# Hyperbolic PolyStep: preregistration

Registered 2026-10-29, before any solver code exists (Phase 0). The benchmarks, budgets, metrics and
decision rules below are fixed. A later change is recorded as a dated amendment at the end of this file,
with its reason. Results computed before an amendment are reported under the rules in force when they ran.

The form follows the open problems of *The Geometric Engine of Time* (§11): the evidence so far, a test,
a decision rule fixed in advance, and what depends on the answer.

## Claims under test

| # | Claim | Tested by |
| --- | --- | --- |
| C1 | On tree-structured problems the hyperbolic step beats the same step at curvature 0, at equal evaluation budget | B1, B3, B4 |
| C2 | Hyperbolic depth acts as built-in annealing: a fixed hyperbolic radius gives shorter Klein steps as a channel commits | B1, B2 |
| C3 | Three channels on the 24-frame schedule beat one channel with three times the particles, at equal budget | B1 ablation |

The properties H1 to H7 of the step are mathematical. They are proved and tested exactly, and are not part
of this preregistration.

## Common rules

- **Budget.** One evaluation is one candidate scored, as in polystep. Every method in a comparison gets the same budget.
- **Baselines.**
  - The primary comparison is the same code at curvature 0 (property H5 makes this exactly polystep's step). It isolates the effect of curvature.
  - polystep 1.1.1 itself (`PolyStep` or `PolyStepOptimizer`) on the flattened Euclidean problem is reported as a second baseline.
  - Random search at the same budget is reported as a floor.
- **Settings.**
  - Before any measured run, each method gets 24 tuning runs on separate tuning instances (seeds 9000 to 9023). The tuning covers ε, the radii and their cosine schedules.
  - The chosen settings are frozen and written to `reference/settings.json`.
- **Instances and seeds.**
  - Instance seeds are 1000 + i. Method seeds are 0, 1, 2.
  - Each instance is run once per method seed. Results are paired by instance and seed.
- **Statistics.**
  - Each effect is the mean paired difference (hyperbolic minus baseline), with a 95% percentile bootstrap interval over the pairs (10,000 resamples, seed 7).
  - The claim is **supported** if the interval lies above 0 and **refuted** if it lies at or below 0. Otherwise it is **inconclusive**.
  - Every cell is reported. No cell is dropped after the run.
- **Authority.** Decisions come from the Python reference. The browser page reruns small cells only to illustrate.

## B1 · Graph 3-colouring

- **Instances.**
  - Planted 3-colourable graphs (equipartition into three classes; edges only between classes, drawn uniformly) with average degree c ∈ {4.0, 4.4, 4.69}.
  - n ∈ {60, 120, 300}, with 20 instances per cell.
- **State.** One point of H² per vertex, so the state lies in the product (H²)ⁿ.
- **Step.** Each step moves a random block of 8 vertices. The corners are an orthoplex in the 16-dimensional tangent space of the block, applied factor by factor.
- **Decoder.** A vertex's colour is the branch (0°, 120°, 240°) nearest the angle of its Poincaré point. A point at the origin takes colour 0.
- **Cost.** The number of monochromatic edges, counted exactly by Helix B.
- **Budget.** 400 · n evaluations per run.
- **Method.** Three channels of 4 particles on the schedule, clutch "membrane" with strength μ from tuning.
- **Primary metric.** Success, meaning a proper colouring found within budget, scored 0 or 1.
- **Secondary metrics.** Final conflicts, and evaluations to success.
- **Ablations.**
  - C3: one channel of 12 particles at equal budget.
  - Membrane off (clutch "free").
- **C2 measure.** For every step of every run, the Klein length of the step and the mean Klein radius |κ| of the block before it. C2 is supported if the Spearman correlation is below −0.3 and its bootstrap interval excludes 0.

## B2 · MAX-3-SAT

- **Instances.**
  - Random 3-SAT at clause ratio 4.27 (polystep's generator).
  - n ∈ {100, 500, 2000}, with 20 instances per cell.
- **State.** Variables in blocks of 8, one H⁸ factor per block.
- **Decoder.** A variable is true when its Poincaré coordinate is positive.
- **Cost.** The fraction of unsatisfied clauses, counted exactly.
- **Budget.** 300 · n evaluations.
- **Primary metric.** Fraction of clauses satisfied at the end of the budget.
- **C2 measure.** As in B1.
- **External reference.** polystep's published MAX-SAT rows are quoted for orientation only. They used other budgets and are not part of any decision.

## B3 · Tree address search (with a negative control)

- **Tree.** The norm-5 tree of the paper, with 6 · 5^(L−1) addresses at depth L, for L ∈ {4, 6, 8}.
- **Embedding.** The tree is embedded in H² by Sarkar's construction, with fixed scale τ = 3.
- **Decoder.** A point decodes to the nearest node at depth L, by greedy descent from the root.
- **Graded objective.** The tree distance between the decoded node and the target. 30 random targets per depth.
- **Membership objective (the control).** Cost 0 at the target and 1 elsewhere.
- **Budget.** 6 · 5^(L−1) evaluations, the size of the depth-L level.
- **Primary metrics.**
  - Graded objective: the fraction of targets found.
  - Membership objective: the mean number of evaluations to the target.
- **Rules.**
  - **C1 on graded.** C1 is supported if the hyperbolic success rate exceeds curvature 0 by the common rule.
  - **Control.** Under membership, every method's mean evaluations to the target must be consistent with (N + 1)/2 for a uniformly placed target (Prop 5.13 of the paper). "Consistent" means that a two-sided 95% interval contains (N + 1)/2, or that the method fails within budget at the expected rate.
  - **Control failure.** If any method beats this bound significantly, the run is treated as a bug (information leaking from the target into the search) and investigated, not reported as a result.

## B4 · Hard-threshold network

- **Data.**
  - Eight classes at the leaves of a depth-3 binary tree. Inputs in R¹⁶ are drawn from hierarchical Gaussian means: each level adds a mean offset of halving size, plus noise σ = 0.5.
  - 2,000 training points and 1,000 test points per seed.
- **Model.**
  - An encoder W: R¹⁶ → R² sends x to exp₀(Wx) in H².
  - There are 8 class prototypes in H².
  - The prediction is the nearest prototype, a hard argmin with no useful gradient.
- **Parameters.** W lies in R³²; the prototypes lie in (H²)⁸.
- **Curvature-0 counterpart.** The same model with prototypes in R² and exp₀ the identity.
- **Budget.** 2 · 10⁵ evaluations per seed, with 10 seeds.
- **Primary metric.** Test accuracy.
- **Secondary metric.** Accuracy at the coarse level of the tree (pairs of leaves), to see where the errors fall.
- **Additional baseline.** polystep's `PolyStepOptimizer` on the Euclidean model.

## What none of this decides

- **P versus NP.** Its status is unchanged: it stays open.
- **SAT encodings.** A solver of this kind would give an integrable encoding of SAT only with a proof, not a benchmark (Theorem 5.11).
- **Black-box search.** No method can beat the volume of the space against a membership-only objective (Prop 5.13). B3's control checks the implementation against that bound; it does not test the bound.

## Amendments

### Amendment 1 · 2026-10-29 · implementation details fixed in Phase 1, before any tuning or measured run

None of these changes a preregistered setting (sizes, budgets, seeds, metrics, decision rules). They record choices
the reference had to make, so that the measured runs are judged against the method as built.

1. **Horizon.** Every factor of a state stays in the closed ball of radius 8 about the origin.
   - In hyperbolic geometry this is hyperbolic distance; at curvature 0 it is Euclidean distance.
   - A point that leaves the ball, after a step or after the membrane, is moved radially back onto it: the nearest point of the ball.
   - It is needed because in B1 and B2 only a point's direction is decoded, so particles can drift outward without limit. Beyond hyperbolic distance about 17 the exact snap's 2^-24 grid, and float64 hyperboloid coordinates, can no longer hold a point.
   - The projection never lengthens a step (H2 still holds) and never changes a decoded answer.
2. **B3 charts.** In B3 each particle is held in the frame of a tree node near it, and re-anchored after every step (a tiling-based coordinate model).
   - It is needed because at depth 8 (distance about 24) float64 coordinates measured from the root lose the sheet constraint.
   - The step is unchanged (H4), and B3's curvature-0 counterpart uses the same anchoring, with straight-line steps in the anchor's tangent plane.
   - The membrane is not used on B3, since it would mix points of different charts; B3 runs with the clutch on "free".
3. **polystep on B3.** The library's `PolyStep` has no re-anchoring hook. On B3 it searches the root chart and is limited by float64 beyond distance about 18 (depth 6). It is the secondary baseline; C1 is decided against curvature 0, which is anchored.
4. **Counting.** polystep's evaluations are counted by wrapping its objective, since its default polytopes differ from the orthoplex. The final scoring of its particles is not counted. It keeps no incumbent; its best final particle is certified.
5. **Halves round up** (`floor(x + 1/2)`), as the browser's `Math.round` does: in the snap, in the exact step length t, and in B1's edge count round(c n / 2). The two implementations then agree on every grid point and every instance.

### Amendment 2 · 2026-10-29 · fixed in Phase 4, before any tuning or measured run

The runner (`reference/run_bench.py`) is revised. No preregistered size, budget, seed, metric or decision rule
changes. These items fix procedures the preregistration left open, and correct one baseline before it is used.

1. **Tuning.**
   - **Cell.** Each method is tuned on one middle cell per benchmark: B1 n = 120, c = 4.4; B2 n = 500; B3 depth 6, graded; B4 its only cell.
   - **Candidates.** The same 24 settings for every method, drawn once from Python's `random.Random(9000)`:
     - ε log-uniform on [0.01, 2];
     - step radius uniform on [0.05, 3];
     - μ uniform on [0, 0.8];
     - a constant or a cosine schedule, with equal chance. The cosine schedule anneals ε and r to a tenth of their starting values over the run's expected number of steps. polystep gets its own `CosineEpsilon`; its radii are multiples of ε, so they follow.
   - **Trials.** Trial i runs candidate i on tuning instance 9000 + i with method seed 0.
   - **Choice.** The best trial by the primary metric wins, with ties broken by the lower certified cost.
   - **What is fixed or not tuned.** Block sizes stay at B1 8, B2 4, B4 6, and all factors for B3. The random floor is not tuned. B3's membership cells use the graded settings, since a membership objective gives tuning nothing to measure.
2. **Evaluations to success** are counted exactly. The objective is wrapped, and the count is the index of the first evaluation whose cost is 0 (a proper colouring, or the target). Before this change, the count was taken at the next zero point.
3. **The random floor.**
   - **The defect.** The draw it would have used took Poincaré coordinates from a cube that reaches outside the unit ball, which made some B4 prototypes invalid.
   - **The fix.** It now draws each hyperbolic factor's Poincaré point uniformly in the ball of radius 0.9, so B1's colours and B2's truth values are uniform. B4's encoder entries are uniform on [−2, 2].
   - **Batching.** Candidates are scored 64 at a time. B4 draws one training batch per 64, as a step draws one per step.
4. **How the rules are read.**
   - **C1** is decided in B1's cells, B3's graded cells and B4: hyper3 − flat3 on the primary metric. B2 and B3's membership cells report the same difference without a decision.
   - **C3** is decided in B1's cells: hyper3 − hyper1.
   - **C2**, per B1 and B2 cell. Take the hyper3 runs' Spearman correlations, their mean, and a bootstrap interval over runs.
     - Supported if the mean is below −0.3 and the interval lies below 0.
     - Refuted if the interval lies above −0.3.
     - Inconclusive otherwise.
   - **The B3 control.** For a uniformly placed target, the first k queries find it with probability at most k/N. Hence every method has E[min(T, N)] ≥ (N + 1)/2, where T is the evaluation that first scores the target and a run that never does counts as N.
     - The control is checked per method and depth with the bootstrap interval of the mean of min(T, N).
     - It fails, and is treated as a bug, only if the whole interval lies below (N + 1)/2.
   - **Descriptive comparisons.** The membrane ablation (hyper3 − hyper3_free) and the comparisons with polystep and the random floor use the same bootstrap. They are reported as effects, not claims.
   - **Floor flag.** A cell where both methods score 0 on every pair is flagged as a floor. The rule still applies to it as written.
5. **The ledger.** A claim's final status is decided over its cells:
   - **supported** if it is supported in a majority of its cells and refuted in none;
   - **refuted** if it is refuted in a majority and supported in none;
   - **mixed** otherwise, with the cells listed.
6. **Speed, not method.**
   - In float runs the membrane moves all factors at once, and the horizon clamp acts on all of a step's factors at once. The maths is the same, and a test checks one membrane against the factor-by-factor loop to 1e-10. Exact runs keep the loop, so the browser's traces are unchanged (regenerated byte for byte).
   - Runs execute in parallel processes. Each run is seeded on its own, so the order does not matter.
   - A run stops at the first zero point whose certified cost is 0 (B1, B3), and random search stops at its first proper colouring. Continuing a solved run cannot change its primary metric or its evaluations to success. polystep keeps no incumbent, so it always runs to the budget.
   - The B3 decoder looks up each node's children in a table instead of recomputing them. Its arithmetic is unchanged, and the tree trace regenerates byte for byte.

### Amendment 3 · 2026-10-29 · a defect found after the measured runs, fixed before any corrected run

**The defect.** The membrane moves each channel's particles by a rigid motion. In float runs the moved points were never put back on the sheet.
- **How it grew.** Near the horizon (x0 ≈ cosh 8 ≈ 1,490) these motions have entries around 2·10^6, so rounding grew from one membrane to the next. ⟨y, y⟩ drifted from −1 by more than 10^6, and states turned to NaN. NaN states decode to arbitrary answers.
- **How it was found.** The C2 records of B1 at n = 300 contained NaN.
- **Separately, a NaN in the browser core's transvection.** It did not guard a zero direction, which arises between points equal to rounding. The Python batched path already guarded it, and the Python factor loop now does too.

**What it affected.** Only `hyper3` on B1, B2 and B4: the only method that applies the membrane to hyperbolic factors. B3 runs without the membrane, and `hyper1`, `hyper3_free`, `flat3`, `polystep` and `random` never use it hyperbolically. In the original runs, the C2 records hold NaN in 2/180, 163/180 and 180/180 of B1's hyper3 runs at n = 60, 120 and 300, and in 3, 14 and 43 of the 60 B2 runs at n = 100, 500 and 2000.

**The fix.**
- Every float state the membrane moves is put back on the sheet by keeping its spatial part and recomputing y0 = √(1 + |ȳ|²). That is valid even when rounding has made ⟨y, y⟩ positive.
- The transvection returns the identity when the direction vanishes.
- Exact runs snap as before, so the traces are unchanged (regenerated byte for byte; parity 169/169).
- A new test runs the membrane near the horizon for 150 joint cycles. It fails on the old code and passes now (82 tests).

**The procedure, fixed now, before any corrected run.**
1. The original results and decisions are kept unchanged in `reference/results/pre_amendment3/`, and are reported beside the corrected ones.
2. `hyper3` is retuned on B1, B2 and B4 by Amendment 2's procedure: the same 24 candidates, the same tuning instances.
3. `hyper3`'s measured runs on B1, B2 and B4 are rerun with the same seeds. No other method's runs are rerun.
4. The decisions and the ledger come from the corrected runs. A claim whose status changes is reported with both statuses.

## Part F · The 0.7 Claude fence · preregistered 2026-11-03, before any measured run

The fence (`reference/hpstep/fence.py`, explainer 11) is Tukey's fence moved into the hyperbolic mass coordinate: the volume of the ball about the robust centre that reaches each point. Its lemmas CF1–CF9 are proved in `explainers/fence-tracer.html` and checked by `reference/tests/test_fence.py` and live in the browser. The runs below test what the lemmas do not settle: how the fence behaves on finite samples, against the baseline anyone would try first.

**Methods.**
- **Claude fence**: the reference `fence()` exactly as traced. Fréchet-median centre; exact verdicts in H², certified verdicts in H³. A point counts as flagged when it is outside or far out (verdict ≥ 1).
- **Distance-Tukey (DT)**: Tukey's 1.5 on the hyperbolic distance from the same centre, with the same order-statistic quartiles. It is computed in the same call.

**Datasets** (the shared generator; seeds 0–59 in every cell; N = 200 inliers):
- **Uniform balls**: `ball` with H² at R ∈ {1, 4, 8} and H³ at R ∈ {1, 3, 5}. These give six cells.
- **Planted outliers**: the same six cells plus 10 `shell` points at mass f · V(R), with f ∈ {1.75, 2, 3}.
- **Wrapped normals**: `gauss` with H² and H³ at σ ∈ {0.5, 1, 2}. These give six cells with no planted points.

**Metrics**, as cell means over the 60 seeds:
- the false-flag rate, the share of inliers flagged;
- the detection rate, the share of planted points flagged.

**Claims and predictions.** The predictions come from the population formulas of CF3 and CF7, not from runs.
- **CF-A (no false flags on hierarchy-shaped clouds).** On uniform balls, the Claude fence's false-flag rate is at most 0.70% in every cell. 0.70% is the share Tukey's fence flags on Gaussian data.
  - *Prediction:* about 0%. The fence sits at about 1.5 V(R), beyond every inlier.
- **CF-B (power where distance-Tukey is blind).** At f = 2, in the four broad cells (H² R = 4, 8; H³ R = 3, 5), the Claude fence detects at least 90% of planted points and DT detects at most 10%.
  - *Prediction:* about 100% against about 0%. With 10 planted points among 210, the Claude fence sits near 1.58 V(R) and DT near 3.5–4.1 V(R).
- **CF-C (a preregistered weakness).** On wrapped normals, the Claude fence flags more than 2% of inliers in every cell, and more than DT does in every cell.
  - *Prediction:* 5–17% against about 1%. A Gaussian cloud is not uniform in mass, and its mass has a long upper tail.
  - The fence is built for clouds spread as hierarchies spread. We expect, and will report, that it over-flags Gaussian clouds.

**Decision.** A claim is **supported** if its inequality holds in every listed cell, **refuted** if it fails in a majority of them, and **mixed** otherwise, with the cells listed. The other cells and the other f values are reported but decide nothing. Results go to `reference/results/fence.jsonl` (one record per run) and `reference/results/fence_decisions.json`.
