Skip to content
Catalyst

creatine.twin · version 1’s report

creatine.twin version 2: external validation

Version 2 is version 1 with absorption that saturates (SHA-256 045bfd02…690d46, version 1’s creatine.ts unchanged at 3e58be5d…ce4770). Protocol: PROTOCOL-creatine-v2.md, commit b08fade2e, task twin-creatine-v2-validation. This is science about a module, not a product, and no figure here is a prediction for any person. Version 1 stays the page’s model. validation-v2.csv · validation-v2.json

Version 2’s absorption constants come from rats: Km and Amax are fixed by two printed points of PMID 29518030 (53% absorbed at 10 mg/kg, 16% at 70 mg/kg) and applied here as human mg/kg. It has no muscle compartment.

Result, as first run

0 observations were scorable, so the primary endpoint (within twofold on plasma creatine Cmax, Tmax, AUC and half-life) has no value, for version 2 or for version 1. The query returned 504 PMIDs (2026-11-16T03:12Z); 504 abstracts were fetched, 22 had none, and the fixed text screen passed 86; after the build and development papers were removed, 84 were transcribed. Two independent readers wrote 31 and 87 values; 13 agreed on every compared field, and 13 of those had a verbatim quote holding the value; the scope rule kept 0.

ModelObservationsWithin 2-foldWithin 1.5-foldWithin 1.25-foldAFEAAFE
Version 20000not definednot defined
Version 1, same rows0000not definednot defined

Primary metrics: Cmax, tmax, AUC0–∞, AUC0–t, t½. Per metric, by regimen and by dose the counts are all zero, and the secondary endpoints are empty for the same reason. No amendment was made, and nothing was re-run or tuned.

Cause

The agreed rows are all out of scope, and most values never reached agreement: the readers agreed on 13 values, and 83 were one reader’s alone or differed on a scope field. Reader A wrote 31 rows and reader B 87; B also transcribed values for other analytes, muscle and renal figures that A did not. Among creatine plasma values that could have been scored, the readers differed on a regimen’s structure (PMID 33900258, 20 g a day; PMID 9629989, 3 × 100 mg/kg/day) and on the population or statistic of others (PMID 1327657, a single 5 g dose in three subjects; PMID 28121472, a microencapsulated form), and the protocol says disagreements are listed and not resolved by judgement. This says nothing about how well version 2 predicts. It says the literature that passes this protocol’s scope and two-reader agreement is close to empty, and that regimen rows, which this version was written to admit, are where the readers diverged. Every disagreement, with its reason, is in validation-v2.json.

Exclusion reason (agreed values)Values
analyte is phosphocreatine, not creatine5
children and adolescents5
disease or patient population3
DisagreementValues
reported by reader B only65
reported by reader A only9
readers differ on health4
readers differ on statistic2
readers differ on statistic, auc_end_time_h1
readers differ on age_group1
readers differ on age_group, statistic1

Source 1, the study records

Study records for creatine recorded after 2026-11-08T17:46:59Z (the latest row version 1’s amendment A2 saw) and before the protocol commit: 0 rows. Query: paper_pk?subject_node_id=eq.CHEBI:16919&and=(recorded_at.gt.2026-11-08T17:46:59.14348Z,recorded_at.lt.2026-11-16T03:09:34Z), read 2026-11-16 at about 03:15Z. Control: the same table holds 31 creatine rows in all, the latest recorded 2026-11-08T17:46:59Z, so the query can return rows and returned none for this window. Source 1 is empty.

Development data, beside and never pooled

These were seen before the protocol, so they are not a test of version 2.

SetModelComparisonsWithin 2-foldIn 5th–95thAFE
The module’s six rows (PMID 35268011, 34445003)Version 16631.24
The module’s six rowsVersion 26321.82
PMID 10622230, single 20 g (version 1’s amendment A1)Version 11101.82
PMID 10622230, single 20 gVersion 21000.49

Saturation moves the 20 g prediction from 1.82-fold high to 0.49-fold (just past twofold low), and moves the 2 to 2.5 g arms of PMID 35268011 from 1.0 to 1.6-fold high to 1.8 to 2.6-fold high. The saturable form with rat constants corrects the high dose and over-corrects the low ones: three of the six rows fall outside twofold under version 2, none under version 1. These rows are not scored for the endpoint.

Method, as written

Two independent readers (separate sub-agents, each given only the transcription instructions and the in-scope abstracts) transcribed each candidate abstract. A value is kept when both report the same PMID, analyte, metric, value, unit and dose and regimen structure, their scope fields agree, and the quote is verbatim in the abstract and contains the value. Scope is applied mechanically, including a regimen’s interval, number of doses and time after the last dose. 1,000 virtual subjects per arm, seed 1. The PMID list was fetched once at the start of the run (generic tool name, no contact detail) and is saved with the data.