Does the podium lead the page?¶
First measured 26 August 2026; baseline corrected and re-measured on
abstracts 28 August 2026. Reproduce with python3 scripts/lead_time.py --papers
--cohort 2024 and --match abstract.
Every monthly briefing in this corpus rests on one assumption: that a meeting programme shows you where a field is going before the journals do, and that the gap is worth having. The assumption was never tested. It could not be — the journal index started at 2025-01 and every conference vault was 2026, so the two layers barely overlapped and no lead was measurable in either direction.
Testing it needed a meeting old enough for the literature to have answered it. AACR 2024 and ASCO 2024 — 14,895 abstracts with full text, from Crossref — against a journal index of 47,362 papers running 2023-01 to 2026-08. The index reached only 2024-01 when this was first written, which is why the baseline below had to be corrected: there were no journal months predating the meetings to measure chance against.
The answer¶
Meetings do lead, by about ten to thirteen months. Almost nothing said at them arrives.
| Median abstract-to-paper lag | 12 months (IQR ≈ 9–20) |
| Stability of that lag | unchanged across every threshold, and unchanged when matching abstracts instead of titles |
| Abstracts becoming papers in these 33 journals | ≈0.2–0.4%, against a chance rate of half that |
Previous estimate in threads.yml |
4–6 months, from five hand-picked cases |
The lag is the robust part. Tightening or loosening the matching threshold moves the hit rate by an order of magnitude and barely moves the median lag, which is what a real effect looks like. The old 4–6 month figure was low by about a factor of two, and it came from five cases chosen by someone who already believed the answer.
The hit rate is the part that should change behaviour. At a Jaccard cutoff of 0.5, 98 of 14,890 abstracts (0.7%) match a paper published later.
Corrected 2026-08-27 — the first honest baseline this test has had
The table below was published with a chance rate that was not a chance rate, and the corrected numbers are much weaker than the ones they replace. The lag above is unaffected.
The old baseline ran the matcher against journal papers published before
the meetings, selected with month < '2024-04' against an index whose first
month was 2024-01. So the "before" arm was three months — 2024-01 to
2024-03 — and those are the months this cohort deposits into: ASCO's GI and
GU symposia in January and February, AACR in March. A journal paper from
March 2024 matching an AACR March 2024 abstract is not coincidence; it is
the same work in both places at once. The arm built to measure luck was
measuring simultaneity. It was also a tenth the size of the arm it was
compared against.
Extending the index to 2023-01 gives twelve journal months that genuinely predate the meetings. Both arms are now thinned to the same 11,724 titles, so the comparison is fair by construction rather than by assumption.
| Match threshold | Matched later | Matched earlier (chance) | Ratio |
|---|---|---|---|
| 0.35 | 181 (1.22%) | 182 (1.22%) | ×1.0 |
| 0.45 | 51 (0.34%) | 23 (0.15%) | ×2.2 |
| 0.50 | 36 (0.24%) | 17 (0.11%) | ×2.1 |
| 0.60 | 15 (0.10%) | 5 (0.03%) | ×3.0 |
The effect is real, small, and only exists at strict matching. At a Jaccard cutoff of 0.35 the meeting arm scores 181 and chance scores 182 — the podium has no measurable advantage at all. The signal appears only as the threshold tightens, reaching about three times chance at 0.60, which is the behaviour of a real but narrow effect: loose matching catches shared boilerplate, which is equally common on both sides of a meeting.
Measured again with a better instrument, and it holds¶
Everything above matches conference abstracts against journal titles, because the journal index held no abstracts. That was the obvious suspect for the ~1% conversion rate, and on 2026-08-28, with abstracts harvested for all 47,362 papers, it was tested directly.
| matched later | chance | ratio | median lag | |
|---|---|---|---|---|
| title vs title, Jaccard 0.50 | 36 (0.24%) | 17 (0.11%) | ×2.1 | 12 months |
| abstract vs abstract, Jaccard 0.10 | 55 (0.37%) | 22 (0.15%) | ×2.5 | 12 months |
The answer does not move. Same lag, same order of magnitude, same ratio against chance. The conversion rate was not a title artefact, which is what a year of caveats in this corpus had assumed it might be.
The thresholds differ because Jaccard divides by the union: two 200-word abstracts share more words than two titles but have a far larger union, so they score lower, not higher. At 0.20 both arms collapse to three matches each and the ratio stops meaning anything. 0.10 is the operating point for abstracts, and comparing the two rows across different cutoffs is only legitimate because both carry their own pre-meeting null.
That a finding survives a change of instrument is the strongest thing that has been said about it here. It is also the first measurement in this corpus that a wider window or a better instrument did not overturn.
The previously published version of this table claimed a 1.2-percentage-point excess at 0.35, where the truth is none, and read as though the effect were strongest at loose matching. It was exactly backwards.
So of the order of one abstract in a hundred becomes a paper in the tracked journals within two years. Reading an abstract book to find them means reading about two hundred abstracts per eventual paper. The meeting is a real leading indicator and a terrible filter, and those two facts have to be held at once.
What this does not measure
These are the corpus's 33 elite journals. Most AACR and ASCO abstracts were never going to appear in Nature or NEJM — they publish in specialty journals outside the window, and that is not a failed prediction. The ~1% is a conversion rate into this corpus, not into the literature. Title-similarity matching also misses any abstract whose title was rewritten for publication, so the true conversion is higher than 1% by an unknown margin. The lag is the number to carry forward; the rate is a lower bound.
The vocabulary version is dead¶
A cheaper version of the premise says you do not need to track individual results — just watch which words are getting loud at the meeting, because that is where the literature goes next. That version is refuted.
Ranking terms by how over-represented they are at the 2024 meetings relative to the 2024 journals, and asking whether that predicts journal growth by 2026:
| Predictor | Spearman ρ |
|---|---|
| Conference over-representation, raw | +0.168 |
| Conference over-representation, register-controlled | +0.091 |
| Null: a random half of the journals' own 2024 output | +0.088 |
The gap between the meeting and a coin-flip split of the journals' own back
catalogue is 0.003. The apparent signal in the raw figure is not science, it
is dialect: mcrpc, pts with metastatic, real world, first-line 1l —
abstract shorthand that journal titles never use whatever the underlying work
does. Require a term to appear in at least one journal title and the effect
evaporates.
The same confound wrecks the naive first-appearance version. Of terms said at
five or more meeting records and absent from every 2024 journal title, 43% later
appear in a journal title and 57% never do — but reading the two lists shows
they contain the same kind of term. mcrpc arrived; mcrc did not. hnscc
arrived; hcc did not. That split is a coin flip on which abbreviation an
editor tolerated.
Why this took nineteen months to check¶
Both predictors in the growth test carry the 2024 journal rate in their denominator, so a term that was rare in 2024 by chance scores high on over-representation and on growth. That artefact alone produces a positive correlation from nothing, and the first version of this analysis reported it as a result. It needed a null model built from the journals' own data to see.
The first version of the first-appearance test was worse: it returned "conference first, 100% of terms, median lead twelve months", which was purely the shape of two non-overlapping windows. The fix for that was harvesting 2024.
Which is the general lesson, and it is not really about conferences. Every "first appearance" claim this corpus has published — five of them — moved earlier the moment the window was extended, and none has ever survived. A bounded window cannot date a beginning. It can only report its own floor.
What follows¶
- The
clinical-practicethread now carries measured numbers instead of an estimate. Seedata/threads.yml. - A meeting abstract is worth logging when it names a specific trial, target or readout that can be checked against the literature in a year. It is not worth mining for emerging vocabulary.
- Ten to thirteen months is the planning horizon the corpus can actually claim.
The conference layer has since grown, and it does not change this¶
On 2026-08-27 the store went from two venues to nine — 18,749 abstracts, adding ELCC, ESMO Breast, ESMO GI, ESMO Gynaecological, USCAP 2026 and two ESMO Open supplements whose congress could not be identified. None of them can extend this measurement, and it is worth being explicit about why, because a bigger number invites the assumption that the finding got stronger.
Everything new is from 2026. The lag being measured is ten to thirteen months, and the journal index ends at 2026-08 — so a March 2026 abstract book has about five months of follow-up against a median lag of twelve. Testing it now would return a low hit rate that means "not enough time has passed", and would be indistinguishable from "these meetings do not lead". The 2024 cohort is the only one with a follow-up window longer than the effect it is measuring, so the numbers above are still computed on AACR 2024 and ASCO 2024 alone.
The 2026 venues also arrive as titles only — the ESMO family and Laboratory Investigation deposit abstract metadata without abstract text. The 2024 test used full text. Re-running it in 2027, when the follow-up window is long enough, will therefore be a weaker test on those venues, not merely a later one.
What the new venues are good for now is coverage of what was actually said in 2026, which is what sections 6 of the monthly briefings report.