here: time to a decision). | What it does deliver, under its own heading. It is not counted in the thesis of §2. | Built off the critical path, at F or M, or not at all. |
The register follows. Numbers in the "what survives" column are ranges read from published work or computed from it (sources in Appendix H). Each is a HYPOTHESIS about this system until the gate of the named section records a measurement.
| ID | Hypothesis | State | What survives | Where it binds |
|---|---|---|---|---|
| H-01 | Sampling control with zero training | NARROWED | A screening rung and an out-of-distribution fallback, at about 0.3–1 GPU-minute per quadruped episode and 1–5 per humanoid episode: cheaper than one training run up to about 30 episodes per quadruped design and 60 per humanoid design at a 20 ms planning step (6 and 12 at 4 ms). The published successes are legged robots with hand-built swing or stepping references; there is no success rate over a varied corpus and nothing for tracked robots. It never stands in for a trained policy as acceptance evidence. Where the sampling controller is itself the delivered controller, its own episodes are the evidence; the break-even against training once is 2.0 GPU-seconds per 20 s episode for a quadruped and 12 for a humanoid, which published settings miss by 10× and 5×. | Reflex has a sampling controller as a screening rung, a fallback and a demonstration source, and a delivered-controller lane with a frozen planner record (§22.8; BET-27). |
| H-02 | A universal policy prior | NARROWED | Within one morphology class, on flat or mildly varied ground with fixed actuation: zero-shot at 70–85% of specialist return; fine-tuning at about 2–25% of a from-scratch run stopped at matched return (1–10% against the fixed budget one source used). Across topology-diverse designs at affordable pretraining: zero-shot at 10–40%, a fine-tuning gain of 2–3× at best and none on flat ground. Within a class on interior terrain, with payload or actuator-card changes: no published figure. | Family policies per morphology class, not one universal policy. A gated experiment on interior terrain, against an early-stopped from-scratch control, that needs an owner-approved compute envelope (§23; BET-15). |
| H-03 | Topology as data | NARROWED | Mixed-topology populations with few worlds per design, in size buckets. One incumbent solver publishes a mixed batch of three body plans at 0.86 of the harmonic mean of its homogeneous rates. That solver is maximal-coordinate, is documented as not deterministic, and is reported as slower than the reduced-coordinate solvers of its framework. Three differentiators are left: measured mixed-batch throughput, the time from an edit to the first GPU step, and run-to-run replay and forks inside a mixed batch, which no incumbent document read claims. | Per-world model variants inside padded capacities; a head-to-head experiment with a per-family-template control arm and a digest check inside the mixed batch (§17; BET-18). |
| H-04 | Supercompiled kernels | REFUTED for recursion-only specialization | The Lie-algebraic recursions are 16–20% of two incumbent step traces, so the ceiling is about 1.2× and reaching it needs the recursions more than five times faster. Derived derivative kernels for CPU solvers survive. Reversal condition: a specializable share of at least one third on our own profile. | Specialization only for survivors, and only past the reversal condition. Derivation stays, for correctness, certificates and CPU solvers (§21; BET-01). |
| H-04b | Structure-exploiting constraint assembly and solve | ACTIVE | Constraint assembly and the solve are 62% of the same incumbent trace: a ceiling of 2.6× if both vanished and 1.7× at a three-fold reduction. Ancestor-sparse rows and a tree-ordered factor cut operation counts 2.6× to 4.1× for the factor and 3.8× to 8.1× per foot row. The shares of those operations inside the solve stage are unmeasured, fewer operations do not imply a faster device step, and the sparsity is carried as data, not compiled per design. | An operation ledger on the CPU tier and a device measurement at M2, with a kill gate. Not counted in the thesis of §2 (§16). |
| H-05 | Gait continuation | NARROWED | Full-order orbit solves take about 4–260 s of CPU at 20–68 states: a screening signal for periodic gaits on regular ground, at about 1% to 63% of a 410 s training run. Not a referee-ready controller. | A gait-orbit analysis with continuation in design parameters, labelled screening (§22; BET-26). |
| H-06 | Region certificates | NARROWED | Certificates per constraint. Bisection is confined to a split set of about six sensitive parameters; the others are enclosed at full range, and a box of any dimension is admitted when that enclosure is narrower than the constraint's margin. Isotropic paving of a 10–20 dimensional box is out of reach. Whether robot constraints have narrow weak-variable enclosures is unmeasured. | Family answers are built from per-constraint certificates, monotonicity and witnesses, not from brute paving (§12, §25; BET-20). |
| H-07 | A rare-event engine | NARROWED | Search finds failures; a bound comes from a proposal frozen before the confirming draws, with bounded weights, or from crude sampling. Published measured gains are 2–20× and about 5× in mean-square error; 10–100× at 10⁻⁴–10⁻⁵ is a TARGET. Proposals may act on static scene parameters as well as per-step disturbances; only snapshot reuse needs per-step noise. A failure mode the proposal misses costs up to 1/ε times crude sampling, so absence is never certified more cheaply. | Nemesis separates exploration from certification, and its bound claims refutation and tightening (§26; BET-32). |
| H-08 | Counterfactual autopsy | NARROWED | Interventional contrasts reported with a flip rate under perturbations at or above round-off scale. Temporary interventions fork from the snapshot; design-parameter interventions re-run from the episode start. Precision follows measured sensitivity: at the published exponent, single-precision round-off stays below 4 × 10⁻⁵ of state scale over fork windows up to about 1 s and reaches 2% at 2 s. Double precision is the default for one failure and the escalation for marginal cases. | Autopsy reports the stability of each attribution, uses a placebo fork, and may attribute on the originating tier under an admission rule (§27; BET-23). |
| H-09 | Incremental evaluation | REFUTED for exact reuse | Exact reuse of rollouts after a mass or geometry edit is at most about 10% of compute. Training avoided by warm starts is a different quantity; it belongs to H-21 and is counted once. A memo graph for analyses, step plans, kernels and scene compiles survives as hygiene. Dependency-directed revalidation of verdicts survives as an evidence rule. | The memo graph extends past model identity; "which verdicts survive an edit" is a verb; a reuse ledger keeps exact reuse, carried evidence and avoided reacquisition apart (§28; BET-33). |
| H-10 | Value of information | NARROWED | 2–5× from stopping and elimination when candidates are separated; about 1× otherwise. | Experiment planning is used at the finalist stage and for choosing measurements (§28). |
| H-11 | An always-on referee | NARROWED | A sampled snapshot referee on any backend. A correctness guard that saves no compute. Its coverage is unmeasured and scales with referee speed: about 0.1–2.4% of GPU world-seconds if the referee runs at 0.2–1.0× real time per core on 24 cores, and more if it is faster. Its samples, drawn with recorded inclusion probabilities, also serve the paired tier correction. | A background referee sampler that reports its coverage and detection delay per hardware and backend pair (§16; BET-34). |
| H-12 | Generative structure | DEFERRED as a decision-time lever | Links 7–30% lighter than engineered parts: 2–8% of robot mass on the one quadruped with a published mass breakdown, and no figure for humanoids. It improves the output, not the time to a decision. | Foundry is an output-quality subsystem (§31; BET-29). |
| H-12b | Member sizing inside screening | ACTIVE | A closed-form check that a declared link wall survives the design's own loads keeps screening from rewarding links that cannot be built. How often it changes the winning actuator or gear ratio is unmeasured. | The beam rung runs at preflight and blocks promotion of under-built designs (§31, §14). |
| H-13 | Requirements replace rewards | NARROWED | Bounded-time requirements as the acceptance metric and as a falsification score. Not a replacement for shaped rewards. | Missions are requirements; training keeps reward terms (§24; BET-24). |
| H-14 | Conformal reality bands | NARROWED | Coverage within one calibrated robot and task family, varying by about six points at n = 20. Nothing for designs that were never built. | Ground issues bands per calibrated family and refuses elsewhere (§30; BET-28). |
| H-15 | Half precision | REFUTED | The condition number times half-precision round-off is at or above one in 37–64% of sampled contact states, and refinement alone costs more than the single-precision factorization. | Not built. Single precision with diagonal scaling (§35; BET-08). |
| H-16 | Living multiphysics in every world | REFUTED as stated | Full-room fluid is out of memory by orders of magnitude, and coupled cloth and cable at batch speed are out of reach on present foundations. A contact-level film-friction law and rigid proxies are the S substitutes; whether they reproduce the failures of coupled models is untested. | The film law and the proxies are S. The proxy-adequacy pilot decides the proxies' claim and whether a coupled island is ever funded; islands are M (§19; BET-36). |
| H-17 | From mission to design by solving | NARROWED | Trajectory optimization with design variables as local refinement (about 12 degrees of freedom, up to 17 parameters, 4–10 minutes). Polynomial synthesis only inside the symbolic library's caps. Monotone catalog selection of actuators, batteries and compute, with no measured timing. The pass rate of synthesized designs under independent closed-loop evaluation is unknown. | A bounded exact synthesis lane, a refinement lane and catalog selection; continuation is M, on the CPU; every synthesized candidate passes the ordinary ladder (§25). |
| H-18 | Information, not the engine, is the lever | ACTIVE | No instrumented design session exists. Where every candidate is trained, training is about 96% of an illustrative campaign and information features that do not remove training cap near 1.04×. Once controller acquisition is amortized, agent turns are about two thirds of elapsed time. Which regime a mission is in is a measurement. | The ten-fold test measures decisions, with a phase ledger and three single-change ablations (§2, §39). |
| H-19 | A thin layer over an incumbent would do | ACTIVE | Decided by a measured recovery ratio, not by a count of gaps. Documented engine-internal gaps: a mixed-topology batch with run-to-run replay; a Metal tier. Unverified: replay of one world outside its batch; solver-internal certificates. In the incumbent documents read, run-to-run determinism, heterogeneous worlds and the fastest solver sit on different solver paths, so each is available and their conjunction is not shown. | The ten-fold test has three arms and reports how much of the gain a composed incumbent recovers; the plan says what it does if that arm ties (§39; BET-14). |
| H-20 | Proof-guided control | NARROWED | Reduced-order barrier shields with a statistically measured margin. No full-order certificates. | An energy-budget tier and runtime filters, classed Tested (§22, §12; BET-25). |
| H-21 | Policy inheritance | NARROWED | No published figure measures a small typed parameter edit from parent to child. Transfer from a multi-embodiment prior to a similar body costs about 2–25% of a from-scratch run stopped at matched return; the one parent-to-child figure is about 50%. Single-limb topology edits: 33–100%. Chained inheritance risks loss of plasticity and narrows morphological diversity. | Inheritance for parameter edits must beat an early-stopped from-scratch control per edit class before it is the default. Resets follow a measured regression of an audit twin, with a depth schedule as fallback (§23; BET-16). |
| H-22a | Rank fidelity between rungs | ACTIVE | No published rank agreement between a shared policy and specialists was found. The nearest figure, 0.40, compares a critic with its own policy. A high rank correlation does not protect the winner: with 40 designs and the true best at cheap rank 3, τ is 0.91 on the top ten and a cut to two discards it. | The rank-agreement experiment is the decisive early experiment of the program, and it reports containment and regret beside τ (§23, §28; BET-17). |
| H-22b | The fast tier as a paired control | ACTIVE | A paired correction gains 1/(1 − ρ²) before cost: 2.0× at 0.7, 2.8× at 0.8, 5.3× at 0.9. The correlation is unmeasured and is reported separately for continuous metrics and for binary outcomes. | The tier-correction ledger of §16 and §29. |
| H-23 | A family per dispatch | NARROWED | Fewer rounds are plausible; measured compute gains in published population methods are about 3–4×. Paired finite-difference gradients are unproven over long contact-rich horizons. | Families per dispatch, with short-horizon paired differences (§17, §28). |
| H-24 | A geometric-algebra tree policy | NARROWED | Invariance to heading and to description conventions is necessary, and augmentation recovers most of it. There is no locomotion evidence for the geometric-algebra specifics: expect 0–1.3×. | An M experiment with a plain morphology transformer as the control arm (§23; BET-30). |
| H-25 | Brackets on what a body can achieve | ACTIVE | A controller-independent upper bound and an achieved lower bound on task utility; controller effort stops only when a checked upper bound falls below a competitor's lower bound. The bound is often vacuous for contact-rich full-order tasks. | §25.32, §28. |
| H-26 | Common randomness compiled from the task | ACTIVE | Separate address domains for world fields, absolute time and task-triggered events keep paired draws paired across retiming and topology changes. A paired contrast's variance scales with 1 − ρ. Any change in a marginal distribution is a failure of the mechanism. | §24.22, §35. |
| H-27 | Step adjoints with short-horizon joint training | ACTIVE | Implicit-function adjoints through the converged contact solve are exact for a fixed active set and cheap. Several incumbents differentiate their step, so this is no differentiator. Gradients say nothing about outcomes decided by a change of contact sequence. | M, off the critical path (§16.21, §23.29). |
| H-28 | Placement and sizing along a prescribed motion | ACTIVE | Inverse dynamics is linear in inertial parameters and contact forces for a fixed motion, so placement and sizing is one cone program. Its certificate excludes a motion, not a design; gear ratios are enumerated; link lengths are not variables. | §25.33. |
| H-29 | Visibility and observability in sensor placement | ACTIVE | Visibility along a motion with the robot's own limbs and payload as occluders is a cheap, missing analysis. The observability index is exact only without bias states. The share of estimated-state failures it explains is unknown. | §20.24. |
The question the program was asked is open: no matched mission has been run on any pipeline. No single mechanism delivers ten-fold, and no conjunction has been shown to. §2 states the candidate conjunction and the two comparators it must be measured against. The register above was challenged after it was first adjudicated, and sixteen of its rows were corrected; the corrections are in the rows as printed. These disagreements remain open, and each has a test that settles it:
- Whether a thin layer over an incumbent recovers the gain. Settled by the recovery ratio of §39 (BET-14).
- Whether design missions are dominated by training or by agent turns and diagnosis. Settled by the phase ledger and the three ablations of §39.
- Whether cheap rungs rank designs as specialists do, and how well tiers correlate. Settled by the rank-agreement experiment (§23, §28; BET-17).
- Whether a family policy holds within a class on interior terrain, and what an inherited edit costs. Settled by BET-15 and BET-16 against early-stopped controls (§23).
- Whether a sampling controller is more than a screen. Settled by BET-06 and BET-27 on the design corpus, and by the delivered-controller lane (§22).
- Whether a structure-exploiting solve beats parity, and whether per-design code adds to it. Settled by the operation ledger and our own M2 profile (§16, §21, §36).
- Whether rigid proxies reproduce the failures of coupled models. Settled by the pilot of §19.10.7.
- Whether formal certificates change any design decision. There is no evidence either way. Lemma records, for every certificate consumed by a decision, whether the decision would have differed without it (§12; BET-11, BET-20).
The same evidence binds three further things. Differentiators that the incumbent audit withdrew are not claimed anywhere in this plan; §1.3 lists them. Numbers that the audit corrected replace the numbers of version 0.2 wherever they appear; §2 and §36 carry the corrected figures. A verdict changes only on new evidence of the kind that produced it: a pilot, a measurement at a gate, or a primary source re-read at a pinned version. The register is append-only, and a changed verdict keeps its history.