diff --git a/simulation/lab/IDEAS.md b/simulation/lab/IDEAS.md index 8a87f63fa..e171694e9 100644 --- a/simulation/lab/IDEAS.md +++ b/simulation/lab/IDEAS.md @@ -3,6 +3,39 @@ Unordered, mined during work. Promote to experiments when picked up. ## Simulator fidelity +- **The foreign balance sheet (READY, graph in hand 2026-07-29):** + dijkstrasden's revised model graph at + `~/codez/data/realistic_graph.json` — describegraph form, 11,255 + nodes / 37,203 edges, snapshot ~2026-07-28, plus per-edge `balance` + AND `balance_certainty` generated from ln-scores data. First + liquidity family NOT authored by us: soft U-shape (16.8% of + channels <5% of capacity, 15.9% >95%, ~10% at 40-60%) where our + generator is sharp-bimodal with an empty middle. Directly attacks + the WHY.md §0 circularity (mainnet tier = our synthetic balances). + Needs: loader flag to read `balance` instead of generating + (routing/sim_load.go, blocked on tree lock), og→new key mapping for + vantages (11/11 sealed-scenario keys map via `pub_key_og`), a + decision on `disabled` flags (1,906 both-sides, 7,447 one-side). + Extra fields (priv_key, announcement, update1/2) are for HIS + simulator; our loader ignores them — file parses as-is today. + `balance_certainty` (median 0.50) doubles as a served-confidence + signal for exp-016-style import experiments. His simulator itself + lands this week: running OUR champions in HIS simulator is the + cross-simulator escape from "simulator-shaped." +- **Fees as a liquidity prior (from dijkstrasden's report, + 2026-07-29, + https://fee-liquidity-correlation.lightning.wiki/):** advertised + fee rates predict depletion on mainnet — P(drained) RISES with a + direction's own fee and FALLS with the peer's reverse fee + (replicated on two snapshots a year apart; lnd-default 1ppm bucket + anomaly at 30.6% drained). No router of ours reads this signal: the + champions treat fees purely as cost, lnd's apriori likewise. A + candidate that seeds its interval prior from the two directional + fees gets a free, gossip-only information edge — and the + realistic_graph.json balances are generated FROM this model, so + that tier is also the one where the signal is guaranteed live. + Breed or hand-write; also an upstreamable idea for bimodal's + scale/mode weights. - Liquidity models from real `lncli querymc` data (dijkstrasden's BalanceHints idea) instead of synthetic distributions. - Mainnet `describegraph` snapshot corpus entries (loader exists;