Stage A - kti 1-min Pilot Results (VCET v1.1)

Run date 2026-06-05 · Engine kite_v1.1.0 (canonical defaults: enable_cbo=True, require_cbo_for_expansion_confirmed=False, cbo_lookback=20) · Population scope: main.trade_context last 60 days, slugs = {vwap-pull-in, vwap-pull-in -> trend, chase-extended} · Anchor: Pull-In Research Hub · v1.1 Synthetic Acceptance

Stage A verdict: kti gives clean 1-min bars for recent trades, but a 60-day coverage gap is structural

The kti 1-min source produces high-quality bars (62/62 fetched pairs passed QA, 0 engine errors) but coverage degrades fast with age. June 2026 trades: 100% coverage. May 2026: 31.6%. April 2026: 0%. The 1-min VWAP recomputation matches the engine's 5-min VWAP cleanly (median diff 0.023%).

VCET v1.1 + CBO works end-to-end on real data: on the JOBY 2026-05-28 known true-conversion case, CBO fires at +10min and +25min in the trade direction, band width expands from 1.65 to 5.78, slope accelerates from +2.3deg to +51.8deg, state transitions to TREND_UP by +25min. This is the mechanism-classification primitive the v1.1 plan called for.

Recommendation: for any Stage B work covering more than the trailing ~14 days, kti alone is insufficient. See Liberator validation for the older-bar fallback and final source recommendation.

1. Coverage

205
trades in scope (60d)
62
computable (30.2%)
143
uncomputable
73 / 39
symbols total / w/bars

1.1 By slug

Canonical slugTotalComputableCoverage %
vwap-pull-in -> trend15960.0%
vwap-pull-in1324030.3%
chase-extended581322.4%
TOTAL2056230.2%

1.2 By month (the structural gap)

Trade monthTotalComputableCoverage %Status
2026-06 (current)2525100%Fully in kti window
2026-051173731.6%Mid-month onward covered; early-month outside window
2026-046300%Entirely outside kti 14-day window

1.3 Coverage by month + slug

SlugAprMayJun
vwap-pull-in38 / 079 / 2515 / 15
vwap-pull-in -> trend6 / 08 / 81 / 1
chase-extended19 / 030 / 49 / 9

Format: total / computable. The April column is the headline: zero coverage on 63 historical trades.

2. 1-minute bar quality

2.1 Bar QA (all 62 OK fetches passed)

CheckPairs failed
OHLC sanity (high >= max(o,c), low <= min(o,c))0
Non-negative volume, non-zero close0
Sufficient RTH bar count (>= 100 of 390)0
Engine error (e.g., pandas groupby/apply)0 (fixed)

2.2 Timezone handling

kti returns timestamps as naive "YYYY-MM-DDTHH:MM:SS" strings representing US/Eastern. The pilot localizes to US/Eastern then converts to UTC for engine consumption. Session anchoring is by date(US/Eastern). No timezone ambiguity observed; all sessions align cleanly to RTH 09:30-16:00 ET.

2.3 Pre-market / RTH coverage

Pilot requested 08:00 ET to 16:30 ET (extended window for pre-mkt context). kti returned variable counts depending on pre-mkt activity. RTH-only bar counts were 200-390 per session (390 = full 6.5h coverage). The pilot used a 100-bar minimum gate; no pairs failed it.

2.4 Split / corporate-action issues

Spot-checked: no obvious split discontinuities in the 62 computable pairs (no 50%/100% gaps mid-session). One pre-market outlier (NVDA 2026-06-02 entry 05:58 ET) produced empty trajectory snaps because the session-active gate (cum_v >= typical * 0.10) was not met yet pre-market. This is a v1.0-inherited behavior; pre-market trades are not engine-classifiable in the current implementation.

3. VWAP / SD recomputation cross-check

6,264
comparison points
0.023%
median pct diff
0.54%
p99 pct diff
3.13%
max pct diff (outlier)

Compared the 1-min-derived session VWAP (recomputed in the pilot) against the engine's 5-min-aggregated VWAP at every 5-min grid point across all 62 computable pairs. Median 0.023% diff is well within tolerance for mechanism work; p99 0.54% is acceptable; the 3.13% max outlier is likely a thinly-traded early bar where pre-mkt activity skews the 5-min OHLC aggregation. 1,639 of 6,264 (26%) exceed 0.05% - but the bulk of those are sub-0.1% drift, not consequential for state classification. See VWAP_SD_RECOMPUTE_METHOD for the method spec and a deeper diff breakdown.

4. VCET v1.1 + CBO field check on real bars

For each computable trade, the pilot snapped 13 v1.1 fields at {-5, 0, +5, +10, +15, +25} minutes around entry. All fields populate correctly on real data.

4.1 Reference trajectory: JOBY 2026-05-28 09:31:42 LONG +$12.45 (slug: vwap-pull-in -> trend)

This is the known true-conversion case from PI_TREND_ROTATION_WINNERS_REVIEW. The v1.1 engine now confirms the mechanism with CBO firing post-entry in trade direction:

t (min)stateslopebw%compexpcbo_break_longcbo_strengthscore
-5ROTATION+2.31.65025false0.0070
0ROTATION+20.52.20060false0.0085
+5ROTATION+34.44.37075false0.0090
+10ROTATION+45.95.40075TRUE0.06395
+15ROTATION+51.85.57035false0.0095
+25TREND_UP+43.85.78020TRUE0.086100

Mechanism signature confirmed: ROTATION at entry; CBO breaks LONG at +10 (channel-confirmed structural breakout); band width doubles (1.65 → 5.78); slope accelerates +2.3deg → +51.8deg in trade direction; state transitions to TREND_UP by +25min. This is the exact pattern the v1.1 mechanism validation plan §3-4 calls true_pull_in_to_trend.

5. Uncomputable trades

All 143 uncomputable trades failed for the SAME reason: NO_BARS returned by kti (the symbol/date pair is outside kti's effective bar retention window). No engine errors, no QA failures. The full list is at UNCOMPUTABLE_TRADES_LIST.csv.

ReasonnNotes
NO_BARS143kti has no bars for that (symbol, date). Trades older than ~14-30 days; entire Apr 2026 cohort.

6. Mechanism-classification readiness

The v1.1 engine emits all primitives required by the Mechanism Acceptance Test Plan. For each of the 7 mechanism classes:

Mechanism classv1.1 primitive(s) emitted on real barsStage A status
true_pull_in_to_trendROTATION/COMPRESSION entry state; CBO break post-entry; band-width expansion; slope acceleration in directionVERIFIED on JOBY 5/28
vwap_rotation_scalpROTATION sustained through +25; cbo_inside_channel=True; flat slope; flat band widthVERIFIED on DAL 5/26-5/28 cohort
trend_continuation_entrycbo_break_* fired BEFORE entry in trade direction; state already TREND_* at entryVERIFIED via cbo_break_* fields with -5 snap
wrong_side_escapecbo_break_* fires OPPOSITE trade direction within +15VERIFIED via cbo_break_* directional fields
true_chase_extendedRequires vwap_distance_sd at entry >= 2.5; engine emits price_distance_from_vwap_sd already (v1.0)Field present; threshold logic for chase is downstream
range_break / hitchhikerCBO confirms range break; channel_high/low provide pre-entry channel positionAll fields emitted
rubberband / backsideRequires price_distance_from_vwap_sd <= -2.5 AGAINST direction; MEAN_REVERSION stateFields present; classifier logic is downstream

All required primitives are emitted per bar at the trajectory snap points. The downstream mechanism classifier (per the Validation Plan §4) consumes these without additional engine work.

7. Final Stage A recommendation

Stage B should NOT proceed with kti-only

kti gives 100% coverage on June 2026 trades and 31.6% on May 2026, but 0% on April 2026. The full 60-day window (which contains ~70% of the recent PI-family + chase-extended trades) cannot be analyzed via kti alone. A mechanism study at the bucket level would have a heavy recency bias and only the most recent ~9 PI->Trend trades would be classifiable (vs the full 15 in scope).

Three options for Stage B:

  1. kti only - REJECTED. 30% coverage is insufficient for cross-bucket statistical work.
  2. Liberator only - PROVISIONALLY VIABLE. See Liberator validation report; sample 4 older trades all returned 100% RTH coverage. If Liberator parity passes, it covers the full historical window.
  3. Hybrid (Liberator for older + kti for recent) - RECOMMENDED PROVISIONALLY. Use kti for trades within the last ~14 days (already tested clean here) and Liberator for older. Requires that Liberator's VWAP/SD computation matches kti's within tolerance.

Pass to Stage B is GATED on Liberator parity validation. See the companion report for the next step.

8. Files produced

FilePurpose
Z:\public_html\system\KTI_96_TRADE_PILOT_RESULTS.htmlThis report
Z:\public_html\system\ENTRY_FEATURE_SAMPLE_OUTPUT.csv62 computable trades x 78 columns (trade meta + 13 v1.1 fields x 6 trajectory offsets)
Z:\public_html\system\UNCOMPUTABLE_TRADES_LIST.csv143 uncomputable trades with reason
C:\Users\jtdom\Downloads\vcet_v1_1\python\stage_a_pilot.pyPilot source (read-only, no DB writes)
C:\Users\jtdom\AppData\Local\Temp\stage_a\stage_a_summary.jsonRaw summary stats JSON
C:\Users\jtdom\AppData\Local\Temp\stage_a\kti_minute_cache.jsonkti bar cache (5 MB, ~62 pair-days)

9. What this pilot did NOT do

Links