MySleepMate / Research log
The working notebook
Research log: how this gets made
The articles on this site state conclusions. This page states how the conclusions were reached, including the wrong turns. It exists because the most useful thing I have found while building this site is not any individual sleep fact: it is the repeated discovery that confident numbers usually have much softer origins than their delivery suggests, and that finding this out requires going and looking rather than reading one more summary.
- Entries are dated and appended. Nothing here gets quietly rewritten.
- Errors stay visible. Corrections link to the corrections log.
- The method is described, including the AI part. See how this is made.
- This is a notebook, not advice. The articles are where conclusions live.
This is the notebook rather than the finished work. Every entry is a thing I set out to check, what I believed before I checked it, what I actually found, and where I was wrong. The mistakes are here on purpose. A site arguing that claims should be traceable to their sources cannot reasonably hide its own working.
Third-year medical student at Manipal Tata Medical College, writing about sleep research. A student, not yet a licensed clinician, so everything here is sourced, and none of it is medical advice. Why you should or should not trust me →
Started 2026-08-31 · Updated as work happens
Researched against primary sources, drafted with AI assistance, then reviewed and approved by the named author. How we write · Corrections log
How this is actually made
Being straight about the process, because a research log that hides its own method would be a strange thing.
I work with an AI assistant. It does the searching, the drafting and the wiring-together of pages. I do the verification: pulling the actual papers, reading them, and checking whether the thing being claimed is the thing the source says. That division is not a detail. It is the whole reason this site can make the claims it does, because it turns out to be exactly where AI-assisted research fails.
| Step | Who | Failure mode it introduces |
|---|---|---|
| Finding candidate sources | AI | Surfaces what is popular, which is often the secondary retelling rather than the paper |
| Reading the source | Me | this is the step that catches things |
| Drafting | AI | Writes confidently regardless of how firm the underlying claim is |
| Checking claims against the paper | Me | – |
| Publishing and correcting | Both | – |
Three of the first four errors on this site came from the drafting step being more confident than the evidence underneath it. None came from a fabricated citation, which is the failure people expect. They came from real papers being described with more certainty than they earn. That is a subtler problem and, I think, a more common one, not just in AI writing but in health writing generally.
If that disclosure makes you trust the site less, I would rather that than the alternative. Every page here carries the same line in its byline, and it has been there since before this log existed.
31 August: The first number on this site that is mine
Entry 6 · Nap calculator
Until today every number on this site came from somebody else's paper. This one is mine, and it changed a default on a live tool.
The nap calculator needs to know how long you take to fall asleep, because a 30-minute timer plus a 2-minute sleep onset is a 28-minute nap and lands you in deep sleep. I had it defaulting to 10 minutes, which I picked because it is the low end of the range textbooks quote. That is an assumption wearing the costume of a fact, which is the exact thing this site exists to complain about.
So I used my own diary: 28 consecutive nights, 1 to 28 February 2026. Lights-out recorded at the time, latency estimated the next morning, with caffeine, alcohol, screens, exercise and stress logged alongside. No nights dropped, including the bad ones, because a log with the bad nights removed is worse than no log.
Median 12 minutes, with 50% of observations between 7 and 19. Ninetieth percentile 31. Worst night 47. The default is now 12.
Three things I did not expect. The first is how wide it is: a single default is right for the middle of my nights and badly wrong for a fifth of them, which is the argument for the input box rather than a fixed figure. The second is that the two biggest subgroup gaps were stress (9 minutes against 24) and late caffeine (10 against 21), which is a hypothesis rather than a finding, because those nights overlap heavily with each other and the subgroups are small.
The third is the result I liked most and trust least. Nights with latency under 10 minutes were followed by mornings I rated 3.8 out of 5; nights over 20 minutes scored 2.7. It looks like falling asleep faster produces better mornings. It much more likely means that stress produces both, or simply that a long latency costs you sleep when the alarm does not move. A diary cannot separate those. I published it anyway, with the three readings laid out, because leaving out the finding I like most would be the wrong instinct for this site.
One correction already: I first gave the dates as 14 to 28 February while telling you it ran 28 nights, which are not the same fortnight. It was the whole month. Caught before it had been anywhere.
What it demonstrates, more than anything about sleep: the difference between an assumption and a measurement is a month of writing things down.
31 August: The pattern in my own errors
Entry 5 · Method
Four corrections in, a pattern is visible and worth writing down before I forget it.
My errors are not about facts. They are about the status of evidence.
- Two percentages attributed to a document that does not print them.
- A measurement point reported as a duration of effect.
- An absence of replication asserted without a literature search.
- "Beat 20 on every measure" where the paper concluded "most effective overall".
Every one of those involves a real study, correctly named, described in a way that overstates what it establishes. Not one is a made-up citation. If I had been checking for fabrication I would have found nothing wrong.
So the check that actually works is not "does this source exist" but "does this source say this, in this strength?" Existence is easy to verify and is not where the failures are. I am now reading the methods section before the abstract, because the abstract tells you what was found and the methods tell you what could have been found.
31 August: Claiming an absence without looking
Entry 4 · How long should a nap be
The nap page went live saying that Brooks & Lack (2006) "has not been replicated at scale."
Nobody had checked that. No literature search was run. It was written because it sounded like appropriate caution, a hedge, added to seem careful.
But it is not a hedge. It is a claim about the state of a literature, and it is a stronger claim than most of the things it was sitting next to. Saying a finding stands alone tells the reader something specific and load-bearing about how much weight to give it. I withdrew it entirely rather than replacing it with a softer version, and the page now says plainly that I have not surveyed replication and will not imply either way.
The thing I want to remember: caution that has not been checked is not caution. It reads as rigour while adding an unverified claim. That is worse than saying nothing.
31 August: When the clock stopped, not when the effect did
Entry 3 · How long should a nap be
Brooks & Lack compared 5, 10, 20 and 30-minute naps. The summary I was working from said benefits of the 10-minute nap were "maintained for as long as 155 minutes." The draft turned that into "benefits lasting over two hours."
Reading the paper, 155 minutes is the last time they measured. Post-nap assessments ran at 5, 35, 95 and 155 minutes. Some effects were still detectable at that final point, which means the study cannot say when they faded, only that they had not faded yet.
Those are different statements and the difference matters. "Lasts 155 minutes" implies an endpoint was observed. What was observed is that the endpoint had not arrived by the time everyone went home.
This one bothers me most of the four, because it is the kind of error that survives every check except reading the methods. The number was right. The sentence containing it was wrong.
The page now carries a callout warning readers against exactly the reading it previously invited, which felt like the honest way to fix it.
31 August: The percentage that came loose from its study
Entry 2 · Where sleep numbers come from
The best-known fact about napping is that NASA found a 26-minute nap improves performance by 34% and alertness by 54%. It is on every productivity blog. It was on this site.
I went to the technical memorandum, Crew Factors in Flight Operations IX, held by the US DOT National Transportation Library. Those two percentages are not in it.
What is in it: 21 pilots, a 40-minute scheduled rest window, a mean of about 25.8 minutes of sleep obtained inside it, and outcome counts, 34 microsleep events against 120 in the no-rest group across the final 90 minutes of flight, and 81 vigilance lapses against 124.
Rosekind has described the work in percentage terms since, and secondary coverage has carried that onward until it became the study's headline finding. The paper's own numbers, which are more concrete and in some ways more striking, disappeared.
Two things I took from this. The first is that "26-minute nap" was never a protocol: it is a measurement of what tired pilots happened to sleep inside a 40-minute window, rounded and reissued as a prescription. The second is broader: a number can be entirely real, correctly attributed to a real study by the study's own author, and still not be a finding that study published. There is no dishonesty anywhere in that chain. The number just came loose.
30 August: Google decided what this site was about before I did
Entry 1 · Strategy
Two weeks after launch, with 2,780 impressions and zero clicks, I went looking at Search Console for what was working. What I found reframed the whole site.
| Query | Average position |
|---|---|
| 90 minute sleep cycle | 79.5 |
| sleep cycles 90 minutes | 64.3 |
| american academy of sleep medicine sleep cycles 90 minutes vary | 8.1 |
| sleep cycle average 90 minutes authoritative source | 6.0 |
Same topic. Same page. Seventy-three positions between the top and bottom rows. The difference is that the queries where this site ranks well contain a word like source or name a specific body like the AASM.
I had built a sleep calculator and assumed the site was competing with other sleep calculators, which it would lose. What Google had actually decided is that this is a place people land when they want to know where a claim comes from. Nobody types "authoritative source" into Google casually. That is writers, students and journalists checking something before they publish it.
I want to be honest about the size of the signal: those two queries have three and seven impressions between them. It is tiny, and I could be reading far too much into it. But it pointed at a lane nobody else in this niche wants, because tracing your own numbers means occasionally finding out they do not hold, which is uncomfortable if your business is confident advice. Everything on the site since has been built on that read, and if three months of data says it was wrong, that will be an entry here too.
What I am checking next
Open questions, listed so that failing to answer them is visible.
- The coffee nap trial. Reyner and Horne (1997) is the study behind every coffee nap article, including mine. I have the figures from secondary descriptions rather than from the paper: 12 participants, 200 mg of caffeine, a 15-minute nap, incidents down to about 9% of placebo against 34% for caffeine alone. Given what happened with the NASA percentages, those numbers need checking against the original before I treat them as settled.
- Sleep latency, partly answered. "Ten to twenty minutes to fall asleep" is repeated everywhere and I still have not traced where it comes from. I do now have 28 nights of my own, which puts my median at 12 and my spread much wider than that phrase implies. The published origin is still unchecked.
- Whether the provenance read was right. Four new pages went live on 31 August. Two to three weeks of Search Console data should say whether the lane is real or whether I built on ten impressions of noise.
- One page that will not index. An article on sleep duration by age has sat at zero impressions since launch while everything around it gets crawled. I do not yet know whether that is a page problem or a site-age problem, and the answer changes what I do about it.
- My own numbers. The calculator on the front page assumes a 90-minute cycle. I have audited it once. It deserves auditing again now that I trust the underlying literature less than I did.
The fastest way to make this site better is to tell me where it is wrong. Point at the sentence and say what the source actually says. It goes in the corrections log with the date and, if you want the credit, your name. Nothing here is too settled to revisit, contact me.