e9b79ce90a
Two gaps left by the previous review. Section 3.4 quoted the attribution bracket at N=20,000 while the only committed evidence carrying those columns was the timing run at N=2,000, so a reader diffing report against data saw different numbers for the same quantity. Added configs/attribution.yaml and a make target: it records both bounds and the graph hop distance at the reported scale, cheaply, since the adversary and deanonymization metrics are closed-form and the hop distance is a property of the topology. It reproduces the section exactly -- L = 2.58 and neighbourhood confidence 0.640 at degree 8, f_adv 0.2. It also surfaces a result the smaller run could not: degree cuts both ways. A sparser graph has longer routes, so it offers the adversary more upstream places to see the message -- L is 4.18 at degree 4 against 1.93 at degree 16, lifting neighbourhood confidence from 0.61 to 0.72. The low diameter that makes propagation fast also starves the adversary, one of the few places where raising the degree helps anonymity rather than hurting it. Section 3.11 was the only section without a figure. Fig 25 plots MAP success against the effective anonymity set for both release designs: the dashed sets separate far faster than the solid best-guess curves, which is the whole argument for not trusting perplexity alone. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
3.2 KiB
Evidence of record
The sweep outputs behind every number in the report. The simulator does not commit
its own runs/ directory — these are the copies of record, kept so that any figure or table can be
re-derived, or challenged, without re-running hours of compute.
Each run directory holds the three tables the simulator writes: propagation.parquet,
adversary.parquet and deanon.parquet.
| directory | config | sampling | backs |
|---|---|---|---|
default/ |
configs/default.yaml |
1 000 rounds × 8 seeds = 8 000/cell | §3.1–§3.5 — delay, observation, eclipse, deanonymization, delivery, coverage |
redundancy/ |
configs/redundancy.yaml |
1 200 × 8 = 9 600/cell | §3.8 — messaging redundancy R = 1…4 |
percolation/ |
configs/percolation.yaml |
800 × 8 = 6 400/cell | §3.5 — the churn threshold u_c = 1 − 1/(degree − 1) |
correlated-churn/ |
configs/correlated-churn.yaml |
800 × 8 = 6 400/cell | §3.9 — correlated AS/region outages vs uniform churn |
fullscale/ |
configs/fullscale.yaml |
64 × 3 = 192/cell | §5 — the 10⁶ scaling check (deliberately lighter; not a source of headline numbers) |
cover-traffic/ |
configs/cover-traffic.yaml |
900 s timeline × 4 seeds | §3.10 — blending, mixing, and the emission-quota stake ceiling. Carries a fourth table, traffic.parquet |
attribution/ |
configs/attribution.yaml |
closed-form + 4 seeds | §3.4 — the attribution bracket at the report's scale: local confidence, attributable fractions, upstream hops, neighbourhood confidence |
timing/ |
configs/timing.yaml |
120 s timeline × 3 seeds | §3.11 — the two release designs under a timing attack, and the minimum-interval control |
The linkability results (§3.6–§3.7) and both deanonymization rates are closed forms over these
tables rather than separate measurements, so they have no run of their own — blend.linkability
derives them and make verify checks them against Monte-Carlo.
Regenerating the report's numbers
python report_numbers.py
prints every quoted value straight from the parquets here -- the §3.1–§3.5 and §3.8 tables with
their across-topology standard errors, and the §3.9–§3.11 tables and the §3.4 attribution bracket
from their own runs.
That is the fastest way to check a table in the report against its evidence. It takes optional
paths (report_numbers.py <default> <redundancy> <percolation>) if you want to point it at fresh
runs instead.
Regenerating the data itself
From tools/simulators/blend: make sweep,
make redundancy, make percolation, make correlated-churn, make sweep-fullscale. Results
land in that simulator's runs/<timestamp>_<label>/. Note that the seed streams depend on the
configuration, so re-running reproduces the statistics, not bit-identical numbers, unless the
config is unchanged — in which case it does reproduce exactly.
Two runs from the same session are deliberately not kept: the smoke runs (throwaway, far too
noisy to interpret) and an earlier 144-rounds/cell redundancy grid that was superseded because its
sampling error produced a non-monotonic delivery curve — the reason redundancy/ samples 9 600.