20 days of OHLCV
Ticker and calendar date are removed. Relative days, OHLCV, and derived features remain.
Recovery98 / 99
TargetTicker + date
Conditioned null0.011%
QuestionCan public market data link a masked numeric packet back to its ticker and date?
MethodFreeze the released representation, candidate database, distance rule, and confirmatory split.
ResultKTD-style and BlindTrade packets reach 98/99 and 94/99 joint recovery.
BoundaryLinkability is not the same as model memory, spontaneous discovery, or financial value.
Each protocol removes the obvious label but leaves a different numeric fingerprint. Only the KTD-style view below uses a frozen numeric packet. The other two are deliberately schema-only.
Ticker and calendar date are removed. Relative days, OHLCV, and derived features remain.
Recovery98 / 99
TargetTicker + date
Conditioned null0.011%
One frozen packet is matched against a frozen cross-provider candidate list. Select a case, reveal a prefix, then change the retained clue. Distance is a matcher output, not confidence.
Answer keyKMI · 2025-01-08
Exact row rank#1
The exact answer-key row ranks first.
Keep the estimands separate. The 5- and 10-day buttons truncate this same packet. They are not the paper's separately sampled exposure experiments.
The nonprimary 5-day length uses n=24. The 20-day primary cell uses n=99. The paper does not print the 80-day sample size beside the figure.
The one-day vector is closest to one source-system step. The 20-day result measures accumulated exposure.
Rebasing is a direct intervention, not a new attack. The left side transforms one frozen packet. The right side reports a separate sealed ablation over 99 packets.
Read the intervention.Rebasing collapses price-level recovery from 99.0% to 5.1%, yet returns and normalized close remain at 100% and OHLC shape remains at 99.0%.
Choose one of the 18 measured protocol, universe, and horizon cells. Every cell recovered the exact ticker and start date for all 100 packets.
KTD returns · U100 / H2100 / 100Unique exact nearest neighbors
Candidate windows48,200Exact count printed in the paper.
No invented counts. The paper prints exact index sizes only for the smallest KTD cell and the two largest cells. Other selections say “Not printed” rather than estimating a value.
In a separate construction study, an agent receives raw market history and generic code execution, but no matcher, network, answer key, or project code.
KTD exposes a 20-day price and volume trajectory. BlindTrade exposes 20 ordered rows of nine indicators.
Different study, different claim. The round budgets differ by declared design. These held-out development packets are not the sealed representation split, and equal recovery does not compare model efficiency.
Move from raw paths to continuous summaries, then to train-fitted bins. Closed-set recovery falls sharply only after discretization.
Nineteen daily returns retain almost the full fingerprint.
Raw-signal improvement over the constant was not established in the 600-firm primary gate.
Absent stocks are still falsely linked under R3.The frozen abstainer has no reliable signal that the truth is missing. R3 is resistant under this closed-set ensemble, not certified anonymity.
What the utility study says. R2 uses six continuous summaries. R3 maps each summary to five train-fitted bins. In the 900-firm temporal replication, R3 beats the constant on a predeclared secondary endpoint: MSE gain .181, unadjusted 95% CI [.159, .204]. That does not rescue the failed primary composite.
Each normalized block appears under five randomized headers. Select one to inspect what the assay holds true or deliberately changes.
The complete, correct header condition.
Do not overread the null. Both intervals lie inside the predeclared ±.10 meaningful-effect region. The task elicited little forecasting skill, so this is not a universal null on financial value.