Marcin Pawlowski ac6a309e58
Review fixes + high-precision design-band delay study
Acts on a correctness/completeness review of the countable uncle model
and its report material.

Correctness fixes in the report:
- s3.4 quoted 0.998 for W_abs=10 at the 8s budget; the run says 0.9963.
- s1 claimed both models >= 0.996 at U >= 1; countable U=2 delta=8 is
  0.9955. Corrected to >= 0.995.
- The s3.2 table presented two cells (U=1 at delta 16 and 32) as model
  differences. They are not resolvable: t = 0.46 and 0.47 over 5
  replicates. The table now carries +-SEM and a t per cell.
- s3.4 claimed the ~7-block-interval floor "carries over unchanged".
  Accuracy is still climbing past W=7 at every delay (8s: 0.989 ->
  0.996), so the claim is dropped. The 32s curve is non-monotonic with
  replicate SD up to 0.22 and is now flagged as noise, not a trend.
- 1-r was attributed to the first-fork restriction alone; it is the
  combined first-fork and capacity loss, which this measurement cannot
  separate. Hedged to match fig32's own axis label.

Completeness: the U=0 negative control was swept but never reported.
With no uncles the two models are identical by construction, yet they
differ by -0.23 at delta_max=32 (t=2.1) because they draw independent
RNG streams. That is the noise floor the rest of the grid must clear,
and it is now in s3.2, s9, fig30 and the config header.

New study (configs/fine-delay.yaml, scripts/plot_fine_delay.py, s3.2a,
fig34/fig35): the design band delta_max 1-5 at 40 replicates, both
models. Findings: every U >= 1 cell of both models lands in
0.998-1.001, flat in delay, while U=0 decays 0.810 -> 0.640. No
individual cell resolves a model difference (widest 95% CI +-0.15pp;
max t=2.59 vs Bonferroni 2.94 over 15 cells). Pooled across uncle caps
the first-fork cost is monotone in delay and separates from zero only
at delta_max=5 (-0.0014 +- 0.0007, t=3.7) -- below 0.15% everywhere in
the band, against +-0.9% per-epoch sampling noise.

Code:
- deep_ref_share is identically 0 on every real countable run: for a
  chain block B the producer's chain below B is the counting chain
  below B, so the counting-side parent-on-chain re-check cannot reject
  what selection emitted. It is a drift alarm, not a rate. Documented
  as such in measure.py, the plot docstring and the config header, and
  pinned by a new end-to-end test.
- Removed annotate_uncles: a second countable implementation that
  production never called, while carrying most of the selection test
  coverage. Tests now drive select_uncles_at_production through an
  annotate_via_production replay helper -- same assertions, live path.
- Added tests for the two previously uncovered branches of the live
  selection: the pmin/below chain walk that resolves parent-on-chain
  for candidates whose parent sits below the window, and the
  occupied-slot exclusion built from the chain walk.
- theory.q_effective and theory.window_miss_prob were unused and
  untested. Now used (the prediction figure reconstructs q_u through
  the identity the report quotes) and tested. The window_miss_prob test
  records that its "~ e^-W" docstring is the f->0 limit: the true decay
  is e^-1.017W at f=1/30, 16% off by W=10.
- Shared sem()/recovery_rate() moved into figures_pernode.py; fig30 and
  fig33 regenerated with SEM error bars and the U=0 control curve.
- Fixed the pre-existing E501 in bootstrap_dynamics.py; ruff clean.

Report prose reworked to read standalone: the countable model is
described as the rules under analysis and the former model as a
labelled "unrestricted" comparison baseline, with no dated banners and
no round-to-round narration.

Tests: 209 passed (was 202).

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-04 20:48:38 +02:00
..
2026-07-30 18:57:10 +02:00
2026-07-30 18:57:10 +02:00
2026-07-30 18:57:10 +02:00
2026-07-30 18:57:10 +02:00
2026-07-30 18:57:10 +02:00

tsi-sim-pernode — Cryptarchia TSI per-node network simulator (Phase 2)

The reduced-model simulators (../tsi-sim/, ../tsi-sim-mc/) collapse the network to one global canonical chain and one scalar D_est per epoch. This package removes that collapse: every one of the N nodes runs TSI individually with its own D_est, from its own partial view of the block tree under explicit message propagation over a peering graph. Its job is to test the reduced model's assumption that all honest nodes agree.

What it models

  • Per-node lottery: node i wins a slot with φ_f(w_i / D_est_i)D_est is a length-N vector, each node self-updating from its own view (the reduced model's key reuse: the sparse sampler already takes a per-node probability vector).
  • Topology (topology): three propagation models over the network.
    • full_mesh: every node one hop away with uniform latency L — reproduces the reduced model exactly (validation baseline).
    • regular: a random d-regular peering graph (configurable degree) with per-link latency (link_latency_dist ∈ {fixed, uniform, exp, geo}, all with mean link_latency_mean). A block reaches a node after the shortest weighted path from its producer (gossip flooding). Models direct block gossip.
    • blend: the same d-regular graph, but a block is first routed through the Blend mixnet before it is public — the producer picks blend_hops distinct relay nodes uniformly at random, the block hops producer → r₁ → … → r_hops over the graph, each relay waiting a Uniform(0, blend_delay_max) mixing delay before forwarding, and the last relay's forward is the final network-wide gossip that makes the block visible. Relays are blind forwarders (they learn the block only from that final gossip). The dominant latency is the per-hop mixing, not the graph transport — this is the multi-slot regime where forks and the stake underestimate appear and uncle references matter. Because the mixing delays are Uniform-bounded, the windowed fork choice stays exact (horizon (blend_hops+1)·max_path_latency + blend_hops·blend_delay_max).
  • Real-world latency (units). Latency is in slots and a slot is 1 s, so measured internet latencies (tenshundreds of ms) are fractions of a slot; arrivals are therefore kept sub-slot (float), not rounded to whole slots. link_latency_dist=geo draws each link from a geographic band mixture (~15 ms metro → ~200 ms antipodal, EU↔EU ≪ EU↔AU), rescaled so link_latency_mean stays the mean-latency knob. So regular runs the realistic sub-slot direct-gossip regime (~0.050.2 slot), where forks are rare, and blend runs the multi-slot Blend-mixnet regime, where per-hop mixing delays dominate.
  • Per-node views: one global block tree plus an (N × n_blocks) arrival matrix A; each node builds on / measures density over the blocks that have arrived at it. Uncle refs are baked at production from the producer's view (faithful — immutable once adopted).
  • Uncle model (uncle_model, CLI --old): the default countable model implements the spec's counting-only rules (cryptarchia-v1-protocol.md): only the first block of a fork (parent on the producer's chain) is referenceable/countable, the window is derived as w_u = window_absorption / f slots (W expected block-intervals, default W = 10 → 300 slots, bounded W ≤ 0.6·k), selection skips slots already occupied on the producer's chain and picks one uncle per slot, and the measurement pass re-checks every rule per reference (rejections tallied as deep_ref_share). Passing --old to tsi-sweep/tsi-verify runs the pre-redesign model unchanged — window = uncle_window slots, any-depth orphans referenceable, every baked reference counted — and bit-reproduces historical runs (the old model's RNG key is byte-identical to the pre-uncle_model key).
  • Metrics: per-node D_est spread (range, IQR), canonical-chain agreement (window prefix vs current tip), mean accuracy, and — with init_dest=heterogeneous — transient re-convergence.

Headline result

Per-node D_est disagreement collapses to zero. Because TSI reads density from a window buried far past k-finality, and all nodes seed the recursion from a common hardcoded genesis D, every node computes the same measured density midentical D_est (range ≈ 0, agreement_window = 1) — even under a sparse graph with high latency and heavy tip-level forking (agreement_tip can drop well below 1). This validates the reduced model. Topology/latency instead shift the shared mean accuracy (via fork rate → q), which uncle references recover just as in the reduced model. (Injected heterogeneous disagreement, which the real protocol never creates, is preserved by the common multiplicative update — a cautionary note, not protocol behaviour.)

Quick start

make install         # venv + editable install
make test            # unit + fast per-node checks
make verify          # per-node validation (parity, spread→0, agreement, topology effect)
# Run any configs/<name>.yaml by its stem (auto-discovered); each writes a dated runs/ folder:
make smoke           # tiny end-to-end grid + figures  (configs/smoke.yaml)
make default         # scaled-k divergence/topology sweep + figures  (configs/default.yaml)
make fullscale       # full-scale (true k) confirmation  (configs/fullscale.yaml)
make figures RESULTS=runs/<dir>/results.parquet          # re-render figures from a run
# Extra sweep flags: make fullscale SWEEP_ARGS="--batch-size 1 --mem-frac 0.6"

Scale & performance

  • Representation: one global block tree + (N × n_blocks) float64 arrival matrix A (sub-slot arrivals); topology path_latency[N,N] (per-node Dijkstra, once per trajectory).
  • n_blocks is NOT ~10·k in general — it tracks block production. n_blocks is the number of lottery wins in an epoch, ≈ E·Σᵢφ(wᵢ/D_est). At equilibrium that is ~10·k (≈ 22k at k=2160), but when D_est is far below the true stake — the collapsed-estimate regime, e.g. a small genesis_d_factorΣ(stake)/D_est = 1/genesis_d_factor is large and block production explodes proportionally. At genesis_d_factor=0.01, genesis epoch-0 produces ~2.0M blocks (100× equilibrium) → A ≈ 15 GB for a single worker; D_est self-corrects to equilibrium within ~2 epochs, so only the earliest epoch(s) are heavy. Raising genesis_d_factor toward 0.10.5 collapses this cost (0.1 → ~0.22M blocks → ~1.6 GB; 0.5 → ~44k → ~0.4 GB) and does not change the equilibrium result, which is measured after burn-in.
  • Sliding-window prune (prune_arrival, default on): the arrival matrix never needs per-node columns for blocks past the horizon — under deterministic latency a block with slot ≤ t H has reached every node, so its column is finalized and dropped. We keep columns only for blocks inside max(horizon, uncle_window) slots in a base-offset buffer, turning the O(N·n_blocks) matrix into O(N · keep-span-blocks). This is what makes the collapsed regime affordable: at N=1000/k=2160/gdf=0.01 the buffer is ~tens of MB instead of the ~15 GB full matrix (fork choice, the parent clamp, uncle selection, and per-node tips all reconstruct exactly from it). It is bit-identical to the full matrix at jitter_mean == 0 (proven by test_prune_matches_full_matrix across topologies/uncles/gdf); with jitter it falls back to the full matrix (whose safety clamp keeps the tree valid). Set prune_arrival: false to force the full matrix (the parity oracle). The measurement pass also argmaxes in node-row bands so it adds only a small temporary. Divergence sweeps run at scaled k=256 (configs/default.yaml); full-scale k=2160 is validated to N ≤ 2000.
  • Worker sizing (auto, RAM-safe): both A (~N·n_blocks, incl. the block explosion above) and path_latency (~N²) grow, so the sweep runner sizes the loky pool to fit a RAM budget (--mem-frac, default 0.7 of physical RAM) instead of blindly using every core. The per-worker estimate realises the seeded stake to compute the genesis-epoch block count (expected_peak_blocks), so it reflects a low-genesis_d_factor explosion rather than assuming ~10·k. A calibration probe measures a real worker's peak RSS (one genesis epoch of the heaviest config in a spawned process) whenever the estimate is heavy or N > 2000 (--calibrate {auto,always,never}, default auto; the probe bounds itself to physical RAM so it fails loud rather than freezing).
  • Fail-loud memory guard (memguard.py): every worker checks size before allocating both big arrays — the (N × n_blocks) A (in build_tree_pernode) and the (N × N) path_latency (in build_path_latency, built first) — and raises ArrivalMatrixTooLarge if it would exceed the budget TSI_ARRIVAL_BYTES_BUDGET. The sweep sets that to each worker's RAM share; unset or 0 is not "unlimited" — it resolves to DEFAULT_BUDGET_FRAC (0.9) of physical RAM, so a bare run_trajectory, tsi-verify, the probe, or a --mem-frac 0 run all keep an absolute per-process ceiling. So a mis-estimated block explosion (or a huge N) fails with a clear message instead of freezing the machine.
  • Cost: dominated by the per-node fork choice (batched per slot) and the arrival-matrix fill; the sparse lottery is negligible. Across-config joblib loky parallelism reused.
  • Measurement optimisation (measure.py): the per-node canonical/density/agreement pass was ~95% of an epoch. It is now deduped by tip (nodes sharing a tip share every derived quantity — high agreement collapses N to a handful of computations) and the per-tip chain walk runs as a cached numba kernel (pure-Python fallback if numba is absent). Exact — bit-identical to the naive loop (test_measure). Measured ~9× end-to-end (heavy config 11.3 s → 1.2 s) and ~14× on measurement-bound configs. numba comes via the accel extra (pip install -e ".[dev,accel]", done by make install).
  • Windowed fork choice (windowed_fork_choice, default on): bounds the block-tree build's per-slot candidate scan to a horizon of the max path latency plus the fully-propagated best tip, turning O(n_blocks^2) fork choice into O(n_blocks*H). Exact when link latency is deterministic (jitter_mean == 0) — bit-identical to a full scan (parity test). With jitter_mean > 0 it becomes a tiny approximation and warns; a safety clamp still keeps the tree valid, and windowed_fork_choice=False forces a guaranteed-exact full scan.
  • Reproducibility: every draw spawns off SeedSequence(hash(config)) — child 0 stake, 1 graph, 2 init, 3+e epoch e. graph_seed/degree/link_latency_* are part of the config identity.
    • numpy-version caveat: the accel extra (numba) requires numpy<2.5, so installing it pins numpy to 2.4.x. numpy's Generator.choice(replace=False) is not stream-stable across the 2.4↔2.5 boundary, and the sparse lottery uses it heavily only in the degenerate collapsed-estimate regime (D_est → 0 ⇒ win-prob → 1 ⇒ count ≈ n_slots). So a run on numpy 2.5 and a run on numpy 2.4 give identical results for all normal configs but can diverge chaotically in that one extreme regime (e.g. degree=4, link_latency=8, where the estimate has already collapsed to ~0.13 — off the safe chart). The differences are tiny (max |Δ mean_ratio| ≈ 3e-3) and change no conclusion; pin numpy if bit-reproducibility across environments is required.

Layout

src/tsi_sim/   constants config rng stake lottery topology blocktree(+build_tree_pernode)
               uncles(+select_uncles_at_production) tsi(+update_D_vec) epoch engine metrics
               theory verify  plotting/{style, figures_pernode, make_figures}
configs/       smoke.yaml  default.yaml  fullscale.yaml
               countable-vs-old.yaml  absorption-window.yaml   (countable-model studies)
               fine-delay.yaml   (delay 1-5 at 40 replicates: the design band, high precision)
tests/         test_{pernode,config,rng,lottery,blocktree,uncles,tsi_counting,stake,
                     theory,latency,theory_convergence,countable_counting,...}.py
scripts/       plot_countable_vs_old.py  (countable-vs-unrestricted comparison figures)
               plot_fine_delay.py        (design-band accuracy + model gap with 95% CIs)