Marcin Pawlowski 4baccd8d4b
Paired design: resolve the design band with common random numbers
The unpaired comparison could not answer the question it was asked. The
two uncle models draw independent RNG streams -- uncle_model is in the
config key, which is what makes --old bit-reproduce earlier runs -- so
the arms differed in stake draw, peering graph and every lottery
outcome, each comparison paid the between-run variance twice, and the
per-cell floor (+-0.0015) sat an order of magnitude above the effect.
Only delta_max = 5 resolved, and only after pooling.

Adds `paired_streams`: the RNG root is derived from the model-
independent part of the key, so a countable cell and its --old twin get
the SAME stake, graph and lottery draws and the uncle rule is the only
difference. Each replicate is then a matched pair and the shared
variance cancels. Trajectories still diverge after epoch 0 through the
genuine feedback (a different counted density changes the next epoch's
difficulty), which is the signal.

The flag is deliberately NOT in key(): it selects which key the seed is
derived from, so including it would perturb every historical seed.
Re-verified that --old still bit-reproduces the committed 2026-07-27
rho-boundary parquet, max |delta| = 0.

Results (configs/fine-delay-paired.yaml, 40 replicates per arm):
- Negative control becomes an IDENTITY check. With U = 0 no reference is
  taken, so shared streams must give bit-identical trajectories. All 200
  replicate pairs differ by exactly 0.0. Unpaired, the same control only
  had to agree within +-0.025 and drifted by 0.016.
- Per-cell SE shrinks by a median 1.6x (1.2-2.1x); widest 95% CI goes
  +-0.0015 -> +-0.0010. 5/15 cells resolve at |t| >= 2 (0.75 expected by
  chance); the largest, U=2 at delta_max=4, is t = 4.32 and clears
  Bonferroni for 15 tests.
- The cost is a STEP, not the ramp the unpaired data suggested:
  delta_max 1-3 unresolved (t = 1.1, 1.8, 1.4), then delta_max 4 AND 5
  both resolve at -0.0011 (t = 4.7) and -0.0009 (t = 3.7). Whole-band
  pooled -0.00060 +- 0.00021, t = 5.7 -- where the unpaired estimate of
  the same quantity (t = 2.8) had failed correction.

So the first-fork restriction costs nothing measurable up to
delta_max = 3 and about 0.1% at 4-5 -- an order of magnitude below the
+-0.9% per-epoch sampling noise.

Two bugs found while building this, both of which would have silently
produced a wrong answer:
- paired_streams was missing from metrics._CONFIG_FIELDS, so it never
  reached the parquet; plot_fine_delay.py falls back to the unpaired
  test when it cannot confirm pairing, so the sweep would have completed
  and quietly reported the old result. Caught before the run finished;
  the sweep was restarted and a test now pins the field.
- The U=0 control check reported FAILS on a PERFECT control: paired, the
  gap is exactly 0 so its SE is 0 and t is 0/0. It now checks the gap
  itself when the streams are shared, and falls back to the t-test only
  when there is real spread.

§3.2a is rewritten around the paired measurement; the unpaired sweep is
retained in §9 as the power comparison that motivated it. Figures 34-35
regenerated, with the control annotation and provenance reflecting the
design actually used.

Tests: 214 passed (was 209). ruff clean.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-05 11:58:26 +02:00

215 lines
10 KiB
Python

import dataclasses
import pytest
from tsi_sim.config import SimConfig, SweepConfig
def test_expand_cardinality_and_u0_strategy_dedup():
# full_mesh, where `latency` (L) IS a live axis, so the x2 latency factor applies.
sweep = SweepConfig(
n_nodes=[1000, 2000],
stake_dist=["uniform"],
topology=["full_mesh"],
latency=[0, 4],
max_uncles=[0, 1, 2],
uncle_strategy=["oldest", "random"],
replicates=3,
base={"k": 16, "epochs": 5},
)
configs = sweep.expand()
# U=0 keeps only the first strategy; U>0 keeps both.
# per (n,dist,lat): U0 x1 strat + U1 x2 + U2 x2 = 5 strat-U combos, x3 reps = 15
# x 2 n_nodes x 1 dist x 2 lat = 60
assert len(configs) == 60
u0 = [c for c in configs if c.max_uncles == 0]
assert all(c.uncle_strategy == "oldest" for c in u0)
assert {c.replicate for c in configs} == {0, 1, 2}
def test_latency_collapsed_for_graph_topologies():
# `latency` is the full_mesh-only uniform-L knob; regular/blend ignore it, so sweeping it
# must NOT emit duplicate (seed-shifted) graph cells.
for topo in ("regular", "blend"):
sweep = SweepConfig(
n_nodes=[100], topology=[topo], degree=[4], latency=[0, 2, 4], max_uncles=[0],
uncle_strategy=["oldest"], replicates=1, base={"k": 8, "epochs": 3},
)
configs = sweep.expand()
assert len(configs) == 1, topo # 3 latency values collapse to 1
assert configs[0].latency == 0
def test_base_propagation():
sweep = SweepConfig(n_nodes=[500], stake_dist=["pareto"], latency=[2], max_uncles=[0],
uncle_strategy=["oldest"], replicates=1,
base={"k": 32, "epochs": 7, "fixed_point": True})
(c,) = sweep.expand()
assert c.k == 32 and c.epochs == 7 and c.fixed_point is True and c.stake_dist == "pareto"
@pytest.mark.parametrize("kwargs", [
{"k": 0}, {"epochs": 0}, {"n_nodes": 0}, {"latency": -1}, {"max_uncles": -1},
{"uncle_window": 0}, {"lottery_chunks": 0}, {"uncle_random_p": 1.5}, {"f": 0.0},
{"f": 1.0}, {"beta": 0.0}, {"genesis_d_factor": 0.0}, {"pareto_shape": 0.0},
{"stake_dist": "zipf"}, {"uncle_strategy": "newest"}, {"topology": "star"},
{"blend_hops": 0}, {"blend_delay_max": -1.0},
{"adversary_period": -1}, {"adversary_withhold_epochs": -1},
# a schedule may not withhold for more epochs than its own period
{"adversary_period": 2, "adversary_withhold_epochs": 3},
# blend needs `blend_hops` distinct relays from the non-producer pool (n-1 of them)
{"topology": "blend", "n_nodes": 4, "degree": 2, "blend_hops": 4},
])
def test_validation_rejects_bad_fields(kwargs):
with pytest.raises(ValueError):
SimConfig(**kwargs)
def test_blend_dedup_does_not_multiply_non_blend_configs():
# blend_hops / blend_delay_max only affect blend runs; sweeping them must not duplicate
# the regular / full_mesh cells.
sweep = SweepConfig(
n_nodes=[100], stake_dist=["uniform"], topology=["regular", "blend"],
degree=[4], link_latency_mean=[1.0], link_latency_dist=["fixed"],
blend_hops=[2, 3], blend_delay_max=[1.0, 3.0], max_uncles=[0],
uncle_strategy=["oldest"], init_dest=["common"], replicates=1,
base={"k": 8, "epochs": 3},
)
configs = sweep.expand()
regular = [c for c in configs if c.topology == "regular"]
blend = [c for c in configs if c.topology == "blend"]
assert len(regular) == 1 # blend knobs collapsed for non-blend
assert len(blend) == 4 # 2 hops x 2 delay_max
assert {(c.blend_hops, c.blend_delay_max) for c in blend} == {
(2, 1.0), (2, 3.0), (3, 1.0), (3, 3.0)}
def test_uncle_window_sweeps_and_collapses_for_u0():
# OLD model: uncle_window is a live axis for U>0, but U=0 references no uncles so it
# must collapse. (The countable model ignores uncle_window entirely — see the twin
# test below.)
sweep = SweepConfig(
n_nodes=[100], stake_dist=["uniform"], topology=["blend"], degree=[6],
link_latency_mean=[0.5], link_latency_dist=["geo"], blend_hops=[3],
blend_delay_max=[4.0], uncle_window=[10, 100], max_uncles=[0, 1],
uncle_strategy=["oldest"], init_dest=["common"], replicates=1,
base={"k": 8, "epochs": 3, "uncle_model": "old"},
)
configs = sweep.expand()
u0 = [c for c in configs if c.max_uncles == 0]
u1 = [c for c in configs if c.max_uncles == 1]
assert len(u0) == 1 # W collapsed for U=0
assert {c.uncle_window for c in u1} == {10, 100} # both W kept for U=1
def test_window_absorption_sweeps_and_ignored_axis_collapses():
# COUNTABLE model: window_absorption is the live window axis; uncle_window is ignored
# and must collapse. And vice versa for the old model (guarded above).
sweep = SweepConfig(
n_nodes=[100], stake_dist=["uniform"], topology=["blend"], degree=[6],
link_latency_mean=[0.5], link_latency_dist=["geo"], blend_hops=[3],
blend_delay_max=[4.0], uncle_window=[10, 100], window_absorption=[2.0, 4.0],
max_uncles=[0, 1], uncle_strategy=["oldest"], init_dest=["common"], replicates=1,
base={"k": 8, "epochs": 3},
)
configs = sweep.expand()
u0 = [c for c in configs if c.max_uncles == 0]
u1 = [c for c in configs if c.max_uncles == 1]
assert len(u0) == 1 # all window knobs collapse
assert {c.window_absorption for c in u1} == {2.0, 4.0} # live axis kept for U=1
assert {c.uncle_window for c in u1} == {10} # ignored axis collapsed
def test_unknown_sweep_key_rejected():
with pytest.raises(ValueError, match="unknown sweep keys"):
SweepConfig.from_dict({"latencies": [0, 1], "base": {}}) # typo: latencies vs latency
def test_from_dict_roundtrip_ok():
sw = SweepConfig.from_dict({"latency": [0, 3], "max_uncles": [0, 2], "replicates": 2,
"base": {"k": 8, "epochs": 4}})
assert sw.latency == [0, 3] and sw.replicates == 2 and sw.base["k"] == 8
def test_key_covers_every_field():
# Guard against the silent shared-RNG bug: key() must reflect all run-affecting fields.
# root_seed enters _entropy separately; windowed_fork_choice is a pure compute
# optimisation (no RNG, identical results) so it is intentionally not in key().
# early_stop is truncation-only (per-epoch RNG streams are pre-spawned, so the epochs
# that DO run are bit-identical to a full run's prefix) — intentionally excluded from key().
# uncle_window is read ONLY by the old model; under the (default) countable model it is
# an ignored field, deliberately left in the base tuple at its old position so that an
# --old run's key stays byte-identical to historical keys.
# paired_streams must NOT be in key(): it selects WHICH key the RNG root is derived from
# (see seed_key), so putting it in key() would perturb every historical seed and break
# --old bit-reproduction. Its own behaviour is pinned in test_rng.py.
ignored = {"root_seed", "windowed_fork_choice", "prune_arrival", "early_stop",
"uncle_window", "paired_streams"}
names = {f.name for f in dataclasses.fields(SimConfig)} - ignored
a = SimConfig()
for name in names:
cur = getattr(a, name)
alt = _perturb(cur)
b = dataclasses.replace(a, **{name: alt})
assert a.key() != b.key(), f"key() does not distinguish field {name!r}"
# ... and uncle_window IS distinguished under the old model, where it is live.
old = SimConfig(uncle_model="old")
assert old.key() != dataclasses.replace(old, uncle_window=old.uncle_window + 1).key()
# paired_streams leaves key() untouched but DOES change the seed derived from it.
a_paired = dataclasses.replace(a, paired_streams=True)
assert a.key() == a_paired.key()
assert a.seed_key() != a_paired.seed_key()
def test_old_model_key_is_historical():
# --old must bit-reproduce historical runs: its key is exactly the pre-uncle_model
# tuple (no uncle_model / window_absorption entries), and the countable key extends it.
old = SimConfig(uncle_model="old")
new = SimConfig()
assert new.key()[: len(old.key())] == old.key()
assert new.key()[len(old.key()):] == ("countable", new.window_absorption)
# window_absorption is ignored (and absent from key) under the old model...
assert dataclasses.replace(old, window_absorption=2.0).key() == old.key()
# ...and live under the countable model.
assert dataclasses.replace(new, window_absorption=2.0).key() != new.key()
def test_effective_uncle_window():
# countable: derived w_u = round(W / f); old: uncle_window taken directly.
assert SimConfig(k=2160).effective_uncle_window == 300 # W=10, f=1/30
assert SimConfig(k=2160, window_absorption=5.0).effective_uncle_window == 150
assert SimConfig(uncle_model="old", uncle_window=42).effective_uncle_window == 42
def test_window_absorption_bound():
import warnings
with pytest.raises(ValueError):
SimConfig(window_absorption=0.5) # W < 1 rejected
with pytest.warns(RuntimeWarning, match="exceeds the spec bound"):
SimConfig(k=8, window_absorption=10.0) # W > 0.6*k warns
with warnings.catch_warnings():
warnings.simplefilter("error")
SimConfig(k=2160, window_absorption=10.0) # full scale: silent
SimConfig(k=8, uncle_model="old", uncle_window=300) # old model: no bound
def _perturb(v):
if isinstance(v, bool):
return not v
if isinstance(v, int):
return v + 1
if isinstance(v, float):
# stay inside [0, 1]-capped fields (e.g. jitter_frac defaults to 1.0)
return v - 0.001 if v >= 1.0 else v + 0.001
flips = {
"uniform": "pareto", "oldest": "random", "full_mesh": "regular",
"fixed": "exp", "common": "heterogeneous", "suppress": "withhold",
"exp": "poisson", # jitter_dist
"sine": "ramp", # churn_mode
"countable": "old", # uncle_model
}
if isinstance(v, str) and v in flips:
return flips[v]
return v