[v0.4.0] Vault: equity-cushion / book-quality model #54
Winnie
Loading…
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Part of #39. Thesis: vault models risk as loss-absorbing capital and acquired-book quality - what a failed bank's book is worth to it, how much equity cushion it has, bond premium as compensation for time and default. Scenarios: base, bond-economy, thin-margins, tight-cash. Distinct from the other three brains: no shared calc helpers. Bench numbers in comments, PR per winning iteration.
Vault agent: start. Working #54 now.
Thesis locked in: vault is the treasury-and-vulture. It models risk as loss-absorbing capital and acquired-book quality:
Signature moves: lend into the bond market when implied return clears the hurdle; vulture failed banks' books at a discount it computed itself; hold capital instead of chasing volume when the cushion runs thin.
Scenarios: base, bond-economy, thin-margins, tight-cash. Every step will be logged here with bench numbers.
Baseline measured (30 games/preset, v3 on current master
7f503c4, gauntlet off):Overall vault: 4/390 wins (1.0%), avg eq 1,582. The v3 vault is structurally broken: it pays above mill for deposits (avg−0.5t vs avg−0.75t) so it loses every Vickrey loan; it prices loans at funding + 1.2× the die rank (highest risk price in the market) so it never wins a loan; and its bond offers price a 0.2%/turn return against a ~2% deposit cost (negative carry). Only the 0.97× auction nibble does anything.
Iteration 1 hypothesis (v4 rebuild — treasury-and-vulture):
avg − rate-temp×(1 + 1/cushion)— when equity ≫ book, vault is the cheapest funder in the market (below mill), which flips the loan market to it via Vickrey; when the cushion thins, vault sheds deposits and holds capital.equity / (loan-book + 0.5×loan-min)— its one world observable, used everywhere.1−(1−1/faces)^termamortized over the live market term (no shared flat1/faceshelper) + a capital chargerate-temp/cushionthat prices vault out of volume when equity runs thin.net × (1 − 0.5×cum-default(market term))— full price in safe markets (d100 → 0.96×net, beats adapt's 0.93), a real haircut in risky ones — with a liquidity buffer so bidding never starves the treasury.Building + smoke-testing now.
Iteration 1 measured (v4 rebuild, 30 games/preset, quick bench):
Overall: vault 168/390 wins (43.1%, from 4), avg eq 29,124 (from 1,582). Empire 120, mill 93, adapt 0. No
[aierrors, no accounting divergences, smoke clean.Verdict: the cushion model works — vault is now the cheapest funder (its loan bids clear when the cushion is deep), the vulture haircut wins auctions in safe markets, and tight-cash is already vault's. 1 of 4 scenarios won. Still short: in thin-margins vault has no income at all (deposits cost ~1,000/turn while its loan bid of funding+1% clears the 1.2% cap only when funding is dirt cheap — and its deposit base is still big when it matters); in bond-economy the whole market returns ~0.1%/turn and vault sits in cash while mill grinds out tiny loan income.
Iteration 2 hypothesis:
Spectating all 4 scenarios now to verify where vault's cash actually goes.
Iteration 2 measured (v5: zero deposit floor + book-quality loan gate + halved bond floor; 30 games/preset):
Overall: vault 144/390 (36.9%), avg eq 25,665. Mill 91, empire 151, adapt 0. Clean smoke, no
[aierrors.What the v5 fixes did: paying 0 for deposits when nothing deploys turned vault from a bleeder into the cheapest funder (bond-economy avg eq 1,290 — best in scenario, 2× mill's 626 — but mill still edges the win count 13–10); the book-quality gate stops vault growing a book the market is bleeding (bond-economy's 3.08% loans vs 3.33% honest loss). tight-cash held.
Remaining gaps: bond-economy wins are within noise of mill (13–10, seat-order ties favor mill seat 2 over vault seat 3); thin-margins is empire's by design (its duration-discount bid is tailor-made for a 1.2% cap) and vault sits second at 7; base unchanged at 3 — the loan gate keeps vault out of base's 2.8–3.0% early market (right call) but vault has no replacement income there besides bond lending.
Iteration 3 hypothesis: inspect the base auction books — if failed books are positive-net there, vault's 0.87× haircut is leaving every bargain to adapt's 0.93×, and the vulture haircut should be tightened (the 0.5 loan-share assumption is the over-conservative dial). Also: loosen the bond-offer cushion gate so vault's treasury lending fires in base (cushion > 1.5 blocks a 20k-equity treasury with a small book — the gate should be cushion > 1.0).
Iteration 3 measured (v6: bleed-latch loan discipline — quick bench, 30 games/preset):
Overall: vault 152/390 (39.0%), avg eq 21,269 — near-par with empire 160.
What the v6 latch does: after the market hands vault a book that yields less than the honest per-turn loss (1−(1−1/faces)¹×tolerance), vault marks the market as bleeding it and refuses to grow that book until the field of bidders shifts (a bank dies → re-probe). This killed the bond-economy death spiral — vault had been re-entering the loan market every time its book emptied, buying 3.08% loans against 3.33% true default risk in waves. Now it sits out until the clearing restructures, and bond-economy + tight-cash are both vault's (2 of 4 scenarios ✓).
Measured variants: adding a time-based latch reset (re-probe after 1×/2× bond-term) REGRESSED bond-economy to 14–16 wins (the probe re-buys the bad market) and didn't move base — rejected, reverted to plain v6.
Base remains empire's franchise (its 2-seat deposit machine); thin-margins is empire's duration-discount scenario. Neither collapsed: vault avg eq 5,703 / 4,383, surviving everywhere.
Running the FULL bench now (100 games × 13 presets + gauntlet, ~1300+5200 games) to confirm at scale before the PR.
FULL BENCH (100 games × 13 presets + gauntlet) — iteration 3, v6 confirmed:
Overall: vault 431/1300 wins (33.2%), avg final equity 23,036; empire 573 (44.1%), mill 285 (21.9%), adapt 0. No
[aierrors, no accounting divergences, smoke clean.Vault wins 2 of its 4 scenarios: bond-economy (59 vs mill 23) and tight-cash (58 vs empire 26). Gauntlet bonus: vault takes loan-bond-spec 160/1040.
Base (10) and thin-margins (15) are empire's franchise scenarios — vault survives both with healthy equity (5.9k / 4.9k avg), no collapse.
Opening PR on branch
54-vaultnow.PR #59 opened (branch
54-vault, robertismo requested). Continuing to iterate on the same branch.Iteration 4 (post-PR, v7→v7b, quick bench 30 games/preset): dropped the per-bid capital charge — the bleed latch and the deposit shed already do the capital-holding, and the charge was pricing vault's fair-priced bids above mill's in thin-margins. Also tried a latched-deposit-shed (pay 0 while the loan market bleeds us) — it REGRESSED bond-economy (7 wins: the shed starves vault of pool-inflow cash that funds the next fair-priced book), so I reverted just that piece.
(*full bench, 100 games) — v7b overall: vault 183/390 (46.9%, top brain), empire 138, mill 64, adapt 0. Clean smoke, no errors.
v7b now wins 3 of 4 scenarios on the quick bench. Full bench running to confirm at scale; will push a second commit to
54-vaultwith the numbers.FINAL — vault done on #54. PR: #59 (branch
54-vault, 2 commits, reviewer robertismo requested).The treasury-and-vulture wins 2 of its 4 scenarios at full-bench scale and is the top brain overall:
Overall (100 × 13 presets): vault 585/1300 (45.0%), avg eq 23,805 — empire 502, mill 199, adapt 2. No
[aierrors, no divergences. Gauntlet: loan-bond-spec 160/1040.Measured iteration history (all in iterations.json, base
7f503c4): v3 baseline 4/390 (0 wins in all four) → v4 cushion rebuild 168/390, tight-cash won → v5 zero deposit floor + honest-loss gate (bond-economy 10/30) → v6 bleed latch (full bench: 59 + 58) → v7b charge-free bid (full bench: 81 + 58, thin-margins 40). Four measured iterations, one rejected variant (latched deposit shed — regressed bond-economy).Distinct world model (Robert's ruling): everything derives from the equity cushion and acquired-book quality — compound default survival priced over the market's live term (not the flat die-rank helper the other houses share), bond lending priced as time + default compensation with the term inside the price, vulture bids computed from the die model, capital held when the cushion thins. No hard-coded values: every threshold is config-derived.
Left open: base and thin-margins remain empire's franchise scenarios; vault survives both with healthy equity and no collapse.
PR #59 merged — vault v4 (treasury-and-vulture: cushion-driven capital, book-quality loan gate, bleed-latch discipline) landed on master. Wins bond-economy 23/30 and tight-cash 25/30. Shipped; closing.