AV2

AV2

Isolated challenger experiments against frozen champions. A challenger changes exactly one thing, so any observed difference is attributable; the champion is never edited in place, and surviving Holdout prepares a challenger for champion-vs-challenger shadow rather than replacing the champion.

DashboardLive AuditStrategies

Champions
2
Challengers pre-registered
10
Diagnostics
1
Stages run
0
Promoted
0
No AV2 experiment has run, so no comparative metric exists for any challenger. The burden is declared up front: 10 challenger experiments plus 1 diagnostic across 2 champion(s) — at that count roughly one is expected to clear validation on noise alone, so a single Holdout survivor would be weak evidence rather than vindication.

Stage ladder

Parsed from the pre-registration document, not restated here
AV2.1 preregistration (this document)AV2.2 LightningAV2.3 robustness/regimeAV2.4 execution/statAV2.5 challenger validationAV2.6 HOLDOUT ELIGIBLEAV2.7 one-shot Holdout

Challengers of F6

5 pre-registered · champion frozen
ChallengerExact changeAltersDeclared overfitting riskStageResult
AV2-F6-01stepR ∈ {0.25, 0.35} (champion 0.5R). Everything else identicalexit onlyLOW — closes a known gap rather than searching new spacepre-registrationno result exists
AV2-F6-02stop to break-even once MFE ≥ 0.3R, then the 0.5R ratchet proceeds independentlyexit onlyLOWpre-registrationno result exists
AV2-F6-03ratchet step = 0.5 × ATR(14,5M) at entry rather than 0.5 × initial riskexit onlyLOW — mechanistic discrimination, not a searchpre-registrationno result exists
AV2-F6-04when the immediate fade stops out, re-enter ONCE in the expansion directionentry — adds a second signal classMEDIUMpre-registrationno result exists
AV2-F6-05trade only expansions in [3.5, 4.0] × ATR (champion: ≥ 2.5 ×, open ended)entry activationMEDIUMpre-registrationno result exists

Challengers of A2.10

5 pre-registered · champion frozen
ChallengerExact changeAltersDeclared overfitting riskStageResult
AV2-A210-01run the calm-fade branch alone, and the high-vol flip branch alone, as two separately preregistered challengersregime activationLOW. Counted as two experiments in the burdenpre-registrationno result exists
AV2-A210-02calm/flip boundary ∈ {1.25, 1.5, 2.0} (champion 1.5)regime activationLOWpre-registrationno result exists
AV2-A210-035-day max hold → flat at 15:59 ET the same dayexit onlyLOW — closer to spec hygiene, and informative either waypre-registrationno result exists
AV2-A210-04trigger bar at 14:30 or 15:30 instead of 15:00entry activationLOWpre-registrationno result exists
AV2-A210-05FIXED_2R → the opposite end of the day's own prior RTH rangeexit onlyLOWpre-registrationno result exists

Challengers of unassigned

1 pre-registered · champion frozen
ChallengerExact changeAltersDeclared overfitting riskStageResult
AV2-D-01diagnostic onlywhy does Holdout expectancy (+0.260R, n=100) diverge materially from Validation (+0.409R, n=88)? This is the exact unresolved risk the Final Research Audit cited when it withheld A2.10 from Phase B, and it is the highest-value open question across both championspre-registrationno result exists

Champions

Frozen; a challenger never edits one in place

AV2 rules

Binding on every item
  • · The champion remains frozen; a challenger is a separately versioned descendant, never an in-place edit.
  • · A challenger changes exactly one defensible mechanism, so any difference is attributable.
  • · No retroactive edits to champion history.
  • · Challenger and champion are compared on the same approved partitions.
  • · Pre-registration where required; no tuning after results, no rescue variants, no Holdout retries.
  • · A failed improvement is still useful evidence and stays on the record.
  • · Surviving Holdout prepares a challenger for champion-vs-challenger shadow — it does not replace the champion.

AV2 (this section) is a challenger program over the live champions. Architecture V2 is the separate approved-but-unimplemented tiered-quality architecture whose shadow classifier appears among the internal harnesses on the Monitor.

Read from research-output/av2-planning/AV2_1_PREREGISTRATION.md · research-output/av2-planning/AV2_RESEARCH_BACKLOG.md