Depth-weighted crown rate band (CROWN_RATE_BAND)

[anderdc]

What

Generalizes the crown's exact-rate tie-break into a rate band: every qualified miner whose quote is within CROWN_RATE_BAND (0.5%) of the best qualified rate holds the crown together, and each interval's credit splits in proportion to collateral depth via the new crown_depth_shares helper.

• A one-tick undercut no longer takes 100% of a direction from a deep miner — evicting the band requires a rate more than 0.5% better, so depth becomes the contested axis inside the band while genuinely better pricing still wins outright.
• Depth is capped at required_collateral(max_swap) — collateral beyond a full-capacity fill buys no extra share (same never-pay-to-win bound capacity_factor applies at 1.0).
• Proportional splitting is sybil-neutral: splitting one collateral pile across N hotkeys yields exactly the same total share, minus N× the fee overhead.
• CROWN_RATE_BAND = 0 restores exact-best-rate behaviour (band off), tunable in constants.py without redeploy.

Why 0.5%

Sized empirically from all 2,011 completed mainnet swaps + live quote ladders (2026-08-03):

anchor	value	role
rate tick (5 sig figs)	0.002–0.01%	sniping granularity the band must dwarf
median executed-rate drift, <1h	0.004–0.03%	honest quote staleness the band must absorb
rate distance takers actually accepted (p50–p75 vs same-6h-bucket best)	0.02–1.5%	differences below ~0.5% demonstrably don't decide routing
gap to the next genuine quote tier	~2.6%	materially worse pricing the band must exclude

The live SOL-TAO ladder's competitive cluster ends at ~0.75% from best; every band width from 0.75% to 2.5% produces identical membership today, and ≥3% starts crediting the 2.6%-worse tier. 0.5% keeps band escape economically live (a genuinely better-priced miner can post +0.6% and take the crown alone) while ending one-tick crown theft.

Consistency

The scoring replay (credit_interval), the live snapshot (snapshot_current_crown_holders), and the persisted intervals (intervals_to_crown_rows) all flow through the same crown_holders_at_instant(rate_band=…) + crown_depth_shares pair, so the dashboard tip cannot diverge from the rewarded ledger. Interval rows now carry per-holder proportional credit (still sums to 1.0 per interval, same DB shape); snapshot rows all carry the band's anchor rate.

Behaviour change

test_rate_tie_broken_by_collateral (deepest-takes-all on exact tie) is superseded by test_rate_tie_splits_by_collateral_depth — an exact tie now splits by depth rather than awarding the whole interval to the deepest miner. Dead-heats on rate + collateral still split evenly (degenerate case of the proportional split).

Tests

• Band membership: near-rates join (both sort directions), band anchors on best qualified rate (a busy leader can't strand the band), unqualified members excluded, band 0 = exact-only.
• crown_depth_shares: proportional split, full-capacity depth cap, all-zero-depth even fallback.
• End-to-end: thin 0.4% undercut inside the band shares 1/6 instead of taking all; deep quote keeps 5/6; a 1.5% quote outside the band earns nothing.
• Full suite: 848 passed.

Out of scope (follow-ups)

• allways-db / UI: crown snapshot & holder rows are shape-compatible but the UI may want to render multiple concurrent holders with fractional credit.
• min_collateral raise (admin tx, no code) and any ε retune once post-band equilibrium data exists — the two levers are deliberately independent.

🤖 Generated with Claude Code