VCET v1.1 Mechanism Acceptance Test Plan

Date 2026-06-04 · Status: PLAN ONLY — no DB / taxonomy / classifier / Scorecards changes · Anchor: Pull-In Research Hub · Scope: design the next acceptance test to decide whether VCET v1.1 + CBO can act as a mechanism classifier (not just a state-at-entry labeler).

What this plan answers (and what it doesn't)

Answers: if v1.1 lands, what is the exact test that decides whether it earns a slot in the workflow — what gets measured, with what thresholds, against what baseline, producing what report.

Doesn't answer: what v1.1's 6 extra checks or CBO algorithm actually do (we have v1.0 on disk, not v1.1). Those are open questions to be resolved when v1.1 source arrives. This plan is structured so it survives both the optimistic case (v1.1 detects CBO cleanly) and the conservative case (CBO is noisy and v1.1 doesn't beat v1.0).

Premise: the existing Quality Gate (RVOL + gap) is the primary pre-trade filter and stays. v1.1 is evaluated as a secondary mechanism-classification layer — the goal is to call which mechanism is in play at and after entry, not to replace the gate.

1. Hypotheses

Each hypothesis is testable, has a directional prediction, and has an explicit null. If the null wins, that's a real finding too — it tells us v1.1 doesn't move the needle on that dimension.

H1 — Mechanism coverage

v1.1 trajectory (entry → +25 min) produces a non-INSUFFICIENT mechanism call at MEDIUM or HIGH confidence on ≥70% of trades where bars are available.
Null: ≤50% coverage (v1.0 trajectory would already get there). Test: classify all PI-family + chase-extended trades with bars; count non-INSUFFICIENT confident calls.

H2 — CBO adds detection

CBO (Compression Breakout) signal flags ≥5 trades as range_break_hitchhiker that v1.0 cannot detect — these trades currently route to ROTATION or EARLY_EXPANSION generic states.
Null: CBO fires <5 times in the sample, or all CBO-flagged trades are already classified by v1.0 numerics alone. Test: count CBO-driven HIGH-confidence range_break classifications.

H3 — False expansion suppression

v1.1's expanded check set (13 vs v1.0's 7) reduces FAILED_EXPANSION_TENTATIVE/CONFIRMED rate on real winners and raises it on real losers — i.e., the engine becomes a better failure detector.
Null: FAILED_EXPANSION rate is uncorrelated with P&L outcome. Test: P&L sign vs FAILED_EXPANSION presence (chi-square or rank correlation).

H4 — Trend continuation entry is detectable

Trades where state at entry-5 was already TREND_* in trade direction and stays TREND_* through +25 are reliably classified as trend_continuation_entry — distinct from true_pull_in_to_trend by lacking a ROTATION/COMPRESSION precursor.
Null: trend_continuation_entry trades have the same P&L distribution as true_pull_in_to_trend. Test: this would mean management style doesn't need to differ — but the Mechanism Audit already showed they were a weak sub-group (LUNR loss).

H5 — Wrong-side escape detection

State transition to TREND_* OPPOSITE trade direction within 15 min, or slope sign flip against direction with magnitude ≥15° within 5 min, identifies wrong-side trades with ≥90% precision (low false positive rate).
Null: rule fires equally on trades that go on to profit handsomely. Test: per-trade audit of flagged wrong-side calls vs eventual P&L.

H6 — True chase requires vwap_distance_sd

If v1.1 surfaces vwap_distance_sd at entry as a first-class output, true chase is detectable as: distance >2.5 SD from intraday VWAP at entry AND state in EXPANSION_CONFIRMED/TREND_* AND no compression precursor. Without this field, true chase remains undetectable in the DB (current state).
Null: v1.1 doesn't expose vwap_distance_sd as a field, or distance threshold doesn't correlate with chase-pattern losses. Test: regression of P&L on vwap_distance_sd × state-at-entry.

H7 — Rubberband is detectable from MEAN_REVERSION state

Rubberband (LONG into oversold, SHORT into overbought) is detectable as: vwap_distance_sd ≥ 2.5 SD AGAINST trade direction at entry-5 or entry + state cycles into MEAN_REVERSION at entry or within ±2 bars + reclaim toward VWAP within 25 min.
Null: MEAN_REVERSION state is too rare in the live sample to detect any rubberbands. Test: count rubberband-classified trades; compare P&L distribution to non-rubberband mean-reversion trades.

H8 — v1.1 beats v1.0 on outcome correlation

v1.1 mechanism classifications correlate more strongly with P&L outcome than v1.0 state-at-entry labels — measured by Spearman rank between (mechanism class) and (P&L deciles).
Null: same correlation (≤0.05 improvement). Test: cross-classify with v1.0 baseline (no CBO, no vwap_distance, 7 checks) — same trades, same trajectory snaps, just v1.0 logic. Compare correlations.

2. Required data

DataSourceStatusNotes
VCET v1.1 engineProvided by upstream (currently only v1.0 on disk)BLOCKER13 checks + CBO. Must arrive before test can run.
v1.1 schema specSame source as engineBLOCKERNeed exact field names for CBO signal, vwap_distance_sd, and 6 new check outputs.
5-min barskti (14-day window)PARTIALSame blocker as Phase 4a — limits coverage to ~14 days. Sample size for full test = ~30-60 PI-family trades + ~10-20 chase trades with bars. Insufficient for HIGH-confidence acceptance on rare mechanisms (rubberband, range_break).
trade_context rows for test windowmain.trade_contextOKAlready populated with RVOL, gap, MTF, catalyst, MA stack. Use as Quality Gate pre-context.
v1.0 baseline for comparisonAlready on diskOKUsed for H8 — same trajectory data run through v1.0 to produce baseline classifications.
Bar archive beyond 14 daysTBD — not currently availableFUTUREWould expand sample size 5-10x. Acceptance test should be designed to re-run cleanly when archive lands.

3. Mechanism labels — the 7 classes to discriminate

Defined precisely so the test rules can be implemented mechanically. Each class lists the v1.1 fields required to call it confidently and the v1.0-only fallback (if any).

1 true_pull_in_to_trend

Hypothesis: compression-before-expansion signature with state transitioning to expansion/trend matching trade direction.

v1.1 fields needed: CBO confirmation (optional, raises HIGH confidence) · v1.0 fallback: works with v1.0 outputs alone.

2 vwap_rotation_scalp

Hypothesis: stays in rotation; profit comes from band-edge mean reversion, not breakout.

v1.1 fields: none required · v1.0 fallback: full coverage.

3 trend_continuation_entry

Hypothesis: trade enters when state was already TREND_* in direction — NOT a Pull-In conversion; a continuation play. Discovered via Mechanism Audit (LUNR, BB, INTC).

v1.1 fields: none required · v1.0 fallback: full coverage. This is the 4th mechanism the v1.0 rule set didn't model.

4 wrong_side_escape

Hypothesis: VCET trajectory moves AGAINST trade direction within 15 min; profit (if any) is luck, not setup.

v1.1 fields: none required · v1.0 fallback: full coverage. HIGH precision target: ≥90% — false positives would tell traders to exit profitable runners.

5 true_chase_extended

Hypothesis: entry late in an already-extended move, far from intraday VWAP, no compression precursor. Requires v1.1 vwap_distance_sd field.

v1.1 fields needed: vwap_distance_sd at entry (CRITICAL — not in v1.0 outputs) · v1.0 fallback: falls back to true_chase_inferred (LOW confidence) using atrs_from_prior_close ≥ 2.0 as a weak proxy.

6 range_break_hitchhiker

Hypothesis: trade enters on a consolidation breakout, CBO confirms the break, expansion follows. Requires v1.1 CBO signal.

v1.1 fields needed: CBO signal (CRITICAL — not in v1.0) · v1.0 fallback: falls back to range_break_inferred (LOW confidence) using compression→expansion transition without CBO confirmation.

7 rubberband_backside

Hypothesis: trade is a mean-reversion reclaim — entered when symbol is extended AGAINST trade direction; expansion fades back toward VWAP.

v1.1 fields needed: vwap_distance_sd (CRITICAL) · v1.0 fallback: uses atrs_from_prior_close AGAINST direction as weak proxy (MEDIUM cap, same as Chase Audit rule P2).

4. Test logic

Sequential, per-trade. Applied in priority order so that mutually exclusive classifications don't double-fire.

STEP 1 — Pre-filter (Quality Gate context, already available)
  Pull RVOL, gap, MTF from main.trade_context.
  Classify tape as GREEN / YELLOW / RED.
  This stays the PRIMARY filter; mechanism classification is secondary.

STEP 2 — Snap v1.1 trajectory
  For each trade with bars in kti window:
    Snap at offsets: -5, 0, +5, +10, +15, +25 min (5-min granularity).
    For each snap, extract:
      vcet_state, reason_code, vcet_score,
      vwap_slope_deg, band_width_pct,
      expansion_score, range_compression_score,
      CBO_signal, CBO_range_high, CBO_range_low,    (v1.1 NEW)
      vwap_distance_sd                              (v1.1 NEW)

STEP 3 — Apply mechanism rules in priority order
  C (wrong_side_escape)     — adversarial check first
  then 5 (true_chase_extended)  — if v1.1 vwap_distance_sd available
  then 6 (range_break)          — if CBO signal fires
  then 7 (rubberband)           — if v1.1 vwap_distance_sd against direction
  then 1 (true_pull_in_to_trend) — conversion signature
  then 3 (trend_continuation_entry) — already-trending entry
  then 2 (vwap_rotation_scalp)     — stayed in rotation
  fallback: INSUFFICIENT

STEP 4 — Cross-check vs outcome
  Tabulate net_pnl, win/loss, hold_time by mechanism class.
  Compute mean / median / WR per class.
  Statistical tests:
    - Mann-Whitney U: true_pull_in_to_trend vs vwap_rotation_scalp (P&L distributions)
    - Sign test: wrong_side_escape mean P&L <= 0
    - Chi-square: mechanism class × outcome (W/L)

STEP 5 — Adversarial v1.0 baseline
  Re-classify SAME trades, SAME trajectory snaps, but with v1.0 logic only:
    - Ignore CBO_signal (treat as null)
    - Ignore vwap_distance_sd (drop from rules 5, 6, 7)
    - Use v1.0 expansion_score / compression_score behaviour
  Produce v1.0_proposed_mechanism for each trade.

STEP 6 — Delta analysis
  Compute (v1.0 → v1.1) reclassification table.
  Test H8: which classification correlates better with P&L?

5. Acceptance criteria

v1.1 PASSES the acceptance test if all of A1-A4 hold AND at least 2 of A5-A6 hold. If any of A1-A4 fail, v1.1 is rejected and Quality Gate remains primary.

IDCriterionPass thresholdMeasurement
A1Coverage≥70% of trades with bars get non-INSUFFICIENT MEDIUM/HIGH classificationCount non-INSUFFICIENT calls / total bar-available trades
A2DiscriminationP&L distributions across mechanism classes are statistically distinct (Mann-Whitney p < 0.10 for top 3 most-populated classes)Pairwise rank tests on classes with n ≥ 5
A3Wrong-side precision≥90% of wrong_side_escape classifications have negative or near-zero P&L (no false-positive flagging of runners)Per-trade audit of flagged wrong-sides
A4v1.1 beats v1.0Spearman rank correlation of (mechanism class → P&L decile) improves by ≥0.10 from v1.0 baselineSpearman on (v1.0 classes, P&L) vs (v1.1 classes, P&L)
A5CBO utility≥5 trades classified as range_break_hitchhiker HIGH-confidence from CBO signal (i.e., v1.0 cannot make this call)Count CBO-driven HIGH calls
A6Trend-continuation discrimination≥5 trades flagged trend_continuation_entry; their P&L distribution is distinct from true_pull_in_to_trend (lower mean)Mean P&L comparison + count

6. Expected report output

One HTML report: VCET_V1_1_MECHANISM_ACCEPTANCE_RESULTS.html. Structure mirrors the existing audit pattern, with adversarial v1.0 baseline shown alongside.

7. Blockers / data gaps

BlockerSeverityWhat it blocksResolution path
B1. v1.1 engine not on diskHIGHEntire test cannot runObtain v1.1 source from upstream. Until then, this plan stays inert.
B2. kti 14-day bar retentionMEDIUMSample size on rare mechanisms (range_break, rubberband, true_chase)Accept partial coverage (A5 + A6 are CONDITIONAL — pass with fewer than threshold if direction is right). Re-run with bar archive when available.
B3. vwap_distance_sd not in trade_contextMEDIUMRules 5, 7 cannot use it from DB; must compute in-memory from bars per test runv1.1 engine should emit it; pull during trajectory snap. Do NOT persist to DB during the test.
B4. CBO algorithm details unknownMEDIUMCannot fully predict CBO precision/recall before test runsRead v1.1 spec when it arrives; pre-test the CBO signal on synthetic compression-breakout cases to validate it fires correctly.
B5. Mechanism rare-class statistical powerMEDIUMrubberband + range_break may not reach n ≥ 5 in 14-day windowAccept LOW statistical power on rare classes; weight A5 and A6 as confirming-not-required (already structured this way above).
B6. No ground-truth mechanism labelsLOWCannot compute precision/recall against external truthUse P&L outcome as the indirect ground truth (rule of thumb: true_pull_in_to_trend should be net positive, wrong_side_escape near zero, etc.). Adversarial v1.0 baseline (H8) is the rigorous test, not absolute precision.
B7. Trader behavior confoundLOWSome trades exit fast / scratch; mechanism may be ambiguous even with full dataReport hold-time alongside mechanism; treat sub-5min scratches as a separate "noise" class even if they pass another rule with LOW confidence.

8. What would prove v1.1 is useful

✓ Strong PASS — v1.1 earns a permanent role

Action if PASS: propose adding optional mechanism_slug column to trade_context (per Refinement Plan §3); compute it on bar-available trades only; surface in Scorecards drilldown as a "mechanism column" alongside canonical_strategy_slug; ADD to Quality Gate as the documented secondary management filter ("if VCET flags wrong_side_escape, exit").

✗ Strong FAIL — drop v1.1 plan, stay with Quality Gate

Action if FAIL: document the failure modes, keep Quality Gate as primary, postpone mechanism_slug indefinitely. Re-evaluate when either (a) bar archive lands and sample size grows 5×, (b) a v1.2 algorithm arrives with different signal generation, or (c) intraday VWAP geometry is added to trade_context independently.

⚠ Conditional PASS — useful but limited

Action if CONDITIONAL: deploy only the working mechanism classes as Quality Gate addenda (e.g., "VCET flags wrong-side → exit"); leave the noisy classes (rubberband, range_break) out of live rules; revisit when more data accumulates. Add a partial mechanism_slug with INSUFFICIENT as the default.

9. What this plan does NOT touch

10. Sequencing — when this plan can execute

  1. Now (this plan): document the test design, blockers, acceptance criteria
  2. When v1.1 source arrives: validate the engine on synthetic compression-breakout cases (5-10 hand-built scenarios); confirm CBO fires where expected; confirm vwap_distance_sd is in outputs
  3. After v1.1 validation: write vcet_v1_1_acceptance.py (kti-only, in-memory, no DB writes) — pull PI-family + chase-extended trades with bars, snap trajectory, classify, score
  4. After test run: produce VCET_V1_1_MECHANISM_ACCEPTANCE_RESULTS.html with the verdict
  5. After verdict: if PASS → propose ADR for mechanism_slug + companion field schema (per Refinement Plan §3); if FAIL → archive the test artifacts and return to Quality Gate as primary

Steps 2-5 do not execute until v1.1 source is delivered. This plan is the staging artifact; the actual test code is not written until the engine arrives.

Links