Five years of not knowing the question

Somewhere in the second year I started keeping a note called "trajectory of focus," which was a year-by-year log of what I was working on, what I was reading, and — this is the part that turned out to matter — what I was confused about at the time and in what specific terms. It was not written for anyone. It was written because I could not tell from the inside whether I was making progress.

Reading it back after finishing is a peculiar experience, because the log makes an argument that I could not have made from memory. Memory smooths a PhD into a narrative where you were working toward the thesis all along. The log says otherwise, with dates.

What the record actually says

2018. Started on "ANC effects" in August, framed as let's do some simple experiments. Learned the basics of cochlear mechanics. Got the first pilot data — active noise cancellation was shifting the threshold of a psychometric function. Received the response "this can be the first chapter of your thesis," which contained no information about what the other chapters would be or what the thesis was about. In a seminar that year, someone asked me what the slope parameter of my psychometric function represented. I did not have an answer and said something vague about task difficulty.

2019. Presented a proposal defence in which I was still calling the experiment by the wrong name, still describing a signal property as though it were an experimental paradigm, and still had no idea how many experiments a thesis requires. Discovered the term "sinusoidal amplitude modulation" in March and immediately worried about how deep the literature looked. Submitted a paper in August with a hastily written literature review that was, I can now say plainly, wrong. It was rejected. Wrote to my supervisor asking, in effect, how thorough an explanation of one's own results is actually required for a PhD.

2020. A committee meeting in January went badly enough that a faculty member asked whether I was having an existential crisis. I built minimum viable versions of four unrelated experimental paradigms and abandoned all four because I could not construct an argument for their relevance. In February, the note reads: the notion of "focus on AM only" dawns but cannot accept it yet.

That sentence is the hinge of the entire document, and it is dated six months before anything changed.

August 2020. The log records an epiphany and, from that point, a mode it calls focus. Three chapters, fixed: a computational-neuroscience chapter, an amplitude-modulation-basics chapter, a vocoded-speech chapter. Everything after that entry is execution — hard, slow, full of the metric-implementation problems I have written about elsewhere, but no longer searching.

The decision was subtractive

Here is what I did not expect to find. The thing that ended the flailing was not the acquisition of an idea. It was the abandonment of most of them.

Everything in the 2018–2020 stretch is additive. Read more, learn more paradigms, build more pilots, find more literature that might justify the direction. The reflex under uncertainty was always to widen. And widening felt like diligence — every abandoned paradigm was real work, honestly done, and each one taught me something.

The August 2020 change is the opposite move. It says: the axis is amplitude modulation, everything not on that axis is out, including things that are interesting and things you have already built. What made it possible was not new knowledge. Every fact needed for that decision was available in February, which is why February's note can state the conclusion and then decline to accept it.

Six months between knowing and deciding. I do not think those six months were wasted exactly — I think that is roughly what it costs to give up on the breadth you have accumulated — but I would like to have spent fewer of them.

Why the log itself was the instrument

I want to be precise about what the note did, because "keep a research journal" is generic advice and this is a specific claim.

The value was not in recording findings. Findings end up in papers and figures. The value was in recording the exact shape of my confusion, in the vocabulary I had at the time. Entries like cannot make head or tail of SAM and worried about the onslaught of parameter value choices are useless as science and diagnostic as data, because when you read them a year later you can see immediately whether the confusion has resolved, migrated, or simply been forgotten.

Three things followed from that.

It made unknown unknowns visible in retrospect. The hardest part of a literature review is that you cannot ask a question you lack the vocabulary for. My log shows me circling a concept for months using the wrong words before finding the right one, then noting the right one with a date. That gap between circling and naming is the actual duration of the problem, and it is invisible without a record.

It falsified the memory of steady progress. Without the log I would tell you the thesis took shape gradually. It did not. It was three years of undirected search and eighteen months of execution, and the transition happened in a specific month.

It let me stop asking people who could not help. One of the later takeaways in the note is blunt: a good rule of thumb is to stop asking questions of people once they have demonstrated they do not have useful advice. That conclusion is only reachable from a record. In the moment, every unhelpful conversation feels like your own failure to ask well. Across a year of logged conversations, the pattern separates from the noise.

The opinion

The most useful thing to write down during a long research project is not what you learned but what you did not understand and in what words — and the reason almost nobody does it is that those entries are humiliating to read back, which is precisely the property that makes them informative.

The polished narrative of any research project is constructed after the fact, and constructing it destroys the evidence about how the work actually proceeded. If you want to get faster at the searching phase — and the searching phase is where the years go, not the execution phase — the only data you will ever have is a contemporaneous record of being lost, written by someone who did not yet know which direction turned out to be right.

I have kept the same practice since, on work that has nothing to do with hearing. The entries are still embarrassing. The February-2020 pattern still shows up: I now recognise, with roughly a month's lag rather than six, the entry that states a conclusion I am not yet willing to act on.