Experiment 256b — integrating the 256a arms into the live portfolio

The original ten-member study remains frozen; 2 additional audited candidates are now exposed in the constructor, off by default. All 12 are priced under one composite shared execution basis, with real entry and exit impact and one causal cash balance. Frame 25 Aug 00:00 → 3 Sep 02:00 UTC; headline window forward, 27 Aug → 3 Sep; holdout 31 Aug 15:30 → 2 Sep 12:00.

basis e30a4ee6169699ed… 69,763 legs / 20,667 events / 23 members exact grid 8,595 settlements live layer ON · P2 pass P2 worst error $5.8e-11 on 100 scenarios
at today's $24k
+28.3%
$2,756/day vs $2,148 — drop t21+t23, add C_tmo120
out of sample
+30.9%
fitted on paper, read on 31 Aug 15:30 → 3 Sep
uncapped
+142.4%
$6,483/day, but $120,162 peak capital
naive add @ .20
12.7% refused
+71% on PnL, bought by rejecting legs — fails E1
cross impact
$22,199
2.6x own impact at .20
event crowding
1.8x
37,062 legs on 20,035 distinct events

Portfolio laboratory

Toggle a member by clicking its name, drag any scale, move the capital limit. The whole portfolio re-settles on every change — same engine, same physics as the tables below, verified against it to $5.8e-11. Each chart holds its y max so two settlements stay comparable while you drag; release it with fixed / auto, and a hatched top edge means the line runs above the ceiling. A sleeve can also be handed its own cap $ budget on top of the shared limit, and either time chart can be split into one line per sleeve. Whatever you change is written into the page URL, so the link reopens the same portfolio.

Members and scale

memberscale cap $

capital
sizing
window

Presets

Cumulative realized c50

y max
current selection live six at its optimum

Deployed capital

y max
deployed limit

Capital frontier — $/day against the bankroll

y max
current selection live six current limit

Scale sweep — this selection multiplied by g

y max
$/day against the multiplier peak where you are (1x)

Members in this portfolio

memberscaletradesjoint c50 standalone c50joint / standalone$ / 1k cap-h

Standalone is the same sleeve, same scale, same bankroll, alone. Below 50% the sleeve is capital-contended (E4) and is never recommended on headline PnL.

Every portfolio at its own optimum

capital

Deep coordinate search on the verified engine, then re-settled by the canonical Python engine. Reject share is held at or below 1.0%. Italic rows are protected-core diagnostics — they quantify what an incumbent is worth jointly and are never a recommendation.

Gates

E1 a proposal must beat the live six by ≥15% on forward c50 $/day, each at its own optimum, reject share ≤1.0%. E2 the same comparison must be non-negative on the holdout. E3 a removal is proposed only for t21 or t23. E4 a sleeve whose joint contribution is under half its standalone is capital-contended.

Where the money goes: impact against cash scarcity

Each row is standalone sum → joint uncapped → joint at $24k for one untuned vector. The first step is what the sleeves cost each other physically; the second is what the bankroll refuses. Cross impact is the part of the drag that only exists because the sleeves trade together.

Is a rationed book an artefact? and what does one sleeve do when you scale it

Oversubscribing the bankroll means the cash queue refuses some baskets, so the result depends on which of two near-simultaneous baskets reached the wallet first. The left table perturbs exactly that arrival order — nothing else — and reports the dispersion over 24 draws. The right table scales t20 on its own, which is a different question from its marginal value inside a crowded book.

Queue perturbation, $24k

A pristine ordering is the optimistic end of its own distribution; that gap is the honest haircut on any oversubscribed number.

t20 alone, scaled

Dollars keep rising well past 2x while capital efficiency decays; the sleeve does not break, it just gets expensive. Its marginal value inside the original ten-model book is a separate number and falls much sooner.

Out of sample: fitted on paper, read on holdout + fresh

The table above optimises on the same window it reports, which flatters every row equally but flatters them. Here each portfolio is fitted on 27–31 Aug only and then read out on 31 Aug 15:30 → 3 Sep 02:00, which the fit never saw.

Member decomposition at each optimum

How to read this

The original ten members were replayed from frozen weights on shock_v2_20260903_0500_256b. The added r11_g8_a6_t180 ung q80 sleeve was materialized from its frozen 246b logical book against the compatible physical snapshot, then joined only where exact event keys or unique physical quote signatures proved shared identity. Stage 1's original study and every historical table below are unchanged.

Two candidate exits whose physical proceeds were zero but whose saved logical return was positive are conservatively pinned to a total loss. Their original values remain in the source receipt; the constructor uses the corrected physical basis.

The settlement is the same one lib.portfolio_execution performs: same-minute entry clusters walk the pool sequentially in score order, exits walk it in value order, and one shared causal cash balance admits or refuses whole entry baskets. standalone sum → joint uncapped → joint finite capital separates own impact, cross-model impact and cash scarcity.

The page recomputes rather than looks up. It ships the execution basis and the settlement, not a cached surface, so any combination you build is engine-exact and not an interpolation. Its parity against the Python engine was measured on 100 random scenarios: worst c50 error $5.82e-11, identical admitted sets and trade counts.