Did the backtest's state ever see the future?

Real market data, pinned

The daily research factors from the Ken French Data Library: market, size, value and the risk-free rate. 1,259 trading days, 2 January 2018 to 30 December 2022. The vendor's own archive ships inside the kit, fixed by its SHA-256 (af8aec07…), fetched 8 July 2026. The vendor restates its data monthly; this pin is the record.

The dates were taken in strictly increasing order. The history went in the way it happened.

One state per day, built only from the days before it

At each date the state is the geometry of everything recorded up to that date: a covariance of the four factors, made invertible with a small fixed ridge, g = (Σ + 0.01·I)⁻¹. A backtest that reads these states at each date reads only what was known on that date, if each state really was built from its own past.

That "if" is the whole question. It is answered by rebuilding every state from its own past and comparing bytes.

All 1,258 states rebuild exactly

For each of the 1,258 dates, the kit takes the record up to that date, and only that, and rebuilds the state. Each rebuilt state's digest equals the digest the history recorded, with no tolerance: byte for byte. The last one is be07f371….

A brake test the receipt has to fail

A test that cannot fail proves nothing, so the kit commits the crime itself. It takes 16 March 2020, the worst day in the record's future (market factor −12.01), and folds it into the state for 31 December 2019, index 502.

The digests break at index 502 and at all 756 indices after it. The state moves by max|Δg| 0.2808, about 1.69× the record's own floor of 0.1657, which is what re-entering a day the state already held would cost. Had the history peeked, the green above was impossible.

What else the kit checks, and says plainly

An ordinary future day. A statistically unremarkable later day (20 October 2022) also breaks the bytes at every date after it: detection is byte by byte. But it moves the state by about 0.032× the floor, inside the record's own noise. It changes bytes, not substance, and the kit says so beside the headline.

A cleaned-up history. Drop the five worst days before the peek, after the fact, and the chain fails to rebuild from index 23, the first dropped day, onward. A history tidied up later cannot forge this receipt.

One small file anyone can check

The published receipt records the question, the verdict and the two digests: the final state's, and the kit's own. It carries its own SHA-256, so a changed character shows.

{
  "case": "point-in-time",
  "question": "Did the backtest’s state ever see the future?",
  "domain": "quantitative research",
  "register": "detection",
  "status": "frontier",
  "verdict": "valid",
  "geometry_digest_v3": "be07f3713ccc406cfd6e880a6b92dadedf6950f503e326ebc2bdfd9f897b2472",
  "kit_sha256": "982ec9f9ca86bc70c9a93f7056f6a67c4e4e05dea39015d511bb9c739db4aeaf",
  "page": "/point-in-time/",
  "note": "Public register only — the verdict and the two digests already shown on this case. This receipt's own SHA-256 is recomputable here in your browser; the geometry g-digest is recomputable from the kit on your bench. No engine mechanism is disclosed, and nothing here is asked to be taken on trust."
}

Receipt SHA-256 7825b964e9b045eb4f33498fd5be32f2cb4759fc9b078f632c3fb934494c88f6

In Habitat's shared receipt format, the same result is a second receipt that also names what it read (the kit and the vendor's file, each by its SHA-256), the window of dates, and every measure above. Its digest is e6fbe490e840a56329e9d3089b4e2403be03a676352456505ab28643081ec56a.

The receipt opened and checked, the data never sent

The fund runs the kit on its own machine. Its allocator opens the receipt on Habitat Desk, which shows the verdict and checks the receipt in the browser. Desk never accepts the data itself.

The limits, stated

• It certifies the state chain, the covariance a strategy reads. A signal computed outside that state can still peek.
• Red detects a peek. It does not say which day peeked.
• It is a snapshot: this record, these pins. It says nothing about the strategy's future conduct.
• It is about a statistic. It says nothing about any model's weights.
• It is a frontier receipt: the history has to be built on the substrate from the start. A history assembled elsewhere cannot be certified after the fact.