Notes & essays
Essays and short notes on measurement — how to decide what to measure, how to tell whether the number that comes back is real, and what to do when it is smaller than you hoped. Mostly hearing and markets, occasionally neither.
- The removal of cognitive muscle Why AI is better read as a change in the price of cognitive execution than as a machine that thinks →
- Dropping 0.61% of the data destroyed 96% of the profit Data integrity as the first filter, and four other ways a backtest lies with correct code →
- The close that arrives in one tick Three days of tick data on a market-structure change: what decayed, what is tradeable, and what the data refuses to say →
- Correcting for how many strategies you tested is a correction to your biography Effective degrees of freedom, mechanism before fit, and the one division of labour I could not automate →
- An exit model that never sees an entry signal Building a test rig that grades exits independently of entries, and the model that failed on it →
- Four ways of asking "is this a bad day to trade" all failed. A fifth worked. Market-regime gates, self-referential gates, and the difference between abstention and selectivity →
- A 5× improvement in prediction bought exactly zero rupees What happens when you optimise the metric instead of the thing you trade →
- Ten independent results, one shared assumption, zero surviving evidence Why a claim is worth the weakest link it rests on, and why no single study can notice this →
- My search found 89 good signals and zero good strategies Cross-sectional IC, within-stock correlation, and a proxy that was uncorrelated with the objective →
- If the agent can write Python, give it something worse Non-Turing-complete DSLs, validation as a type error, and why expressive power is the thing you want to give up →
- Bhaav Bhagwan Che — sixteen months with one phrase How a slogan about model failure survived contact with the argument against it →
- The sophisticated method taught me the problem; conditionals solved it On matrix profiles, DTW, a published negative result, and why twenty labelled examples changes everything →
- Never hand a language model a number it could be handed as a fact Symbolic reduction as a correctness property, and the specific ways models fail on structured state →
- Before you detect a pattern, decide what counts as one What happens when you implement an informal visual convention exactly as written →
- Give the agent a shell and an API, not a prompt and a hope Statelessness, documented contracts, and why the boring interface is what makes agents swappable →
- The max pain curve is a bowl because of arithmetic, not because of manipulation What the U-shape actually proves, what would have to be measured to test the theory, and where gamma pinning differs →
- A regime classifier that flickers is worse than one that is wrong Why hard regime labels fail at boundaries, and what a membership vector buys you →
- Auditing your implementation against the textbook, one definition at a time A boring review format that catches errors no test suite is looking for →
- A trade can pass two gates and die on the third Regime fit, structural validity and path constraints are different rejections with different repairs →
- For a tool that advises an expert, a miss is free and a false alarm is fatal Why advisory detectors should optimise precision to the exclusion of nearly everything else →
- Write the questions before you build the dashboard Question banks as a specification, and what open interest analytics look like when you start from the question →
- The baseline you skipped is the result you needed On why the simple baseline wins, and the less comfortable reason it wins →
- The specification is the artefact the agent and I both edit Contract-driven agentic development, and why the review target matters more than the code →
- The same geometry, computed twice, in two languages Coordinate frames, index offsets, and why a strict port can be worse than the original →
- What even is edge Two theories of where returns come from, and which one survives contact with a cost model →
- The labels were the hard part Why "what should this row's target be" is a harder question than any model choice →
- When a 2018 note met a 2025 paper Pattern reuse from file systems to neural networks, with experiments →
- What keystrokes tell us about the mind Forty years of cognitive science hiding inside the inter-key interval →
- I spent months on a cost calculator before I had a strategy What rounding to the nearest rupee does to the space of viable trading strategies →
- I read for eight months and shipped one app On the failure mode of breadth-first research, from someone who ran into it at full speed →
- Five ways to ask the same question Procedure invariance fails predictably, and the RLHF stack inherited a UI nobody defended →
- A reward model that agrees with people about a third of the time On what "human preference" means when a scalar model claims to have learned it →
- The comparison graph was in ten pieces Why the analysis method is an experiment-design constraint, not a later decision →
- Re-deriving an ANOVA instead of arguing about it On the moment when re-deriving is cheaper than agreeing →
- Notes are a retrieval problem, not a writing problem What five years and a few hundred linked notes taught me about writing for a reader who has forgotten everything →
- Five years of not knowing the question What a dated record of your own confusion is good for →
- The test that ignores the logic of your experiment Why the standard test was the wrong instrument, and what I used instead →
- Eight metrics, one recording, no ground truth What happens when you point a shelf full of objective quality metrics at a signal none of them were calibrated on →
- The same metric, twice, disagreeing On the day I wrote "all comparison plots before this may be wrong" in my own notes →
- The feature didn't matter; the headphone did When the control variable outperforms the independent variable →
- The kink at −28 dB A feature in the data I could explain but not demonstrate →
- The button I pressed without telling anyone On measuring audition rather than expectation →
- ENV, TFS, and why the modulation filterbank is the right altar to pray at The decomposition that turned a vague question into a measurable one →
- ANC is not just for music A tour of the unglamorous places ANC shows up →
- Why ANC and the cochlea are accidentally well-matched A frequency-range coincidence that does most of the perceptual work →
- The one-line trick that makes a manipulation real Why scaling an envelope changes nothing, and a dc offset changes everything →
- Small effects, honestly reported On the difference between statistically significant and perceptually loud →
- Treating the headphone as a black box Why I evaluated the four ANC headphones without touching the algorithm →
- Why the amplitude envelope, and not the spectrum Picking one dimension of distortion and committing to it for a whole thesis →
- Using a brain model as a test instrument Inverting a computational neuroscience model into a hardware measurement device →
- The measurement nobody makes Why the ANC literature measures the noise and ignores the signal →
- I called it breadth-first. Five of ten students called it too fast The same failure mode, five years before I recognised it in my research →
- Dragons, pizza, and payoff matrices On assessing the step that comes before the mathematics →
- The question I put on every exam On teaching the one idea that survived into everything I do since →
- My students rated my logistics higher than my teaching What a ten-respondent course evaluation can and cannot tell you →