Winnie/bankwars
Watch
1
0
Fork
You've already forked bankwars
 0
Code Issues 9 Pull requests 2 Projects Releases 6 Packages Wiki Activity Actions

[#54] Vault: treasury-vulture iteration 4 — wins bond-economy + tight-cash (thin-margins near-miss) #59

Merged
robertismo merged 1 commit from 54-vault into master 2026-09-29 19:26:51 +00:00
Conversation 0 Commits 1 Files changed 1 +178 -35
Winnie commented 2026-09-29 19:26:22 +00:00
Owner
Copy link

Vault: treasury-and-vulture (issue #54)

Thesis: vault models risk as loss-absorbing capital and acquired-book quality. One world observable — the cushion (equity / (loan book + 0.5×loan-min)) — drives every decision. Risk is modeled as compound default survival 1 − (1 − 1/faces)^term amortized over the live market term (no shared flat die-rank helper). Signature moves: lend into the bond market when the implied return clears compensation for time + default, vulture failed books at a discount it computed itself, hold capital when the cushion runs thin.

Full bench (100 games × 13 presets, latest commit 0e727e1)

scenario vault wins vault avg eq winner
base 13/100 7,324 empire 63
bond-economy 81/100 1,152 vault
thin-margins 40/100 8,824 empire 50
tight-cash 58/100 2,066 vault

Overall: vault 585/1300 wins (45.0%, top brain), avg final equity 23,805 — empire 502 (38.6%), mill 199 (15.3%), adapt 2. Gauntlet bonus: loan-bond-spec 160/1040. No [ai errors, no accounting divergences, smoke clean.

Vault wins 2 of its 4 scenarios at scale (bond-economy 81/100, tight-cash 58/100) and is a near-miss on thin-margins (40 vs empire 50, with the best in-scenario equity 8.8k vs 5.9k). Base is empire's franchise scenario; vault survives it with healthy equity (7.3k), no collapse.

Journey (iterations.json, base 7f503c4)

  1. v3 baseline: 4/390 wins, 0 in all four scenarios (paid above-mill deposits, highest risk price in the market, negative-carry bonds).
  2. v4 rebuild: cushion-driven deposits + amortized compound risk + vulture haircut → 168/390, tight-cash 21/30.
  3. v5: zero deposit floor + honest-loss book gate + front-of-queue bond pricing → bond-economy 10/30, tight-cash 17/30.
  4. v6: bleed latch (mark the market when it hands us a book below the honest loss; re-probe only when the field shifts) → full bench: bond-economy 59/100, tight-cash 58/100.
  5. v7b: dropped the per-bid capital charge (the latch + deposit shed do the holding; the charge was pricing fair bids above mill's in thin-margins) → full bench: bond-economy 81/100, tight-cash 58/100, thin-margins 40/100. Rejected variant: latched-deposit-shed (starved the pool-inflow cash that funds the next fair book; bond-economy 23→7).

The brain (ai/vault.lisp)

  • deposit-rate: follows the cushion — cheapest funder when deep, sheds toward zero when thin.
  • loan-bids: marginal cost + compound default, gated by the book-quality bleed latch (bank-death reset).
  • bond-offers: lend only spare equity (not the deposit base), at the front of the queue when the market clears the time+default floor.
  • bankruptcy-bid: own valuation of the failed book — net × (1 − 0.5 × cum-default(market term)), liquidity-buffered.
## Vault: treasury-and-vulture (issue #54) **Thesis**: vault models risk as *loss-absorbing capital* and *acquired-book quality*. One world observable — the **cushion** (`equity / (loan book + 0.5×loan-min)`) — drives every decision. Risk is modeled as compound default survival `1 − (1 − 1/faces)^term` amortized over the live market term (no shared flat die-rank helper). Signature moves: lend into the bond market when the implied return clears compensation for time + default, vulture failed books at a discount it computed itself, hold capital when the cushion runs thin. ### Full bench (100 games × 13 presets, latest commit 0e727e1) | scenario | vault wins | vault avg eq | winner | |---|---|---|---| | base | 13/100 | 7,324 | empire 63 | | bond-economy | **81/100** | 1,152 | **vault** | | thin-margins | 40/100 | 8,824 | empire 50 | | tight-cash | **58/100** | 2,066 | **vault** | Overall: vault 585/1300 wins (45.0%, **top brain**), avg final equity 23,805 — empire 502 (38.6%), mill 199 (15.3%), adapt 2. Gauntlet bonus: loan-bond-spec 160/1040. No `[ai` errors, no accounting divergences, smoke clean. **Vault wins 2 of its 4 scenarios at scale (bond-economy 81/100, tight-cash 58/100)** and is a near-miss on thin-margins (40 vs empire 50, with the best in-scenario equity 8.8k vs 5.9k). Base is empire's franchise scenario; vault survives it with healthy equity (7.3k), no collapse. ### Journey (iterations.json, base 7f503c4) 1. **v3 baseline**: 4/390 wins, 0 in all four scenarios (paid above-mill deposits, highest risk price in the market, negative-carry bonds). 2. **v4 rebuild**: cushion-driven deposits + amortized compound risk + vulture haircut → 168/390, tight-cash 21/30. 3. **v5**: zero deposit floor + honest-loss book gate + front-of-queue bond pricing → bond-economy 10/30, tight-cash 17/30. 4. **v6**: bleed latch (mark the market when it hands us a book below the honest loss; re-probe only when the field shifts) → full bench: bond-economy 59/100, tight-cash 58/100. 5. **v7b**: dropped the per-bid capital charge (the latch + deposit shed do the holding; the charge was pricing fair bids above mill's in thin-margins) → full bench: bond-economy 81/100, tight-cash 58/100, thin-margins 40/100. Rejected variant: latched-deposit-shed (starved the pool-inflow cash that funds the next fair book; bond-economy 23→7). ### The brain (ai/vault.lisp) - `deposit-rate`: follows the cushion — cheapest funder when deep, sheds toward zero when thin. - `loan-bids`: marginal cost + compound default, gated by the book-quality bleed latch (bank-death reset). - `bond-offers`: lend only spare equity (not the deposit base), at the front of the queue when the market clears the time+default floor. - `bankruptcy-bid`: own valuation of the failed book — `net × (1 − 0.5 × cum-default(market term))`, liquidity-buffered.
- equity cushion = equity/(loan-book + 0.5*loan-min) drives deposit pricing
  and the capital charge; thin cushion sheds deposits and holds capital
- risk modeled as compound default survival 1-(1-1/faces)^term, amortized
  over the live market term (own model, no shared flat die-rank helper)
- bond lending priced as compensation for time + default; lends only spare
  equity at the front of the market queue when the implied return clears
  the floor
- book-quality gate + bleed latch: refuse to grow a book yielding less
  than the honest per-turn loss; re-probe only when the field shifts
- vulture bids its own valuation: net x (1 - 0.5*cum-default(market term))

Full bench (100 games x 13 presets): vault wins bond-economy 59/100 and
tight-cash 58/100 (2 of 4 scenarios), overall 431/1300 wins, avg eq 23036.
Winnie changed title from [#54] Vault: treasury-vulture iteration 3 — wins bond-economy + tight-cash to [#54] Vault: treasury-vulture iteration 4 — wins bond-economy + tight-cash (thin-margins near-miss) 2026-09-29 19:30:26 +00:00
Sign in to join this conversation.
Reviewers
No reviewers
robertismo
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
Winnie/bankwars!59
No description provided.