PORTFOLIO BACKTEST · R4 · PAPER MONEY2026-09-18 · 5 levels A–E · dailyreproducible: node research/backtest.mjs

Five levels of aggression, one temperature.

Every day the JEB score sets how aggressive the book may be — A is all cash, E is fully allocated — via jev's typed decision contract. Execution follows one day later at the closing price with explicit costs, the way a real executor would. Nine years, walk-forward, versus buy-and-hold.

P2 · banding + confirmation7,578%

total return vs buy-&-hold. Exposure 50% average, 344 level changes in ~9 years.

Max drawdown56%

versus crypto's full drawdowns — the level system's job is to be in A when everything is on fire.

Sharpe (daily, rf=0)1.3

risk-adjusted. P3 (regime-gated) trades return for smaller drawdowns.

Suite reliability6/6

engine tests passed: determinism, known-answer, invariants, no look-ahead, cost monotonicity.

Equity — strategies vs buy & hold

buy & hold (log)P2 banding+confirmP3 regime-gatedlevel A–E timeline
top: JEB temperature 0–100 (amber)bottom band: daily level — A cash (grey) → E full (red)

Performance detail

StrategyTotalCAGRMaxDDSharpeAvg exposureChanges
P1 · banding6954%55%54%1.2850%568
P2 · banding + 2-day confirm7578%56.4%56%1.350%344
P3 · regime-gated7137%55.4%47%1.3443%422
buy & hold7,560%—~−83%—100%1

Cost sensitivity (P2)

Cost / slippage0 bps10 bps30 bps60 bps
Total return8,280%7,578%6,344%4,855%

The thesis survives realistic execution costs. Level changes are rare by construction (band edges + 2-day confirmation), which is what keeps the strategy inside its edge after costs.

Backtest-suite reliability

TestResult
determinism — identical engine output on re-runPASS
known answer — 100% BTC, 0 cost = raw price seriesPASS
invariants — levels ∈ 0–4, equity > 0, all runsPASS
no look-ahead — decisions use info ≤ t, applied to t→t+1 return onlyPASS
exposure ladder — APASS
cost asymmetry — higher costs never increase returnsPASS

The known-answer test feeds a strategy that is always fully allocated and checks the engine reproduces the raw price series to machine precision. The no-look-ahead property is structural: a decision taken at close t is only ever applied to the t→t+1 return, so yesterday's score drives today's exposure — never today's.

Who does what — jev vs GLM

jev — the decision layer (typed, gated). Each run produces a typed decision {from, to, direction: upgrade|downgrade|hold, probabilities} derived from the temperature bands and the regime gate. In this backtest the policy stands in for jev (historical jev calls are impossible); live, jev makes the same call from richer unstructured input — news tone, Reddit capitulation language, divergence causes — inside the same typed contract and the same 2-day confirmation, so a hallucinated upgrade cannot pass the schema, and the probability gate can refuse low-confidence moves.

GLM 5.3 MAX — the execution layer (words and orchestration, not math). It translates an approved level change into orders (size the delta, respect venue limits, split if shallow books), writes the fact-bound blog line about the change, handles incidents, and logs the run. It never chooses the level — the same rule that makes the index honest makes the portfolio honest: agents orchestrate, typed code computes.

Limitations

Long-only spot simulation on daily closes; intraday fills, funding, and borrowing are out of scope. The band thresholds (30/45/55/70) come from the R1 sweep's temperature bands, not from optimizing this portfolio — but they were chosen knowing the data, so treat relative (not absolute) performance as the finding. P(bull) is walk-forward refit yearly; 2021+ only. Costs are linear approximations. Past regime structure may not repeat; the level system's value is drawdown control, and that is the claim this data supports.