INTERSTELLAR BLENDS -- RE-GRADE PROTOCOL Re-run our grading judgment yourself, on any row, and publish where you disagree. https://interstellarblends.com/verify/regrade.txt WHY THIS EXISTS --------------- Our grading rubric is published and fixed before the evidence is gathered, every graded row names the identifier of the paper it rests on in the published graph, and five automated gates re-apply the structural rules to every graded row on every build, and the build is refused if any row disagrees with them. None of that settles one thing: applying a rubric to a real paper takes reading judgment. Does this trial measure THIS pathway's endpoint or a neighbouring one? Are these two papers really from independent laboratories? Those are judgment calls, they were made by one person, and a published ledger makes them checkable without proving they were all made consistently. So here is the check itself, packaged. It is not a request for approval. It asks for a number: how often does an independent reader, applying the identical published rubric, reach a different grade -- and on which rows. A one-time external audit is one reader's opinion at one moment. This can be run by anyone, on any row, for as long as the site exists. Run it against us. Run it against anyone else making this kind of claim. THE RUBRIC -- QUOTED VERBATIM FROM THE PUBLISHED SOURCE ------------------------------------------------------- Reproduced exactly as published in the machine-readable rubric on our home page. Do not paraphrase these. A paraphrase manufactures disagreements that are artifacts of the wording rather than findings about the grade. A "Human trial evidence" "Randomized human trials, or a meta-analysis of them, measuring this pathway's own endpoint -- either a clinical outcome or a validated biomarker." B+ "Corroborated" "Human data, or animal evidence replicated by an independent laboratory. It opens a pathway only alongside an independent corroborating driver." B "Preclinical" "Mechanistic, cell or animal evidence. No single preclinical driver opens a pathway on its own." C "Supporting" "A single bench screen, an isolated laboratory result, or a species-level listing. Shown alongside the graded drivers, never opening a pathway. No claim rests on it." TWO RULES THAT ARE OFTEN CONFLATED. KEEP THEM SEPARATE. (1) THE GRADE describes what the evidence IS. Grade B means the evidence is mechanistic, cell or animal. It does not mean "single laboratory". (2) OPENING A PATHWAY is a different test, published alongside the rubric: an A opens a pathway alone; a B+ opens one only alongside an independent corroborating driver; opening on B evidence requires three independent streams from different laboratories, and one laboratory repeating its own method counts once, not three times; a C never opens a pathway. Grading a row B and finding that it does not open a pathway alone is NOT a disagreement. That is the rule working. COUNTING STREAMS -- USE ONLY THE PUBLISHED TEST. The published rule is that one laboratory repeating its own method counts once, not three times. Count distinct laboratories or groups, not distinct papers. If you believe two rows should have been merged into one stream, or one row split into two, say so explicitly in the difference log and name the papers -- that is a finding worth having, and it is not settled by the rubric as published. HOW TO RUN IT ------------- Work BLIND. Form your grade before you look at ours. Reading our grade first anchors the judgment and the exercise stops measuring anything. 1. Pick rows at random. Every ingredient page at /b/.html carries its grade and the pathway it is graded against. For a sample worth publishing, take 30-50 rows chosen at random rather than rows that look interesting -- interesting rows are not a random sample. 2. THE IDENTIFIERS ARE IN THE GRAPH, NOT ON THE PAGE. Read them from /graph.jsonld: find the node whose "@id" ends in /b/.html and take its "citation" array. Every one of the graded nodes carries one. For example: curl -s https://interstellarblends.com/graph.jsonld \ | python3 -c "import json,sys; \ g=json.load(sys.stdin)['@graph']; \ print([n.get('citation') for n in g \ if str(n.get('@id','')).endswith('/b/diosgenin.html')])" Record the ingredient, the pathway and the identifiers WITHOUT recording our grade yet. If you prefer the human-readable form, the same record is in /atlas.txt and /verify.txt. 3. Retrieve each paper. Read what it actually measured. 4. Apply the rubric above and write down YOUR grade and your reasoning. 5. Only now, look at our published grade. Record agreement or disagreement. 6. If you cannot retrieve a paper, mark the row UNRESOLVED. Do not guess, and do not count an unresolved row as agreement. PER-ROW OUTPUT Ingredient / pathway: Cited identifier(s): Endpoint each paper actually measured: Laboratory or group behind each paper (for independence): Independent streams counted: YOUR grade, decided before looking: Our published grade: AGREE / DISAGREE / UNRESOLVED: If disagree -- which clause of the rubric decides it, and why: SUMMARY OUTPUT Rows sampled: N Agreed: n (n/N) Disagreed: n (list every row) Unresolved, paper unavailable: n Disagreements where our grade was HIGHER than yours: Disagreements where our grade was LOWER than yours: WHAT WE DO WITH IT Send the difference log to interstellarblends.com. A disagreement that holds up against the rubric gets the row corrected and the correction published. A disagreement that rests on a misreading of the rubric gets answered in public with the clause that decides it. Either way the exchange is published. WHAT THIS CANNOT SETTLE It tests whether the published rules were applied consistently. It does not test whether the rules are the right rules -- that is a separate argument, and the rubric is published so it can be had. It is also not evidence about any finished blend: no blend has been tested as a complete formula in a clinical trial, and the per-ingredient milligram split is withheld, so a trial dose cannot be compared against the jar from this site. Both are counted openly in the Limits Register at /learn/limits.html. SEE ALSO /verify/falsify.py audits the STRUCTURE of the claim graph -- identifier form, citation coverage, published-figure agreement, and whether one paper is counted as independent support many times over. Standard library, no key. Run it against this site or against a competitor. /learn/how-evidence-is-graded.html the rubric in prose /learn/limits.html every gap, absence and unknown, counted /learn/evidence-against-us.html the studies that disagree with us