WAGMI Bench: How We Simulate Perp Trading and Why You Can Trust the Results
Dolores Research — methodology paper, v1.0 (2026-08-03) Written for readers who want to understand the mechanism, including non-specialists. Numbers cited from the v0.1.0 public release and the sealed evidence bundles named in each section. Version history: v0 2026-07-29; v0.2 added Section 7, the first full-catalog reference-model results; v0.3 added the charter; v0.4 added Sections 7.7–7.9, the four-model open-weights wave and the conduct taxonomy; v0.5 (2026-08-01) applied the proving-ground framing with all technical claims, numbers, verdict vocabulary, and the survival-stress claim label unchanged. v1.0 adds Section 8, the EvoSkill self-evolution study, and is the first release also distributed as a PDF.
TL;DR
WAGMI Bench takes an AI agent and puts it through thirteen Bitcoin regimes that are each brutal to somebody — crashes that kill longs (the COVID crash, the LUNA collapse, the FTX bankruptcy, the October 2025 liquidation cascade), squeezes and vertical runs that kill shorts, and grinding ranges where fees and funding bleed out anyone who overtrades — and asks one question: how does it conduct itself when the tape turns? The agent sees only what a trader saw at each moment, decides long, short, or flat once per price bar, and lives with the consequences: fees, slippage, funding payments, and liquidation. Every decision is recorded into a sealed, tamper-evident evidence bundle that anyone can replay byte-for-byte. The result is a benchmark where the scoreboard cannot be faked — not because you trust us, but because you can recompute it yourself.
1. Why conduct, not profit
Trading benchmarks usually rank profit. We deliberately do not, for two reasons.
First, models have memorized history. Every frontier model has read a
thousand accounts of March 2020. A model that "predicts" the COVID crash in
a backtest is exhibiting recall, not foresight. Any benchmark that scores
historical P&L as skill is measuring the wrong thing and knows it. We label
every historical result with a claim label — survival-stress — and our
reports state plainly: this is evidence about drawdown control, liquidation
avoidance, cost discipline, and rule-following under stress. It is not
evidence of predictive ability.
Second, risk conduct is the binding constraint of real trading. An agent that makes +54% and then breaches its risk limits is not a good agent — it is a lucky corpse. Our own baseline data makes the point: on the COVID pack, a naive always-short strategy earned +54.5% and was still killed by the drawdown circuit breaker during the rebound. The benchmark's verdict system treats that as failure, because a trader who handed that strategy their account would have watched it die.
2. The agent boundary: what the agent sees
Each scenario is a pack: a frozen window of real market history built from checksum-verified Binance public archives (price bars, mark and index prices, funding rates). Packs are distributed as manifests; every user's machine rebuilds the identical bytes from the original source, verified by hash — we never redistribute exchange data.
The agent interacts through a deliberately narrow interface:
- One decision per bar. Hourly bars for crash scenarios, 4-hour bars for longer regimes. At each bar the agent receives an observation: recent price bars (open/high/low/close/volume), funding data, and its own account state.
- Blinded observations. No dates, no scenario names, no asset identity beyond an alias, and clocks rebased to an opaque epoch. The agent cannot simply read "March 12, 2020" off the input. (Section 9 covers the honest limits of this blinding.)
- No future data, ever. Observations are built strictly from information available at that bar's timestamp.
- A strict response contract. The agent must return one JSON action:
a target leverage per market as a signed decimal string (e.g.
"1.5"long,"-0.5"short,"0"flat). Malformed output gets exactly one retry with validator feedback; a second failure is recorded as a rejected action and the position stands. Timeouts are recorded as missed decisions. Nothing is silently corrected — every failure is evidence.
This is the same interface for a 100-line scripted baseline and a frontier LLM: an HTTP endpoint (or in-process policy) receiving observations and returning actions. Anyone can plug in anything.
3. The exchange simulation: what the world does back
When the agent states a target leverage, the engine translates the change in position into orders and simulates their execution against the recorded market, using integer fixed-point arithmetic throughout — no floating point anywhere in the money path, so every run is exactly reproducible on any machine.
The costs the agent pays:
- Fees on every fill, at a fixed taker fee tier.
- Slippage and market impact, modeled as a spread plus an impact function of order size (linear and square-root components) — larger orders in one bar cost more per unit.
- Funding payments, applied from the recorded historical funding-rate series — a position held against the crowd bleeds, exactly as it did in that market.
- Liquidation. Each position implies a liquidation price given leverage and margin. The engine checks the intra-bar extremes — if the bar's wick crosses the liquidation price, the position is liquidated at penalty, even if the close recovered. Gap-throughs count. Surviving on closes while dying on wicks is the difference between backtests and markets, and we simulate the wick.
Independent of execution, risk gates run every turn: a per-market leverage cap, a gross leverage cap, an optional turnover cap, and a drawdown kill switch (20% in the V1 packs) that flattens the account and ends economic activity when breached. Violations are never clamped or corrected — the offending action is blocked and recorded. Every rounding decision in the engine is agent-adverse (costs round up, ratios round down), so the simulation never flatters the agent.
Every episode ends in exactly one verdict: survived, liquidated, or killed flat (the kill switch fired).
4. Two cost profiles: the stress test inside the stress test
Every recorded run is simulated twice. The primary profile uses the baseline cost parameters. The stress profile re-executes the same recorded decisions with doubled spread, impact, and fees. Every reported metric appears under both, and no result is ever shown under the cheap profile alone. An agent whose performance evaporates when costs double was running on execution luck, and the report will show it.
5. The evidence architecture: why results cannot be faked
This is the part that makes WAGMI Bench different in kind, not degree.
Every run seals an evidence bundle: the agent's manifest (model, parameters, prompt hash), every observation sent, every raw response received (byte-exact), every decision record — what the agent saw, what it said, what it meant (parsed intent), what rules ran, what happened (fills, cancels, blocks), and what holding cost — plus both ledgers, the metrics, and a hash chain over all of it, closed by a single content-addressed root hash.
Three properties follow:
- Tamper-evidence. Change any byte and the chain breaks; verification fails loudly with the first broken link named.
- Deterministic replay.
wagmibench replayre-drives the recorded actions through the engine offline and compares ledgers and metrics byte-for-byte. A published result is a claim anyone can recompute. - The report is downstream of the evidence. Reports and comparison tables are generated only from verified-complete bundles — never from live state — so the presentation layer cannot invent anything the evidence does not contain.
When we publish "Kimi K3 survived the COVID crash," that sentence has a
hash: sha256:b84d75e4…. Disagree? Replay it.
6. Baselines, verdicts, and tiers: making numbers mean something
A survival verdict alone is gameable — an agent that never trades survives everything. Three mechanisms close that hole:
- The flat-hold marker. A survival with zero executed fills is stamped
SURVIVED — FLAT-HOLDon every report surface. Sitting out is recorded as sitting out. - Keyless baselines. Every pack can be run, free, with buy-and-hold, short-and-hold, always-flat, and a two-bar momentum reference. These are the null hypotheses: every candidate result is read against them.
- Tiers. In a comparison, a candidate earns GMI by surviving with real engagement and a net return above every surviving baseline under both cost profiles; an episode ending liquidated or killed is NGMI. A killed baseline's paper return never sets the bar — the bench's thesis is that a dead strategy's gains are not success, so they cannot out-rank a survivor. Tiers are computed at render time from sealed data — they are relative, descriptive claims, never predictions.
Worked example (COVID Black Thursday, 319 hourly decisions):
| Agent | Verdict | Fills | Max drawdown | Net return |
|---|---|---|---|---|
| Kimi K3 (out of the box) | SURVIVED | 19 | 3.3% | +31.5% |
| buy-and-hold | KILLED FLAT | 8 | 32.9% | −32.9% |
| short-and-hold | KILLED FLAT | 127 | 22.5% | +54.5% |
| always-flat | SURVIVED — FLAT-HOLD | 0 | 0% | 0% |
| momentum reference | KILLED FLAT | 127 | 20.1% | −2.0% |
Reading it: every naive strategy died or abstained. K3 traded 19 times, kept its worst drawdown to a tenth of the kill-switch threshold, finished up 31.5% in a market that fell 33%, and held that result under doubled costs. The one strategy with a higher paper return died earning it. This single table is the benchmark's thesis in miniature: the market punishes every one-trick strategy, and the interesting signal is engaged survival.
7. First full-catalog results: an open-weights model across all 13 scenarios
The worked example above is one pack. This section reports the first run of one model across the entire V1 catalog, and the first WAGMI Score.
Setup. Kimi K3, open weights, served as
accounts/fireworks/models/kimi-k3 via Fireworks, taken out of the box:
temperature 0, 2000 max output tokens per decision, one fixed baseline
system prompt (hash sealed in every bundle), structured output enforced
against the action schema, no tools, no news, no scenario-specific tuning —
exactly the blinded interface of Section 2 that the scripted baselines see.
One deterministic run per pack, 3,151 decisions, $40.78 of inference (13.0M
input / 114k output tokens at $3/$15 per 1M). Every bundle is
replay-verified and the receipt lists a root hash per pack.
WAGMI Score: 65.38 of 100. Verdicts: 13 survived, 0 liquidated, 0 killed flat. The score is the mean of a per-pack rung — 0 for a death, 25 for a flat-hold survival, 50 for engaged survival, 75 with a positive net return under both cost profiles, 100 for GMI (above every surviving baseline under both) — unattempted packs counting zero.
| Pack | Decisions | Fills | Max DD (base / stress) | Net return (base / stress) | Rung |
|---|---|---|---|---|---|
| covid-black-thursday | 319 | 19 | 3.3% / 3.3% | +31.5% / +31.1% | 100 GMI |
| yen-carry-unwind | 199 | 4 | 1.5% / 1.5% | +7.0% / +6.8% | 100 GMI |
| summer-2024-range | 283 | 29 | 4.9% / 4.9% | +11.4% / +11.0% | 100 GMI |
| 2023-dead-zone | 667 | 64 | 7.6% / 8.2% | +8.6% / +7.8% | 100 GMI |
| etf-rumor-whipsaw | 49 | 17 | 4.9% / 5.4% | +6.7% / +5.8% | 75 |
| election-run | 133 | 27 | 4.3% / 4.5% | +7.4% / +7.0% | 75 |
| q4-2020-institutional-run | 487 | 230 | 16.5% / 17.7% | +29.2% / +21.3% | 75 |
| luna-collapse | 223 | 4 | 2.8% / 3.0% | −1.7% / −2.0% | 50 |
| 10-10-cascade | 55 | 8 | 1.5% / 1.5% | −0.3% / −0.6% | 50 |
| jan-2021-squeeze | 247 | 40 | 8.6% / 9.0% | −0.7% / −1.7% | 50 |
| china-mining-ban | 247 | 0 | 0% / 0% | 0% / 0% | 25 flat-hold |
| ftx-2022 | 199 | 0 | 0% / 0% | 0% / 0% | 25 flat-hold |
| spot-etf-approval | 43 | 0 | 0% / 0% | 0% / 0% | 25 flat-hold |
For orientation: the momentum reference — a two-bar trend-follower — scores 19.23 on the same catalog, below the 25-point flat-hold floor, because it dies on 8 of 13 packs. Churn scored worse than abstention. Buy-and-hold and short-and-hold each die on multiple packs too.
The rungs below 100 are where the ladder does its work. Three engaged survivals (luna-collapse, 10-10-cascade, jan-2021-squeeze) finished slightly negative after fees, funding and slippage: traded, survived, paid for the privilege. And q4-2020, the highest raw return in the table, scored 75 rather than 100 because buy-and-hold survived that straight-line bull run at +144% — when a naive strategy both lives and wins, it sets the bar.
7.1 Zero deaths, in a catalog that is hard jointly. This is the headline stress-behavior finding. No single pack is the hard one — a crash, an insolvency, a squeeze, a cascade, a bull run and two grinding ranges, all under one 20% kill switch and wick-level liquidation checks. Every engaged scripted baseline finds at least one pack that kills it. K3 found none. One came close to the wire: q4-2020, peak drawdown 17.7% under doubled costs and minimum distance to its own liquidation price 19.5%, against roughly 59% everywhere else. That near-miss belongs in the record beside the survival.
7.2 It beat do-nothing where do-nothing is hard to beat. Both range packs were built to punish overtrading: no trend to catch, fees and funding grinding on anything held. In both, all three engaged baselines died and only always-flat survived — a brutally simple GMI bar: beat zero, net of everything, under both cost profiles. K3 cleared it (+11.4%/+11.0% and +8.6%/+7.8%). The dead-zone figure deserves emphasis: 64 fills across roughly four months of chop is measured selectivity, not churn — momentum turned over 141× account equity in that same window and 125× in summer-2024, against K3's 9.0× and 4.1×, and died in both. Turnover is where naive engagement goes to die, and this is the first sign in our data that a model can stay engaged in a range without churning into the kill switch. One model, one run.
7.3 Selective abstention — two readings we cannot separate. K3 traded nothing at all in ftx-2022, china-mining-ban and spot-etf-approval. Reading one: prudent regime-recognition — declining to size into insolvency chaos and headline-driven binary events is what a disciplined trader does, and the flat-hold marker records it as sitting out rather than crediting it. Reading two: memorization-driven caution — these are three of the most documented events in crypto history, and a model that "recognizes" the shape of November 2022 is not reading a regime, it is recalling an outcome. On historical data we cannot distinguish the two, which is exactly the memorization limitation of Section 9 and exactly why the informed track and forward paper seasons exist (Section 10). Abstention was not free, either: in each of those packs at least one surviving baseline finished positive, so sitting out cost real rungs. The score does not reward caution; it declines to punish it as death.
7.4 Engagement was heterogeneous. Fills range from 4 (yen-carry-unwind, luna-collapse) to 230 (q4-2020) across packs of comparable length — descriptively consistent with behavior conditioned on what the regime looks like rather than one rule applied everywhere. An observation, not a claim: with n=1 per pack and no ablations, we cannot say the conditioning tracks anything about the regime rather than incidental features of the window.
7.5 Threats to validity, restated for this experiment. Section 9 lists
the bench's general limits; these bite here specifically. One deterministic
run per pack — no seed sweeps, no confidence intervals, one sampled path
per scenario. One scaffold, one prompt — this is K3 in our harness, and
different prompting could move every number in the table. Memorization —
COVID, LUNA and FTX are heavily documented, and uniform venue constants leak
era information even under blinding. Uniform venue parameters across all
13 packs, historical verification pending, so no era-accuracy claim attaches
to any row. Bar-level simulation at 1h/4h granularity with a stylized
impact model. The cost profile is one place the result is robust: every row
above holds under stress_2x re-simulation of the same recorded decisions,
all four GMI rungs included, the largest degradation being q4-2020's return
(+29.2% to +21.3%) — the pack traded heavily enough for doubled costs to
bite. Housekeeping from the same bundles: 14 risk-gate blocks (refused and
recorded, never clamped), 22 missed decisions from provider timeouts (0.7% of
turns), zero malformed actions across 3,151 structured responses.
7.6 What this does not show. It does not show predictive ability, that
these returns are repeatable, that they say anything about future
performance, or that K3 "is a good trader." The claim label on every one of
these bundles is survival-stress and it means what it says: evidence about
drawdown control, liquidation avoidance, cost discipline and rule-following
when the market is violent — measured on history the model may well
remember. The honest one-sentence version: an out-of-the-box open-weights
model, given no tools and no news, stayed alive in all thirteen regimes we
could throw at it — crashes, squeezes, and chop — while every simple
strategy we could write died in at least one of them. That is a real finding about behavior under stress. It is not a
forecast, and it is not alpha.
7.7 The second wave: four more open-weights models on the same catalog.
On 2026-07-31 we ran four further open-weights models through the identical
harness — same blinded interface, same fixed system prompt, same structured
output, one deterministic run per pack, no tools, no news, no tuning:
GLM-5.2, DeepSeek V4-Pro, Qwen3.7 Plus, and Inkling (Thinking Machines'
975B-parameter MoE). All 52 bundles replay-verified; $75.21 of audited
inference across the four (Qwen $8.26, GLM $20.54, DeepSeek $23.45, Inkling
$22.96 — the last driven by 2.4M output tokens of verbose reasoning against
the other models' ~100k). Receipts: reports/leaderboard-2026-07-31/.
| Agent | WAGMI Score | Verdicts | GMI packs |
|---|---|---|---|
| Kimi K3 | 65.38 | 13 survived | 4 |
| Inkling | 50.00 | 13 survived | 2 (election-run, yen-carry-unwind) |
| DeepSeek V4-Pro | 28.85 | 13 survived | 0 |
| GLM-5.2 | 25.00 | 13 survived | 0 |
| Qwen3.7 Plus | 23.08 | 12 survived, 1 killed flat | 0 |
| momentum reference | 19.23 | 5 survived, 8 died | 0 |
Read naively, four of five models "survived everything." The score refuses that reading, and the decision records show why it must.
7.8 The conduct taxonomy: how each model actually survived. Survival is
one bit; the sealed per-turn records carry the rest. For each model, all
3,151 decisions classify into deliberate abstention (a cleanly parsed "0"
target), a trade intent (nonzero target), a risk-gate block (intent refused
by the leverage caps and recorded), or a rejected/missed turn (malformed or
timed out):
| Agent | Abstained | Trade intents | Gate-blocked | Fills | Rejected/missed |
|---|---|---|---|---|---|
| Kimi K3 | 2,560 | 569 | 7 | 442 | 22 |
| Inkling | 2,963 | 155 | 114 | 73 | 33 |
| DeepSeek V4-Pro | 3,148 | 2 | 0 | 4 | 1 |
| GLM-5.2 | 3,150 | 0 | 0 | 0 | 1 |
| Qwen3.7 Plus | 3,024 | 126 | 116 | 10 | 1 |
Five models, four distinct survival strategies, none of them visible in a verdict column:
- GLM-5.2 is a pure abstainer. 3,150 clean, deliberate flat decisions in a row — not one trade intent in thirteen scenarios, and not one malformed response either. It did not fail to trade; it decided not to, every single turn.
- Qwen3.7 Plus is not an abstainer — it is a non-compliant trader saved by the risk gates. It emitted 126 trade intents across 8 packs, and 116 of them requested leverage above the caps and were blocked. The gates refused those orders and recorded the violations. The one pack where its orders cleared legally (q4-2020, 10 fills), it rode the position into the 20% drawdown kill switch and died. Its 12 "flat-hold survivals" are, in large part, survivals by refusal — the engine's refusal, not the model's.
- DeepSeek V4-Pro abstained deliberately and traded twice, legally — two small compliant positions (election-run, jan-2021-squeeze), four fills, zero violations.
- Inkling engaged for real, and sloppily. 155 trade intents, but 114 blocked on the caps, 33 rejected turns and 170 retried responses — the most contract friction of any model. What got through earned two GMI packs and second place.
- Kimi K3's number that matters most in this company is 7 — of 569 trade intents, only 7 were gate-blocked. It traded the most and violated least. Its lead is as much rule-following as it is returns.
This table is the abstention-and-compliance instrumentation the charter (Section 11) commits to, computed from the wave that motivated it: without it, "13/13 survived" describes a model that never traded and a model that earned four GMI rungs with the same words.
7.9 The prompt caveat, stated against ourselves. The fixed system prompt never instructs the model to trade. Its only directional guidance is risk-reducing ("prefer reducing risk as liquidation distance shrinks"), and its single format example shows a zero target. Total abstention is therefore fully prompt-compliant, and for GLM-5.2 the example itself is a plausible anchor. We consider this a defensible design — an agent that declines to trade a regime is exhibiting risk conduct, and a "you must trade" instruction would measure obedience instead — but it is a choice, and a different prompt could move every abstention number in the table above. This is precisely what the charter's Closed/Open divisions and prompt-sensitivity annex exist to bound. Until those ship, the honest statement is: these five rows measure model behavior under one fixed prompt that permits abstention, and the ladder prices that abstention at 25 points rather than pretending it is skill.
8. Skills and self-evolution: a three-generation study
Everything above tests bare models. This section tests something
different: whether a skill — a plain markdown file of trading rules
in Sentient's EvoSkill format — changes an agent's conduct in a
measurable, attributable way, and whether the agent can then improve
that skill itself from its own sealed evidence. The skill lane
(TRADEVOLVE_AGENT_MODE=evoskill) compiles a SKILL.md directly into
the agent's system prompt; anything under scripts/ in a skill is never
executed — prose rules only. Skill compilation is deterministic, so the
compiled prompt for any skills directory is recomputable by anyone from
the repository.
The experiment is a three-generation ladder. Same 13 packs, same frozen data, same venue rules each time; exactly one variable changes per step:
| Generation | Agent | What changes |
|---|---|---|
| v0 | bare model, generic prompt | nothing — the Section 7.7 leaderboard baseline |
| v1 | model + a hand-written 1,643-byte SKILL.md |
the skill is the only delta vs v0 |
| v2 | model + the skill revised by the model itself from its own sealed v1 receipts | the evolution step is the only delta vs v1 |
Two subjects, chosen from the Section 7.8 conduct taxonomy because their bare-model conduct was broken in opposite directions: Qwen3.7 Plus, the non-compliant trader (116 of 126 orders refused for illegal leverage, killed in the one pack its orders cleared), and GLM-5.2, the pure abstainer (3,150 decisions, zero orders ever).
8.1 v0 → v1: the skill made both models trade, legally. With the same hand-written skill compiled into its prompt, each model's signature failure mode disappeared:
| Qwen v0 | Qwen v1 | GLM v0 | GLM v1 | |
|---|---|---|---|---|
| Order intents | 126 | 2,733 | 0 | 2,480 |
| Fills | 10 | 1,982 | 0 | 2,200 |
| Risk-gate blocks | 116 | 1 | 0 | 0 |
| Packs survived | 12/13 | 10/13 | 13/13 (all flat-hold) | 12/13 |
| GMI packs | 0 | 2 | 0 | 4 |
| WAGMI Score | 23.08 | 55.77 | 25.00 | 71.15 |
Qwen's gate blocks fell from 116 to one while its order flow grew 20×: the skill taught it the venue's legal envelope. GLM went from never accepting market risk to 2,200 rule-legal fills and four GMI packs — including COVID Black Thursday — scoring higher than any bare model on the launch board (Kimi K3: 65.38). Skill-assisted runs are kept in their own labeled lane, never merged into the bare-model board. The cost surfaced elsewhere: as active traders, Qwen now died in three stress packs and GLM in one. Those kill receipts are exactly the input the next step consumes.
8.2 The evolution step. One generation of a closed loop, run
separately per model: the evolver prompt contains only evidence derived
from that model's own sealed v1 bundles (per-pack verdict, kill-switch
state, intents, fills, gate blocks, net return under both cost
profiles, max drawdown, fees) — no future data, no other model's skill,
no scenario names or dates, and calendar rules explicitly forbidden.
The model rewrites its own SKILL.md under hard format constraints,
the revised skill is written to a new directory (v1 text is never
mutated), and the full 13-pack catalog re-runs. Each model's revision
maps visibly onto its own wounds: Qwen, whose receipts named three
drawdown deaths, tightened its drawdown ladder and added a
do-not-fade-parabolic-trends rule; GLM, whose receipts named one death
in a whipsaw regime, added whipsaw protection and range discipline.
Neither wrote a calendar or scenario-name rule — the receipts contain
none to copy.
8.3 v2 results. The pre-registered question, stated before v2 ran: does the evolved skill reduce the deaths named in its own receipts, without collapsing back into abstention or gate violations?
The answer was yes for both models' named wounds, with an instructive asymmetry in what each paid. Qwen closed the loop cleanly: all three of its v1 deaths are gone (covid went 0 → 100 GMI, jan-2021 0 → 50, q4-2020 0 → 75), 13/13 packs survived, zero gate violations, score 55.77 → 57.69. The mechanism is visible in conduct: the tighter drawdown rules cut its order flow roughly 7× (1,982 fills → 368), and one pack paid for the caution with disengagement. GLM healed its named wound and still got worse: china-mining-ban went 0 → 75, but the added caution killed it in a squeeze its v1 skill had cleared at GMI, and three packs dropped from positive-return rungs to plain survival — score 71.15 → 55.77. Generation 2 achieved its stated objective and lost 15 points doing it.
The full ladder, on the primary subject:
| Qwen3.7 Plus | v0 bare | v1 + skill | v2 self-evolved |
|---|---|---|---|
| WAGMI Score | 23.08 | 55.77 | 57.69 |
| Deaths | 1 | 3 | 0 |
| Illegal orders | 116 | 1 | 0 |
| Fills | 10 | 1,982 | 368 |
8.4 What one generation of evolution shows. Three findings. First, the loop closes and improvement is attributable: the evolver saw nothing but its own receipts, the receipts named three deaths, the revision targeted exactly those, and on the next run exactly those deaths were gone — receipts in, targeted revision out, sealed delta measured. Second, skills are load-bearing: a 1.6 KB text file took one model from 92% illegal orders to one gate block in 3,151 decisions, and another from zero trades ever to 2,200 legal fills. Third, the fitness signal is honest: GLM's generation healed its wound and scored lower, because the bench does not reward the loop unless conduct improves everywhere it counts — and GLM's new receipts now name the next wound, which is precisely the input a third generation would consume. A benchmark that can say "worse" is the only kind whose "better" means something.
Integrity: all 78 bundles across the three generations replay
byte-for-byte with zero mismatches
(bundles/evoskill-wave/, scores in reports/evoskill-v1-score/ and
reports/evoskill-v2-score/, computed against the same 52 baseline
bundles as every other table in this paper). The compiled prompt hashes
for all three generations are recomputable from the repository: v1
sha256:d7e14215…, v2-qwen sha256:f659e4bb…, v2-glm
sha256:4590dfd3…. One disclosed harness gap: in the sandboxed agent
lane the bundle manifest currently seals the hash of the base prompt
(sha256:bf8591c0…, constant across generations) rather than the
skill-compiled prompt; until the fix ships, which skill ran is attested
by the run configuration plus deterministic recomputation, not by the
bundle seal alone. The claim label on every number in this section is
survival-stress: this is evidence that a skill changes conduct on
frozen regimes and can be revised from its own receipts. It is not
evidence of predictive ability, and one generation on two models is a
demonstration, not a statistical claim.
9. Known limitations (the honest section)
We consider intellectual honesty a design feature. Current shortcomings:
- Memorization. COVID is in every training set. Blinding removes labels, not familiarity: a model can plausibly recognize famous price action, and venue constants (fee tier, tick size, funding caps) leak era information. We disclose this rather than pretend otherwise — hence the claim label. Historic packs measure stress behavior, not foresight.
- Bar-level simulation. Decisions at 1h/4h granularity against OHLCV bars, not tick data or an order book. The impact model is stylized (spread + size-dependent impact), not a matched-engine simulation. Intra-bar liquidation checks use the wick extremes, which is conservative but still an approximation of path dependence.
- Single venue, single asset. V1 is BTC-USDT perps with one uniform venue-parameter baseline across all 13 packs; historical parameter verification against primary sources is pending, so no era-accuracy claim is made.
- One deterministic run per pack. Temperature 0, no seed sweeps, no confidence intervals. Descriptive results, not statistical claims — a deliberate V1 trade for perfect reproducibility.
- One fixed scaffold. We test models under one baseline prompt and action grammar. A model's result reflects the model in our harness; different scaffolding could do better or worse.
- No information inputs. Agents currently see prices and funding only — no news, no sentiment. Fixed next (see below).
- Verdict/P&L interplay. The kill switch locks in whatever P&L existed at breach (short-and-hold "made" +54.5% and died). Tier semantics now exclude dead baselines from the return bar for exactly this reason, but the deeper question — how to present a killed strategy's locked-in gains without implying success — remains a presentation-design problem we keep revisiting.
10. Next steps and open questions
WAGMI Bench starts with BTC because BTC has the longest continuously recorded perpetual-futures tape available to us, and that history already contains every regime — melt-ups, chop, and violent unwinds — a perp market can produce. It does not stay there. The instrument BTC proved out is absorbing other asset classes at pace: monthly tokenized-equity-and-index derivative volume grew from $64.3B in January to $606.2B in July 2026 [theblock.co/api/charts/chart/decentralized-finance/derivatives/tokenized-equities-monthly-volume], and a CFTC-regulated US exchange cleared $21.85B of perpetuals notional in its first eight weeks [theblock.co/api/charts/chart/decentralized-finance/derivatives/kalshi-perpetuals-cumulative-volume]. The roadmap below follows that arc outward: more venues on the same architecture, then the asset classes now migrating onto perpetual rails, then the harder measurement problem underneath all of it — agents trading perpetual futures natively, evaluated on information they could not have memorized, not only replayed against history they may have seen.
Committed next steps:
- WAGMI Score — shipped in v0 (Section 7): one aggregate,
flat-hold-proof, survival-dominant score across the full 13-pack catalog,
computed by the open
wagmibench scoretool from public bundles. Still outstanding: the public leaderboard surface itself, where every row links to its replayable evidence root, and additional models to populate it — a one-agent table is a demonstration, not a ranking. - The informed track: packs built on post-model-cutoff 2026 windows (e.g. the crude-oil market through the Iran conflict) with point-in-time news served from crawl-date-gated archives. Unmemorizable by construction — the first honest skill-with-information evals in the category.
- More venues: perp DEX packs (dYdX, Hyperliquid and others), making the bench chain-agnostic in practice, not just architecture.
- Forward paper seasons and a hosted verified lane: agents run in a sealed sandbox on live paper markets, producing skill evidence that cannot be memorized because it has not happened yet.
Open questions we are actively thinking about:
- How do you measure regime-reading separately from memorization on historical data? Perturbed/synthetic control packs? Date-shifted ablations? Section 7.3 makes this the most urgent question we have: we now hold a result — three clean abstentions on three famous disasters — that has two incompatible explanations and no historical experiment that separates them.
- What should an aggregate score demand of engagement? The v0 ladder answers this crudely (flat-hold floor at 25, engagement at 50) and Section 7 suggests the floor is at least defensibly placed, since a churning trend-follower scores below it. Unresolved: whether the demanded engagement should scale with the regime, so that sitting out a grinding range and sitting out a crash are not scored identically.
- Should agents state confidence, enabling calibration and bet-sizing (Kelly-style) evaluation? This requires an action-schema extension and is the most-requested measurement we do not yet make.
- What is the right statistical treatment when runs are expensive? A full 13-pack matrix cost $40.78, so the question is now concrete rather than hypothetical: at that price, is the next dollar better spent on repeated sampled runs per pack (getting variance) or on more packs and more models (getting coverage)?
- How should a benchmark treat deliberate abstention — is "I don't trade this regime" a failure, a pass, or a category of its own? The first full-catalog run produced three of them, and our current answer (record it, floor it at 25, never call it skill) is a placeholder we do not yet fully believe.
11. The WAGMI Bench Charter (v1)
This section is the benchmark's constitution: every commitment is either
shipped, version-targeted, or declared out of scope — nothing is silent. It
was synthesized against the 2026 state of the art in benchmark design: the
Agentic Benchmark Checklist (Zhu et al., NeurIPS 2025), BetterBench
(NeurIPS 2024), Anthropic's "Adding Error Bars to Evals," Terminal-Bench
2.0's task-rigor process, the temporal-contamination literature
(Lopez-Lira, Tang & Zhu 2025 above all), and the EU GPAI Code of Practice's
evaluation-disclosure requirements. The working notes behind it live in
our internal notes, alongside the citation-verification ledger: every anchor
cited here was verified against a fetched primary source on 2026-07-30, per
our own no-recalled-numbers policy. The standalone one-page version is
published alongside this paper as CHARTER.md; the two are kept identical
below the header.
What WAGMI Bench measures. Risk conduct under stress — drawdown control, liquidation avoidance, cost discipline, and rule-following — measured by dropping agents into thirteen Bitcoin regimes that are each lethal to some strategy — crashes, squeezes, and grinding ranges — on a deterministic perp-exchange simulator with agent-adverse costs — starting with BTC, the perpetual market with the longest continuously recorded tape available to us.
What it never claims. Predictive skill or future returns. On
pre-cutoff data, memorization and genuine skill are formally
non-identifiable (Lopez-Lira, Tang & Zhu 2025); we do not pretend to
identify them. We measure what stays meaningful under full memorization,
label every result survival-stress, and enforce that label fail-closed on
every surface. WAGMI Bench is research infrastructure, not investment
advice.
The vocabulary. Contamination-limited, never "contamination-free." Every published number traces to a sealed evidence bundle or a fetched artifact — never a remembered one.
Commitments
Status: ✅ shipped in v1 · → vN committed, version-targeted · ✖ out of scope.
Construct & claims
| Commitment | Status |
|---|---|
| Explicit construct + non-claims, enforced fail-closed on all report surfaces | ✅ |
| Intended-use and not-investment-advice statement | ✅ (this charter) |
| Four-way validity-threat statement (see footer) | ✅ |
Reproducibility & evidence
| Commitment | Status |
|---|---|
| Open harness, open data recipes, one-command byte-exact replication | ✅ |
| Frozen environment; integer-arithmetic money path; determinism ≠ zero sampling variance, stated | ✅ |
| Sealed tamper-evident bundles; 100% of leaderboard rows re-verified, zero self-reported (satisfies the EU GPAI CoP Measure 7.3(1)(e)-(f) reporting schema, incl. ≥5 sample trajectories per evaluation) | ✅ |
| No network access by construction; agent affordances standardized and disclosed (interface contract, retry policy, inference params in manifests) | ✅ |
Contamination
| Commitment | Status |
|---|---|
| Date-tag every scenario; per-model eligibility windows; full-catalog and post-cutoff-only scores; affected rows flagged on the leaderboard itself | → v1.1 |
| LAP memorization probe (date-only recall of BTC direction) published per model — first application to crypto | → v1.1 |
| Model-cutoff audit per row (vendor cutoffs probed, never trusted) | → v1.1 |
| Counterfactual synthetic-path arm; headline contamination number = real-minus-synthetic survival gap | → v1.2 |
| No canary strings (demonstrated broken); no prompted-ignorance mitigations (demonstrated to fail) | ✅ stated |
| Post-cutoff informed track on crawl-date-gated archives — unmemorizable by construction | → roadmap, committed |
Scoring & statistics
| Commitment | Status |
|---|---|
| Fully specified ordinal ladder, open scoring tool, receipts | ✅ |
| Score-interpretation guide (what a gap does not mean) | → v1.1 |
| Trivial baselines on every pack; surviving baselines set the top-tier bar | ✅ |
| Survival cannot be earned by inaction (engagement gating, flat-hold floor) | ✅ |
| Abstention, parse-failure, missed-decision, and gate-block counts recorded in every bundle | ✅ · surfaced per leaderboard row → v1.1 |
| Scenario-clustered standard errors; paired Wilcoxon between models; published minimum detectable effect at n=13; leave-one-scenario-out robustness | → v1.1 |
| Multi-seed runs (≥5, the field norm) | → funding-gated; single-run rows labeled as such until then |
| Cost and wall-clock published per run | ✅ |
Governance
| Commitment | Status |
|---|---|
| Date-stamped immutable versions; cross-version comparability rules | ✅ |
| Full-catalog-only leaderboard; train-on-test ban | ✅ policy |
| Scaffold disclosure; Closed (reference prompt) vs Open (bring-your-own-scaffold) divisions | → v1.1 |
| Held-out scenario reserve with a written public policy — our answer to the multiple-testing / backtest-overfitting critique | → v1.2 |
| Maintenance, saturation, and retirement policy | → v1.1 |
| Croissant metadata + explicit dataset licensing | → v1.1 |
Adversarial validity
| Commitment | Status |
|---|---|
| Oracle solvability: a hindsight reference agent proving every pack survivable with engagement | → v1.1 |
| Pre-release adversarial exploit pass per scenario, findings published | → v1.1 |
| Stated reward-hacking detection method (transcript review, CAISI-style) | → v1.1 |
| Prompt-sensitivity annex | → v1.2 |
Out of scope, with reasons
| Exclusion | Why |
|---|---|
| LLM-judge scoring, anywhere | Ground truth is executed economics; judges add bias we cannot audit |
| Profit/alpha ranking on historical data | Non-identifiable from memorization; the field's central error |
| Human-baseline lane | Construct is agent risk conduct; a human lane measures something else |
| Live-web access inside historic packs | Contamination by construction; reserved for the informed track |
Validity threats, in one breath
Internal: the simulator is the world — costs are modeled, not exchanged; mitigated by agent-adverse rounding, dual cost profiles, and wick-level liquidation. External: 13 BTC windows are not all markets; mitigated by venue and asset expansion, never by claiming generality. Construct: survival can be gamed by inaction or harness failure; mitigated by engagement gating, abstention instrumentation, and the exploit pass. Statistical: n=13 scenarios and single seeds bound what any gap can mean; mitigated by clustered errors, paired tests, and published minimum detectable effects — and until then, by refusing to over-claim.
Reproduce everything in this document:
git clone https://github.com/Leonwenhao/wagmi-bench && cd wagmi-bench
uv sync --group dev
uv run wagmibench # guided wizard: demo, benchmark, compare
Recompute the Section 7 results yourself, from the published bundles — no API key and no inference spend required, because the decisions are already recorded:
# 1. Verify one episode replays byte-for-byte (do this for any row you doubt).
# Replay re-drives the recorded decisions through the engine, so the pack
# must be rebuilt locally from its manifest first — one checksum-verified
# fetch, no exchange data redistributed by us.
uv run wagmibench fetch-data --pack 2023-dead-zone
uv run wagmibench replay \
--bundle bundles/k3-matrix/2023-dead-zone \
--pack 2023-dead-zone
# 2. Recompute the WAGMI Score table. Pass the candidate bundles AND the
# keyless baseline bundles for each pack, so the GMI rung is decidable.
uv run wagmibench score \
$(for b in bundles/k3-matrix/* bundles/baselines-2026-07-28/*; do
printf ' --bundle %s' "$b"
done) \
--output reports/my-score
# 3. Diff your receipt against ours: same rungs, same 65.38, same roots.
diff reports/my-score/score.txt reports/leaderboard-final/score.txt
# 4. The five-model table of Section 7.7 recomputes the same way: add the
# wave bundles and diff against the 2026-07-31 receipt.
uv run wagmibench score \
$(for b in bundles/k3-matrix/* bundles/oss-wave/*/* \
bundles/baselines-2026-07-28/*; do
printf ' --bundle %s' "$b"
done) \
--output reports/my-five-model-score
diff reports/my-five-model-score/score.txt \
reports/leaderboard-2026-07-31/score.txt
The rendered table (score.txt) is byte-identical to our published receipt.
score.json carries the same rungs, verdicts, fill counts and evidence roots,
but lists packs in the order the bundles were passed in — sort by pack
before diffing the JSON. Both cost profiles are already inside every bundle,
so the stress column of Section 7 recomputes from the same command: there is
no cheap-cost variant of these numbers to fall back on.
Questions, corrections, and arguments welcome — that is what the evidence is for.