A finding is an answer. An investigation is the question, and the work — which means it exists before there is anything to conclude, and it can close without ever producing an answer. Recording only the inquiries that found something would report a filtered history, and the filter runs in exactly the direction that flatters the platform.
| Open | Concluded with a finding | Closed with none | Trials spent, no conclusion |
|---|---|---|---|
| 1 | 3 | 2 | 6 |
The last column is the one that matters arithmetically. Deflation counts every configuration that was tried, not every configuration that worked, so trials spent on an inquiry that concluded nothing still raise the bar every later result in that lineage has to clear. An unrecorded search does not remove those trials — it only removes the platform's knowledge of them.
Do any of the ensemble regime features derive, directly or through a rolling transform, from observations later than the point at which the regime label is applied?
The chronological split with embargo closed the obvious leak at training time. It does not rule out a feature that is itself computed with lookahead, which would survive every split-level check because the leak is inside the feature.
methodology/hrp@3
Does cross-sectional momentum's out-of-sample performance improve materially at lookbacks other than the 252-session window inherited from Jegadeesh & Titman?
The lookback was adopted from the source paper without being tested on this universe. An untested inherited parameter is an assumption wearing a citation.
assumption/skip-most-recent-period@1 · evidence/jegadeesh-titman-1993@1 · methodology/xsmom@1
No finding. The effect looked for is not there — a result, not an absence of one.
Six lookbacks were evaluated (63, 126, 189, 252, 315, 378 sessions). None produced a deflated Sharpe distinguishable from the 252-session baseline once the six trials were counted, which is the outcome the deflation is designed to produce: the apparent spread across lookbacks is within what six attempts on the same data would generate by chance. Nothing changed, and that is the result. The six trials are recorded here and count toward every subsequent DSR on this lineage — an unrecorded sweep would have left the next published Sharpe overstated by exactly the amount this search inflated it.
Do the `rules` and `universe_filters` a methodology declares actually constrain what the executor does, or are they documentation that happens to be hashed?
Opened as a deliberate audit rather than in response to a defect. Every previous erratum had been a hidden choice that turned out to matter; this asked the inverse question — whether any visible choice turned out not to.
evidence/inert-declarations-2026@1 · methodology/hrp@3 · methodology/xsmom@1
Concluded with a finding. Rules now carry `enforced_by`, the contract verifier refuses a declaration with no realization, and `unrealized_declaration` is a hard publication blocker. The audit found the failure mode it went looking for, which is the argument for running it again on the next artifact type.
Is the flat 10bps transaction cost in protocol/standard@1 an adequate stand-in for realized costs across this universe, or does it systematically flatter instruments with wider spreads?
A flat cost applied to a universe spanning treasury ETFs and crypto is unlikely to be right for both. The concern was raised when execution lag and costs were first introduced and was deferred rather than answered.
assumption/one-session-execution-lag@1 · methodology/hrp@3 · methodology/xsmom@1 · protocol/standard@1
No finding. The evidence available could not settle the question, which is different from concluding there is nothing there.
Could not be settled with the data available. Answering it requires instrument-level historical spread and volume data that this platform does not hold, and substituting a modelled spread would have replaced one undeclared assumption with another wearing more decimal places. Recorded as inconclusive rather than closed. The 10bps figure stands as a declared assumption whose adequacy is explicitly unverified — which is a weaker position than a validated cost model and a much stronger one than an unexamined constant.
Does joining equity and crypto price series into one panel introduce sessions that the equity instruments never traded, and if so does it change published annualized figures?
A reviewer asked what `periods_per_year` meant for a panel spanning two markets. The question had no answer in the code, which is usually a sign the choice was never made.
assumption/nyse-sessions@1 · calendar/crypto@1 · calendar/nyse@1 · methodology/hrp@3
Concluded with a finding and an erratum. Trading calendars became addressable artifacts and protocols now name one, so the annualization basis is declared rather than inferred from whatever the panel happened to contain.
Why does methodology/hrp@1 report a Sharpe ratio of 6.59 on a diversified multi-asset universe, when published multi-asset risk-parity results cluster an order of magnitude below that?
The figure was not challenged by any check in place at the time. It passed every test, satisfied its output contract, and was arithmetically correct. It was opened because the number was implausible, which is a judgement no automated gate was making.
assumption/sample-covariance@1 · evidence/lopez-de-prado-2016@1 · methodology/hrp@1 · methodology/hrp@2
Concluded with a finding. The inquiry also produced the degeneracy diagnostics now applied to every result — concentration, degenerate volatility, implausible Sharpe and effective breadth — so the next instance is caught by a gate rather than by someone noticing.
Every other artifact here describes something that happened. An investigation is the only one that can describe something that was tried and did not work out, which is why it had to be its own type rather than a status on a finding: a finding can record that something was concluded, but has no way to say an inquiry ran carefully and concluded the effect is not there.
Two failure modes become visible as a result. Silent abandonment — configurations tried and dropped without record, each one still a trial for deflation purposes. And conclusions with no inquiry — a finding asserting that work was done without saying what work. The first inflates every Sharpe ratio in a lineage; the second is a declaration with no realization, which this platform treats as a defect everywhere else.