The first two IONQ matches use ten stored observations, but those observations span 13 and 22 minutes. Counting bars and measuring elapsed time answer different questions.
The nearest two matches to IONQ's noon chart both contain ten stored bars. That sounds like an equal-length comparison. Their timestamps show otherwise.
HYFM, original rank 1, has a cutoff of 2:04 p.m. Eastern on March 25, 2021. KOD, original rank 2, stops at 3:19 p.m. Eastern on May 11, 2020. IONQ's cutoff is noon on May 12, 2026. We are using the same frozen sample as the first Casebook article, with no new retrieval and no prices after those cutoffs.
Two ways to count the window
| Pre-cutoff measurement | IONQ seed | HYFM, rank 1 | KOD, rank 2 |
|---|---|---|---|
| Stored observations used by the matcher | 10 | 10 | 10 |
| Clock minutes covered by those observations | 10 | 13 | 22 |
| Observations inside the final ten clock minutes | 10 | 7 | 5 |
| Observations inside the displayed hour | 60 of 60 | 34 of 60 | 27 of 60 |
| Change across the ten stored bars | -1.41% | -2.92% | -1.79% |
For HYFM, three of the matched observations lie before the final ten clock minutes. For KOD, five do. The matcher reaches further back to collect the same number of observations. Compared with IONQ, the represented intervals cover 1.3 and 2.2 times as much elapsed time.
That is a difference in the evidence going into the comparison. Ten stored bars do not automatically mean ten uninterrupted minutes.
Why the blank cells matter
The original measurement export leaves the ten-clock-minute return blank for HYFM and KOD. Their complete ten-minute paths are unavailable under this review's completeness rule. Those blanks are not zero returns.
We can describe the change from the first open to the last close of the ten stored observations. We should not silently rename it a ten-minute move, fill the missing minutes with flat prices, or compress the gaps out of the chart. Dividing the move by elapsed minutes would not recover the missing path either.
The wider hour adds another limitation. HYFM is missing 26 expected minute rows, and KOD is missing 33. Their observed bars show late weakness, but a smooth line through every gap would assert continuity the source does not provide.
Missing rows do not establish a trading halt. They also do not establish that no trades occurred. This examination identifies absent observations in the captured source; it does not determine their cause.
Keep the examples, qualify the claim
Both matches remain in the published 50-member set, at their original ranks. Removing inconvenient cases after seeing them would hide a property of the retrieval. Their presence is useful: it shows that elapsed time and observation coverage need to be read alongside similarity rank.
The stored feature vectors reproduced from the captured inputs in the original audit. That check and the timing problem can both be true. Reconstructing what a model consumed does not prove that every input represented the same amount of market time.
In the full set, 41 of 50 matches have a complete ten-minute window. The other nine require a different reading. This article examines the first two ranks because they make the clock mismatch concrete, not because of anything that happened afterward.
The next step is to specify which time basis a future comparison requires and how incomplete observations will be handled before collecting a new sample. This frozen set stays unchanged. We have not tested whether either timing convention predicts a better trading outcome.
Inspect the original 50-chart gallery. Check every count and measurement in the original CSV.
Figures
Selection and evidence
Follow-up to CB-001 using original ranks 1 and 2, with the entire frozen membership retained. The timing and coverage comparison uses only the published input measurements and pre-cutoff figures; no later performance informed selection.
Source snapshot SHA-256: f6c1b05d94656906c23484c33be9d06e0093940756a8099fecebdd860b25178c. Original protocol and reconstruction receipt.
