trading / confluences

The Confluence Courtroom

Every confluence this desk has ever put on trial — the six live gates of the P1 Master method, the 2026-08-07 candidate screen, the bits that were dropped before they cost a multiplicity slot, and the ones it is now forbidden to re-run. Permanent record; nothing here has ever changed the live method.

BURNED DATA — 2019–2026 was selected on before this lab existedNOT TV-VERIFIEDSCREENING ≠ CERTIFICATIONOPTIMISTIC FILL MODEL — touch fills, zero slippageMES ONLY · P1 M2 managementSOLO LEDGER IS NOT A BOOK — never summedPROMOTIONS TO DATE: 0
Confluences on record
29
6 live · 15 screened · 4 dropped · 4 do-not-retest
Screen population
2,882
candidates · 2,289 filled · MES · 2019–2026
FDR survivors
2
of 14 scoreable at q=0.25 · both benched
Bits promoted to live
0
the seat has never been spent

What this is

The Courtroom is a permanent, cheat-resistant way to ask one question — does this confluence actually help? — and, more importantly, a way to not fool ourselves when the answer looks like yes. Everything the lab produces is a ranking. A ranking is not a decision. Nothing changes the live method until it has walked the whole gauntlet, and most things never will.

Its central scarcity rule is the one-challenger law: at most ONE candidate bit is under test at any moment — not a family, not a sweep, one. The registration is written into a hash-chained, append-only ledger before the bit is scored, carrying its promotion rule, its minimum effect and its quarantine start date. This is enforced in code, not by good intentions: the registry refuses a second open challenger, an incomplete registration, a bit whose source hash no longer matches what was registered, and any attempt to re-open a closed experiment. Because a forward quarantine is measured in months of wall-clock, the seat is the programme's scarcest resource — which is why spending it on a case you expect to lose is itself a ruled-against move.

Failures are permanent. A rejected bit is appended to the ledger with its verdict and is never deleted, never quietly re-run with different settings, never retried "because the market changed". A genuine retest is a NEW registration that cites the old one — so the count of attempts, and the multiplicity it creates, stay visible.

The gate pipeline — all seven, in order, no skipping

  1. Gate 1 · Time machine. The ledger is rebuilt in cold processes from mutated raw data — determinism, truncation, suffix poison, and a bar-level probe from a candidate's exact decision minute. Two deliberately broken bits ride along in every rebuild; the harness MUST flag both, or the lab refuses to score anything.
  2. Gate 2 · Six-year report card (EXPLORATORY ONLY). Conditional association vs a per-bit prevalence-matched null. All of 2019–2026 is burned data, so a good number here means "worth pre-registering", nothing more.
  3. Gate 3 · One pre-registered challenger. The one-challenger law, in code. Grade-layer changes only, on an already-certified emission stream.
  4. Gate 4 · Forward quarantine + paired evaluator. The challenger must win on data that did not exist when the hypothesis was written. The forward window is cut out of the INPUT, not filtered out of a report. Bar: ≥ +$1,500 net paired per 12 months, pro-rated from the registration date, AND no worsening of max EOD drawdown.
  5. Gate 5 · Engine A/B. Implemented as a knob that defaults to legacy-inert; the champion book must re-run 225/225 bit-identical and btdb verify must print 16 passed, 0 failed.
  6. Gate 6 · TradingView Pine referee. An independent port must reproduce entries AND per-trade PnL on a Deep Backtest export. Until then the only correct sentence is "not TV-verified yet".
  7. Gate 7 · Failures stay on the ledger forever.

The screen that produced most of this page is the Gate-2 instrument applied as a screen, with parameters fixed in a sha-pinned design document before any value was computed, and BH-FDR at q=0.25 across the declared 15-bit family. Its own warning, verbatim:

HISTORICAL SCREEN, NOT CERTIFICATION. 2019-2026 is burned data. Nothing here is a forward test, an engine A/B or a Pine referee; nothing here may touch the live method. It nominates at most ONE bit for the Courtroom's challenger seat, and only Arman's registration puts it there.

Did the instrument work? · calibration, run before any candidate was read

ControlRequirementResultPasses
ctl_plantedplanted +0.30 R must be DETECTED (p < 0.05)Δ +0.2979 · p=0.0000YES
ctl_null ×20≤ 3 of 20 permuted columns below α=0.051 of 20YES — calibrated
ctl_null (single)the design doc's literal wording: p ≥ 0.05 on one designated seedp=0.0196NO — reported verbatim; a single null draw lands below α with probability α by construction, so this is not the instrument the verdict rests on
leak self-teststwo deliberately broken bits must be CAUGHT3,817 + 3 violations flaggedYES

Status legend

LIVEIn the live P1 Master gate stack today.
SURVIVOR-BENCHEDCleared the screen's FDR gate; benched, not registered.
DEADScored and did not clear its own noise floor.
DEGENERATEZero variance on the filled rows — not scoreable.
UNDERPOWEREDOne arm too rare to refute anything.
DROPPEDKilled before scoring, in the open, with a written reason.
DO-NOT-RETESTFDR-dead 2026-07-21; re-running it is forbidden by design §1.

Every confluence ever tested

Live-gate statistics come from the six-bit commissioning family (exp-0001, 2026-08-05); screen-slate statistics come from the fifteen-bit screen family (screen-2026-08-07). The two q columns are corrections over different families and are NOT comparable to each other — both are exploratory sorting keys, never evidence.

ConfluenceStatusPrevalenceΔ mean Rmatched pqHeadline
Live gates — the P1 Master stack, commissioning stats
c1_trendLIVE32.2%+0.08080.17400.348WEAK
c2_vwapLIVE63.3%+0.05880.31190.416INDISTINGUISHABLE
c3_volumeLIVE80.6%+0.06630.34680.416INDISTINGUISHABLE
c4_sweepLIVE75.7%+0.13120.04350.261NOMINAL
c5_htfLIVE28.2%+0.08860.15250.348WEAK
c6_roomLIVE9.6%+0.03730.69400.694INDISTINGUISHABLE
Screen slate 2026-08-07 — FDR survivors (q ≤ 0.25)
b_onbreakSURVIVOR-BENCHED21.4%+0.16640.01410.197survivor — benched, seat not spent
b_actfloorSURVIVOR-BENCHED78.5%+0.14440.03150.221survivor — benched, seat not spent
Screen slate 2026-08-07 — scored and dead
b_gapadrDEAD69.5%+0.11060.06600.308failed FDR gate (p=0.066) — no A/B earned
b_pmorDEAD76.9%+0.10120.12610.44187.4th pctile of its own null
b_volhardDEAD87.7%-0.10280.22760.63777.2th pctile of its own null
b_pdbreakDEAD28.1%+0.06540.29040.67871.0th pctile of its own null
b_adxDEAD72.3%-0.05100.41180.82458.8th pctile of its own null
b_orwcapDEAD83.2%-0.04500.54630.87445.4th pctile of its own null
b_dwellDEAD70.6%-0.02650.66270.92833.7th pctile of its own null
b_blackoutDEAD38.0%+0.01740.75760.95224.2th pctile of its own null
b_stopbandDEAD82.7%-0.01520.84020.95216.0th pctile of its own null
b_volregimeDEAD43.4%+0.00840.88370.95211.6th pctile of its own null
b_nr7DEAD16.0%-0.00400.95750.9584.2th pctile of its own null
Screen slate 2026-08-07 — not scoreable
b_avwapdayDEGENERATE100.0%degenerate — consumes no verdict
b_dryupUNDERPOWERED2.5%+0.10340.56180.874underpowered — consumes no verdict
Dropped in the open — before any outcome value was computed
b_closeconfDROPPEDstructural tautology — 400/400 sampled candidates
b_roomstopDROPPEDduplicate of the live c6 at the same 2.0 multiple
SW-31-breakout-latency-windowDROPPEDreduces to FDR-dead n4-earlySignal
EN-12-signal-bar-range-capDROPPEDsource states no number — would be tuning-by-authorship
Do not retest — FDR-dead in the 2026-07-21 deep-dive
n1-narrowORDO-NOT-RETESTNarrow opening range as a positive signal — rewards OR narrowness.
n2-gapAlignDO-NOT-RETESTOvernight gap direction aligned with the trade direction.
n3-displacementDO-NOT-RETESTSignal-bar displacement / bar-size as a quality proxy.
n4-earlySignalDO-NOT-RETESTHow soon after the opening range the signal fires.

Why the four dead bits may not be re-run

BitStanding
n1-narrowORFDR-dead in the 2026-07-21 confluence deep-dive. Two later candidates were built specifically to NOT be it and both said so in writing: b_orwcap (wide side only, excludes rather than rewards) and b_nr7 (prior-DAY range, not OR width). Both then failed on their own merits.
n2-gapAlignFDR-dead 2026-07-21. Note b_gapadr is a different mechanism — gap MAGNITUDE as a band, not gap direction — which is why it was allowed a slot; it then failed the screen's FDR gate anyway (p=0.066, q=0.308).
n3-displacementFDR-dead 2026-07-21. It is the reason EN-12 was dropped (a bar-size cap is adjacent to it) and the reason b_dwell's registration insists in writing that it is persistence, not bar size.
n4-earlySignalFDR-dead 2026-07-21. It is the reason SW-31 was dropped without being measured.

2026-08-08 tribunal ruling — live stands

The ruling: do nothing. Case closed.

Two independent judges, five reviews (two with full primary-data access, three packet-only), ruling on the 31-cell config grid. Keep the live six-gate stack untouched. Register no challenger. Record the result as "live stands" — which the config-challenge workflow explicitly calls a success of the process, not a failure.

All three of the grid's both-era winners are dead under the study's own pre-declared Stage-1 bar ("cells failing any item are DEAD"):

The promotion arithmetic forecloses the seat independently. Over the true 7.16-year span the burned, in-sample, optimistic-basis deltas annualise to $661/12m (actfloor→c6), $544/12m (CONFIG_9_k5) and $380/12m (CONFIG_7_k4) against a $1,500/12m bar. The best cell's most flattering number is 44% of the bar, and forward performance will be lower. Registering any of them is a pre-paid rejection.

Disposition: the live stack stands; b_onbreak goes to the bench as the one object the study produced that deserves to live, re-testable only on a future pre-registered grid over data it has not seen (post-2026-07-31), with a zero-risk shadow-fill log available to resolve its fill-conditioning without spending the seat; b_actfloor is benched behind a required drop-c6 ablation; b_gapadr is not to be resurrected in any form. The challenger seat stays open.

Final line, verbatim: "live stands; all three winners dead under the study's own law; b_onbreak to the bench; the seat stays open. A smaller true number beats a bigger fragile one — and this grid's bigger numbers are not merely fragile, they are mechanically explained."

Standing caveats — attach to every number on this site

screen-2026-08-07 · schema conflab.screen.v1design doc sha256 e2ba7ca77acce01c…bits.py sha1 5473adbbf694…harness finished 2026-08-07← trading