Info — what everything on this dashboard means
How to read this dashboard
Each run is one research cycle: the agent studies live Grand Exchange data and produces strategies.
Every strategy is then paper-traded — re-evaluated against live prices until its evaluation window
ends (the evaluator ticks every 5 minutes; F strategies evaluate at most every 15 minutes, B hourly). The scoreboard aggregates those outcomes per archetype and automatically feeds the next run's brief.
Between runs, the collector sweeps the market hourly and queues signals for the next run to investigate.
The strip under the nav shows what's in flight right now: the active run (if any), open and armed positions, and pending signals.
Strategy lifecycle
openA live paper position, currently being evaluated (every ~15 min for F, hourly for B).
armedA volume-anomaly (V) strategy waiting for its trigger condition to fire — no position yet.
confirmedSurvived its full evaluation window ≥80% healthy AND delivered ≥half its projection. A win for its archetype.
killedA sustained kill signal — the kill price crossed, or the live cycle capacity fell below the lane floor — over consecutive evaluations (F: 3 × 15 min, B: 6 × 1 h). Closed early.
expiredThe evaluation window ended without confirmation, or an armed trigger never fired.
vetoedRejected at ship time — never traded. The reason (on the run page) names which vet rule fired.
Ship-time vetting
Every strategy the agent ships passes six mechanical rules before it enters the book; failures become
vetoed rows with the rule in the reason:
weight 0The run brief set its archetype's weight to 0 ("do not pitch") — currently true for B.
dedupThe item already has an open/armed strategy of the same archetype.
kill pre-breachThe stop was already crossed at ship time.
capital / floorDoesn't fit the 50M per-opportunity budget, or misses its lane's absolute floor (F: 400k gp/cycle = 100k gp/hr, raised from 200k on 2026-08-01; B: 100k/cycle).
EV sanityThe claimed per-cycle gp exceeds 2× a live recomputation from current margin and fillable size.
persistence (F)The claimed post-tax margin held ≥ half its width in fewer than 40% of the last 24 hours — a momentary spike, not a standing spread. Added 2026-08-01 after the first fortnight showed spike-top pitches losing 60% of their margin within 15 minutes of shipping.
Live verdicts
On Actionable now, each strategy is re-checked against current prices when the page loads:
healthyAll health checks pass — the thesis still holds at current prices.
degradedSome checks fail (stale prices, thin volume, entry/exit out of reach) — proceed with caution.
kill_signalThe kill condition is met right now — the strategy's own rules say get out / don't enter.
Confidence
The agent's own rating of how well the data supports the thesis:
high medium low
insufficient_history — the last means the item lacks enough price history to judge.
Archetypes
FVolume flip (the flagship) — a persistent post-tax spread on a deep commodity market (100k+ units/day), captured a full buy limit per 4h cycle. Ships only when the margin's 24h persistence ≥ 0.4 — a spread that held ≥ half its width for at least 40% of the last day.
BHigh-value flip — a 10M+ item flip. Currently OFF (weight 0, vet-enforced): the first fortnight went 3 winners in 33 at −67.5M paper gp, and lost −21.6M even in the most generous counterfactual. May return re-specced as a round-trip-gated quick arb.
VVolume anomaly — arms on a detected anomaly, then waits for a trigger (volume z-score or price move) before entering.
CConversion — buy inputs, sell outputs across a crafting/processing relation; inert while prod has no relations table.
UUpdate / event — positioned ahead of a dated game update or event expected to move the item.
Retired: S (seasonal window) and H (swing hold) — 0% of closed strategies confirmed at this data age; their analytics live on as timing/qualification evidence for F and B. Letters A–G in old scoreboard rows are pre-re-architecture history.
Health-check chips
Each evaluation runs archetype-specific checks: vol_ok means passed,
vol_ok means failed. Any failing check degrades the verdict; kill checks trip it to kill_signal.
kill_price_okPrice has not crossed the kill level.
legs_freshAll legs' prices traded recently (not stale).
buy_leg_fresh / sell_leg_freshThe buy/sell side specifically traded recently.
legs_pricedEvery leg has a price at all (nulls mean nothing traded that side).
entry_reachableCurrent price makes the planned entry realistic.
exit_reachableCurrent price makes the planned exit realistic.
vol_okEnough recent volume to actually fill the planned units.
margin_aliveThe live post-tax margin is still positive (F/B).
floor_okThe live cycle capacity (post-tax margin × fillable units) still clears the lane's absolute floor — the floor that ships a strategy is the floor that keeps it alive. Breaching it sustained is what kills most flips.
margin_okConversion margin is still positive after tax.
out_of_windowCurrently outside the strategy's buy/sell hour-of-week window (informational).
windows_observedThe seasonal windows have been observed in recent data.
armed_waitingV strategy is armed and its trigger hasn't fired yet.
trigger_firedV strategy's trigger condition has fired.
Which checks apply depends on archetype — F/B: kill_price_ok, legs_fresh, margin_alive, entry_reachable, exit_reachable, vol_ok, floor_ok ·
V: armed_waiting, trigger_fired, kill_price_ok, legs_fresh, vol_ok, exit_reachable ·
C: legs_priced, legs_fresh, vol_ok, margin_ok ·
U: kill_price_ok, legs_fresh, entry_reachable ·
retired S/H rows used their own historical check sets.
Runs
runningThe agent is mid-cycle — its run page streams live progress.
succeededFinished with an accepted report and strategies.
failedDied or its report was rejected — see the fail reason on the run.
Only one run can be active at a time; triggering a new one while a run is active redirects you to the active run.
Signals
Kinds (which collector lens produced it): vflip volume-flip candidate (lane F) ·
hvflip high-value-flip candidate (lane B — informational while B is off) ·
volume volume anomaly · band price at the edge of its band ·
seasonal recurring hour-of-week pattern (timing evidence for the flip lanes).
pendingQueued — will be assigned to the next run's brief.
assignedPicked up by a run currently investigating it.
investigatedThe run shipped a strategy from it.
dismissedInvestigated but not worth a strategy, or expired unused.
Numbers glossary
gp/1h projThe projection the paper-trade is judged against. Since 2026-08-01 this is the harness's own recomputation at ship time — the same haircut arithmetic (slippage, tax, 15% participation) the evaluator realizes with — not the agent's claim; the claim is only the fallback on older rows. Before that change, agent claims ran ~8× the haircut math and nothing could ever confirm.
gp/1h nowWhat paper-trading says it's actually paying, with the self-impact haircut: 15% volume participation and 0.5% slippage. Treat paper results as modeled upper bounds.
roiProjected return on the capital tied up.
realized / projectedScoreboard ratio of median realized gp/1h to projected. Now apples-to-apples: ≥ 0.5 over a healthy window is the confirmation bar; 1.0 = the spread delivered exactly what the harness projected.
buy/sell windowHour-of-week ranges in UTC (e.g. "Tue 02:00–Tue 06:00") when the seasonal pattern says to trade.
triggervolume_zscore or price_move_pct going above/below a threshold over a window — what fires an armed V strategy.
buy / sellLeg sides: what to buy and what to sell, with quantities and the prices at spec time.
A dash (—) means no data: null prices are signal here — nothing traded that side in that window. They are never zero-filled.