BETA · descriptive · not a signal
The Engine Read track record
Engine Read is our internal multi-layer directional research engine. This page is
its measured record — the hit-rate history the accuracy loop computes
every round, published with the same rule as everything else we ship: every number
carries its sample size and its confidence bound, and the unflattering parts print
first. It is not the liquidation-map accuracy ledger (that lives at
/proof), it is not a signal feed,
and nothing here has cleared the promotion kill-gate.
window: 7 days · headline horizon 1.0h
measured 2026-08-16T20:19:02Z
anti-Goodhart: CLEAN
The served stream — and the caveat that comes first
Read this before the numbers. In this window the
engine labelled 100.0% of resolved hours “sideways”, so its
3-class hit rate 87.5% (sample size n=8,973, Wilson 95% lower
bound 86.8%) tracks the sideways base rate 87.5% with a
residual of 0.0%. In plain terms: calling “sideways” almost
always is what produced this number — it is honest, it is measured, and it is
not evidence of directional skill. The directional test is the shadow stream
below, and today it is BETA.
| Segment | 3-class hit rate | n |
Moved fraction | Directional calls | Directional wrong |
|---|
| last 24h | 95.2% | 980 | 4.8% | 0 | 0 |
| 24–48h ago | 89.7% | 1,343 | 10.3% | 0 | 0 |
| prior 5 days | 85.9% | 6,650 | 14.1% | 0 | 0 |
Segments of the same 7-day window. “Moved
fraction” is the share of hours price moved beyond the sideways tolerance
— the hours a sideways call gets wrong.
The directional shadow stream — pre-registered verdict
Every directional call the engine would have emitted is logged and graded against
a pre-registered graduation ladder (unvalidated → beta → validated)
with fixed evidence floors — the ladder cannot be moved to fit the results.
The verdict below is computed by the same function the guardian journal records;
it fails closed to the stricter verdict on any degraded input.
Verdict: BETA · outlook (descriptive
annotation, never a gate input): converging_negative. Shadow directional hit rate
10.1% [9.0%, 11.4%]
over n=2,415 resolved calls
(17 resolved days) — against a sideways
base rate of 82.0% (Wilson low
80.4%). BETA — enough resolved evidence to publish the track record with its bound; the beat-the-base-rate gate has NOT cleared (no edge claim).
| Direction | Shadow hit rate | n |
|---|
| downward | 3.5% | 774 |
| upward | 13.2% | 1,641 |
Conditioned on hours that actually moved (n=435),
the shadow stream’s directional hit rate is 56.1%.
- beta — cleared: n resolved 2,415 / 60 ✓; resolved days 17 / 14 ✓
- validated — not cleared: n resolved 2,415 / 120 ✓; resolved days 17 / 28 ✗
Calibration record — every horizon, every confidence bucket
combined corpus: 153,149 samples
111,706 scorable
combined 3-class: 50.6%
combined directional: 51.5%
(random baseline 50.0%)
| Horizon | Hit rate | n scorable |
Directional hit | High-conf hit | n high-conf |
|---|
| 1h | 85.6% | 30,613 | — | — | — |
| 4h | 85.5% | 30,454 | — | — | — |
| 24h | 76.9% | 29,343 | 67.3% | 60.7% | 3,578 |
| 168h | 76.9% | 21,296 | — | — | — |
| 720h | 0.0% | 0 | — | — | — |
Per-horizon hit rates from the calibration artifact
(model v1.cal-20260816, calibrated
2026-08-16T21:02:47Z, scoring basis
1h_canonical_per_T1_0d). “Hit” at short horizons is
dominated by sideways calls — see the disclosure above.
Is the engine’s own confidence calibrated? Measured, not assumed
| Confidence bucket | Expected hit rate |
Measured hit rate | n |
|---|
| 50-60 | — | 52.5% | — |
| 60-70 | — | 32.1% | — |
| 70-80 | — | 27.9% | — |
| 80-95 | — | 37.3% | — |
Expected = the bucket midpoint the confidence claim implies.
Today the higher-confidence buckets UNDERPERFORM their implied rates —
the engine’s confidence is not yet calibrated, and this page says so
rather than borrowing credibility from a number.
P&L directional edge check (counterfactual):
net as-emitted -0.67 USD over
232 actual directional
calls — edge positive: no.
The engine has not demonstrated a positive directional edge after costs.
Where it is worst — published, not hidden
- LINKUSDT — 71.9% (n=641, 180 sideways-calls that moved)
- ADAUSDT — 75.3% (n=641, 158 sideways-calls that moved)
- AVAXUSDT — 75.8% (n=641, 155 sideways-calls that moved)
- DOTUSDT — 82.2% (n=640, 114 sideways-calls that moved)
- POLUSDT — 84.7% (n=641, 98 sideways-calls that moved)
The weakest symbols in the current window, straight from the
measurement. A track record that only shows the wins is marketing.
The honesty block
Validation path: CONTINUE-BETA. Gate evidence today:
counterfactual net as-emitted P&L -0.84 USD (negative is a fail);
the blocked-stream (calls the engine withheld) hit 52.5%
over n=2,328. Outstanding:
consecutive_clean_windows 0 < 2; counterfactual as-emitted edge_positive != true (net -0.837949).
Kill-gate not cleared. Nothing on this page has cleared
the promotion gate (net_usd>0 ∧ ci95_low>0) that guards every
Hunter Killer strategy claim. This is a research engine’s measured record,
published because measured honesty is the product — not a signal, not a
forecast, not financial advice.
Basis: walk-forward, out-of-sample grading by the
accuracy loop; this page reads the loop’s measurement artifacts
read-only and never recomputes a figure. Historical, not predictive.
Related public records: the liquidation-map accuracy
ledger at /proof and /accuracy, the cascade anatomy at
/contagion, the hypothetical reference desk at /reference-desk.
Measurement stamp — loop state 2026-08-16T20:19:02Z ·
verdict BETA · outlook converging_negative