Papers, memos, and long-form notes.

A 31-gate evaluation of which cost components (spread, slippage, funding, borrow, exchange fees) are load-bearing for a sub-$250k systematic book versus which inflate the gate count without changing the pass/fail rate. The thesis: the first six gates are the only ones that move the decision for a book of this size.

Replay eight reference strategies from /projects/quant with a calibrated cost model (project 08). Compare pass/fail under three cost scenarios: zero-cost, fee-only, and fee-plus-half-spread. Report the gate-by-gate delta and rank gates by marginal information.

Six cost components (spread, slippage, funding, borrow, exchange fees, latency penalty) account for ~92% of the cost-induced pass/fail movement. The other six cost gates are stable across scenarios and can be moved to a second-pass audit without losing decision quality.

Walk-forward validation leaks information through the parameter grid; embargoed OOS is statistically clean but discards data. The memo proposes a hybrid: a walk-forward stage that screens the gate stack, followed by an embargoed OOS stage that scores the survivors. The hybrid is cheaper to run and harder to game than either alone.

Replay the deflated-Sharpe sweep (project 01) under three protocols. walk-forward only, embargoed OOS only, and the proposed hybrid. Compare false-positive rates, time-to-decision, and reproducibility across cold starts. The reproducibility test is the load-bearing one.

Hybrid reduces false-positive rate from ~24% (walk-forward alone) to ~6% (hybrid) and only adds ~15% to time-to-decision. Embargoed-only is cleaner statistically but doubles the data requirement; the hybrid is the right default for a small book.

A block-bootstrap null with the same autocorrelation as the source is the cheapest defensible null hypothesis for a systematic strategy. The memo standardizes the block-length choice (Andrews 1991, Politis-Romano), the burn-in rule, and the rank-1 / rank-N reporting convention used across /projects/quant.

Run the standardized bootstrap-null routine on the 9 /projects/quant strategies. Report the rank-1 strategy's z-score against the null for each. Strategies above +1.65σ are flagged for embargoed OOS; strategies below are tabled.

Three of the 9 strategies clear the +1.65σ threshold (deflated-Sharpe sweep, funding-carry, pairs-cointegration). Six are within ±1σ of the null mean. The bootstrap-null step is the single most informative 20 lines of code in the /projects/quant repo.

The same gate stack that defends a quant book (mechanical-validity, OOS discipline, multiple-testing correction) defends a multi-agent LLM system. The memo maps the 31 quant gates to the 31 AI gates used by the orchestrator and shows that the failure modes are isomorphic: look-ahead in time is look-ahead in the conversation history.

Replay the /solutions/01 to 11 agent logs through the 31-gate evaluation harness. Compare false-positive rate, time-to-decision, and gate-by-gate attribution. Report the per-gate failure-mode mapping between the quant and AI stacks.

22 of 31 gates map 1:1 across stacks. The 9 non-trivial mappings are precisely the gates that the 11-agent platform was designed to enforce mechanically. they are the ones that fail closed in production and pass in tests.