The Auditable Track Record & Honors Review

A track record is a claim, and a claim is worth exactly as much as your ability to check it.

This is the whole of your capstone in one sentence. Across this institute you have learned to ask, again and again, when a number is genuine evidence that an asset will move rather than noise dressed up as signal. Now you turn that same skepticism on yourself. You are about to assert that your thesis worked — that you predicted something, that the prediction came true, and that this was skill and not luck. The person reading that assertion has every reason to doubt you, and they should. Your job in the honors review is not to be believed. It is to make belief unnecessary, by handing over something they can verify for themselves.

That something has a name in computer science, and we borrow it deliberately. It is called a proof-of-work credential. The phrase is precise. A credential is a token that vouches for a claim. Proof-of-work means the token is backed by effort that is hard to fake and cheap to check. The classic example is mining a block in a blockchain: finding the answer takes enormous computation, but anyone can confirm the answer is correct in a single step. The asymmetry is the entire point. Hard to forge, trivial to verify. That asymmetry is what we want for a research track record, and almost no track record in the real world has it.

So let us build intuition for why ordinary track records fail before we make ours pass.

Imagine you let someone trade in private for a year and then, at the end, asked them to show you their best ten trades. They would have a spectacular record — because they chose which trades to show you after seeing the outcomes. This is survivorship and selection bias, and it is the original sin of performance claims. The numbers are real; the inference is worthless, because the selection happened with knowledge of the answer. The fix is not better statistics applied afterward. No amount of cleverness recovers honest evidence from a record that was edited with foresight. The fix has to happen before the outcome is known.

That fix is pre-registration. You commit to a falsifiable prediction, in writing, time-stamped, before the market reveals the answer. You name the asset, the direction, the horizon, and the threshold that would count as right and the threshold that would count as wrong. Then you seal it. Once sealed, it cannot be quietly removed if it fails or sharpened if it nearly works. This is exactly the discipline you met in chapter one when you wrote your thesis as something the world could prove false. Pre-registration is what turns a falsifiable thesis into a falsifiable record.

Now the mathematics of trust becomes simple. A sealed prediction is a commitment. In cryptographic terms, you publish a hash of your prediction — a short fingerprint that any change to the prediction would alter completely — so the prediction is locked the instant it is made, even before anyone reads it. When the horizon arrives, the realized outcome is pulled from the same price feed everyone shares, the one you have trusted since the foundations course. The verifier then re-runs one deterministic comparison: did the sealed prediction, scored against the realized outcome by the rule you fixed in advance, hit or miss? There is no judgment in that step. It is arithmetic. Given the sealed claim and the public outcome, the score is forced. That is what makes it a proof and not a testimonial.

The credential, then, is not a badge someone awards you because your work looked impressive. It is a statement that a specific chain re-derives to the same answer every time it is run: these predictions, sealed at these timestamps, scored by these pre-committed rules against this public outcome data, yield this hit rate — and a hit rate this strong, over this many independent pre-registered calls, would arise by chance with probability below the honors threshold. The word independent is doing real work there: the chance probability is computed as if each call were a separate coin flip, so the calls have to be approximately independent bets — distinct assets, non-overlapping horizons — for that number to mean anything. Stack ten correlated calls (the same name ten times, or one asset at horizons that overlap) and you have not made ten bets; you have made one or two, dressed up as ten, and the true probability of your run is far larger than the naive count suggests. Spell that probability out plainly. If a coin-flipping novice would reproduce your record one time in a thousand, your record is evidence. If they would reproduce it one time in three, it is a coin flip with good marketing.

Notice what this rules out, because the exclusions are the value. It rules out editing losers out of the history, because the history is append-only and sealed. It rules out moving the goalposts, because the scoring rule was fixed before the outcome. It rules out cherry-picking your best window, because every pre-registered call counts, including the ones you would rather forget. The honors review does not reward the prettiest story. It rewards the record that survives an adversary who is trying to find the trick.