Stage A - Liberator/CloudQuant Source Validation

Run date 2026-06-05 · Companion to kti pilot results · Dataset: minute_bars on austin environment · Query interface: kti liberator query --dataset minute_bars --date YYYY-MM-DD --symbols X · Anchor: Pull-In Research Hub

Verdict (Stage A initial): Liberator is OHLC-equivalent to kti for overlapping dates; covers older trades cleanly

Direct comparison on NVDA 2026-06-02 (a date inside the kti window so both sources are available): OHLC matches exactly (max close diff = 0.0000, median = 0.0000 across 390 RTH bars). Volume differs by a small fraction (~5K shares median per minute, likely off-exchange/odd-lot filtering differences). Bar count, timezone, and structure are equivalent.

For 4 older-trade samples (AAOI 4/10, CAR 4/13, WULF 4/13, CRWV 4/13) - all of which kti returned NO_BARS for - Liberator returned 100% RTH coverage (390/390 bars each). The historical bar gap that blocks kti-only Stage B is closed by Liberator.

Recommendation: Adopt Hybrid (Liberator for >~14d old; kti for recent) as the Stage B source strategy. Liberator-only is also viable if a single-source pipeline is preferred. Either way, Stage B can proceed once the data source is selected and a small full-population recompute is run.

⚠ Stage A.1 update (2026-06-07): Liberator entitlement was revoked

The 20-pair parity test could not run because Liberator's minute_bars dataset is no longer in the user's entitlement list. Entitled datasets dropped from 27 (during Stage A initial pilot at 15:00 ET on 2026-06-05) to 10 (at 19:55 ET on 2026-06-07). The remaining 10 are fis_* (lending), s3_partners, and wonda - no market_data.

The single-pair parity result above (NVDA 6/2) remains valid; it was cached before entitlement changed. But the 20-pair statistical baseline the user requested cannot be produced today.

Stage B is BLOCKED on Liberator entitlement restoration. See the 20-pair parity report for the full blocker analysis and the queued retry plan.

1. Sample coverage tests

SymbolDateTotal barsRTH bars (of 390)kti statusLiberator status
NVDA2026-06-02390 / 960390 (100%)OKOK (overlapping; direct compare)
AAOI2026-04-10957390 (100%)NO_BARS (pre-window)OK
CAR2026-04-13760390 (100%)NO_BARSOK
WULF2026-04-13838390 (100%)NO_BARSOK
CRWV2026-04-13960390 (100%)NO_BARSOK

Liberator's total-bar counts include extended hours (~04:00 to 20:00 ET). RTH-only filtering recovers the standard 390-bar session window cleanly.

2. Direct comparison: kti vs Liberator on NVDA 2026-06-02

Both sources were queried for the same (symbol, date). RTH-window filtering applied to both. 390 RTH bars in each source, matched by minute timestamp:

0.0000
max close diff ($)
0.0000
median close diff ($)
5,251
median volume diff (shares)
9,220
mean volume diff (shares)

2.1 Spot-check on 3 sample bars

Time (ET)SourceOpenHighLowCloseVolumeVWAP (provided)
09:30:00kti227.19227.55225.47225.563,844,202-
09:30:00Liberator227.19227.55225.47225.554,074,658226.80
10:00:00kti231.42231.97231.18231.26735,041-
10:00:00Liberator231.42231.97231.18231.26762,631228.41
15:55:00kti222.48222.67222.38222.48712,079-
15:55:00Liberator222.48222.67222.38222.48727,290226.17

OHLC identical on every spot-check. Volume off by 1-6% per bar, attributable to off-exchange / odd-lot inclusion differences between the karb price server (kti) and CloudQuant. For VCET features that depend on price (band geometry, slope, CBO channels, range break), the two sources are equivalent. For features that depend on volume (RVOL, expansion_score's RVOL component, volume_expansion_score), there may be ~5-10% derivative drift but the directional signals are preserved.

3. Liberator field structure

Each Liberator minute bar carries 20 fields (kti's minute mode carries 8, minute_expanded mode carries ~15). Liberator's extra fields are research-relevant:

FieldTypeNotes
open, high, low, closefloatStandard OHLC; matches kti exactly
volumeintSlightly higher than kti (off-exchange inclusion)
vwapfloatBar-level VWAP; can cross-check recomputation
bvwapfloatBid-side VWAP; useful for spread analysis
spreadfloatMean bid-ask spread during the bar
bidvol, askvolintBid-side / ask-side volume split
countintTrade count in the bar
avgdeltafloatMean delta (bid lift vs ask lift signal)
timestampstrEnd-of-bar (e.g., "09:30:59.999"); Eastern time
_seq, _dsname, msg, msg_len, muts, length, symbolmiscLiberator internal metadata; ignore for VCET

3.1 Timestamp convention difference

Liberator timestamps are end of bar with millisecond precision (e.g., "2026-06-02 09:30:59.999"). kti timestamps are start of bar ("2026-06-02T09:30:00"). Both represent the same 60-second window. The Liberator loader in any Stage B pipeline must subtract 60 seconds (or equivalently, set time to bar-start) to align with kti and the v1.1 engine's expectations.

3.2 Pre-market and post-market

Liberator returns bars from ~04:00 to ~20:00 ET (16 hours). kti's pilot window (08:00-16:30) returns less. For VCET classification, the session-active gate filters pre-mkt by volume threshold; both sources work. Liberator's wider window gives more pre-mkt context if needed (e.g., overnight gap analysis).

4. Source comparison summary

Dimensionkti (karb price server)Liberator (CloudQuant)
Coverage window~14 days (varies; last ~30d partial)Full history (validated April 2026 cleanly)
OHLC qualityExcellentIdentical to kti on overlapping dates
VolumeExchange-only (lower)Off-exchange + exchange (higher; ~5-10% more)
RTH coverage100% when in window100% on all 4 older samples
Bar-level VWAPOptional (--expanded)Always included
Bid/ask, spread, countOptional (--expanded)Always included
Timestamp conventionStart-of-bar, naive ETEnd-of-bar (.999), naive ET
Query batch limit~30 calls/min (--wait auto-retries)No observed rate limit on small queries
Response size~5 KB per (symbol, date)~400-500 KB per (symbol, date)
AuthenticationLocal configLocal config (~/.kti/config.yaml)

5. Stage B source recommendation

Option A: Hybrid (Liberator older + kti recent) - RECOMMENDED

Option B: Liberator-only - PROVISIONALLY VIABLE

Option C: kti-only - REJECTED

Option D: Defer until better coverage exists - NOT RECOMMENDED

6. Open work to complete Stage A

  1. Volume-drift characterization. The 5-10% volume difference between kti and Liberator needs to be measured on more pairs (not just NVDA 6/2). A 20-pair Liberator/kti overlap study would establish whether the drift is uniform (constant fraction = safe to use either) or symbol-specific (per-symbol calibration needed).
  2. Liberator timestamp normalization. The Stage B pipeline needs a thin loader that converts Liberator's end-of-bar ".999" timestamps to start-of-bar UTC timestamps matching kti and v1.1 engine expectations.
  3. VWAP cross-check using Liberator's bar-level vwap field. Compare the engine's recomputed VWAP vs Liberator's bar-level VWAP (Liberator gives per-bar VWAP for free). If they agree, that's stronger confidence than the 1-min vs 5-min self-consistency check in VWAP_SD_RECOMPUTE_METHOD.
  4. Pre-market behavior. Liberator covers 04:00-20:00 ET; the pilot's NVDA 6/2 entry at 05:58 ET produced empty trajectory because the engine session-active gate didn't fire. Decide whether pre-market trades should be classifiable (requires engine update) or excluded (label as insufficient_data_premkt).

7. What this validation did NOT do

Links