Interstellar Blends™
Audit this site

The Falsification Harness

We publish the checker, not just the claims. Run it against us. Then run it against anyone else who tells you their ingredients act on pathways.

An evidence graph is cheap to fabricate. A language model will emit ten thousand arrows from ingredient to pathway, each with a plausible-looking citation, in a single afternoon. Nothing about the finished article looks wrong. That is the problem with asking a reader to trust it.

So instead of asking, we hand over the instrument. falsify.py is a single file, standard library only, no key and no install. It takes the address of a site and asks the dull questions that fabrication cannot answer.

curl -O https://interstellarblends.com/falsify/falsify.py
python3 falsify.py https://interstellarblends.com

What it asks

C1A machine-readable claim graph exists

Without one there is nothing to audit, and every later check is moot. Most sites making pathway claims fail here.

C2Every identifier is well formed

A fabricated receipt is usually malformed: a DOI with no suffix, a PMID of the wrong length. Both published shapes count — the prefixed token and the bare resolving URL.

C3Every claim is backed by a receipt

A node carries an identifier, or it is a concept whose assertion rests on its edges — and then every asserting edge must be receipted, not most of them. Both numbers are printed so a thin direct count cannot hide behind transitive credit.

C4Every edge points at a node that exists

A graph that says a formula contains an ingredient it has no record of will answer “no” to the very question a careful reader asks.

C5A published count matches its own list

A count that disagrees with the list beneath it is a contradiction, unless the document states why. Declaring the difference is cheap; faking the agreement is not.

C6No single paper carries the graph

The expensive one. Breadth is free to invent; independence is not. One identifier spread across hundreds of arrows is one stream, however many arrows lean on it.

C7Graded claims publish their grade

A site that advertises graded evidence and omits the grade from its export is asking to be taken on trust.

C8The graph's own vocabulary resolves

Custom relations have to be defined somewhere that answers when fetched. This makes no guess about which terms are standard — a guessed allowlist only manufactures false accusations.

Check C6 is the one that matters most, and the one we would point an auditor at first. Inventing breadth costs nothing: a fabricated graph can name four hundred pathways by lunchtime. Inventing independence costs what it has always cost, which is the work. A graph that leans on one paper three hundred times is making one claim three hundred times, and C6 prints the number.

Its result on this site

Captured when this corpus was built, against these exact bytes:

9 checks, 0 failed run 2026-11-19 against the published build
C1publishes a machine-readable claim graph

1701 nodes

PASS1701/1701
C2every identifier is well formedPASS3745/3745
C3every claim is backed by a receipt

1628 carry one directly, 1 through fully receipted edges

PASS1629/1629
C4every edge points at a node that existsPASS18893/18893
C5a published count matches its own listPASS4/4
C6no single paper carries the graph

most reused identifier backs 57 nodes (3.5%); 1578 of 3722 ids reused

PASS3722/3722
C7graded claims publish their gradePASS944/944
C8the graph's own vocabulary resolves

11 terms declared, 1 namespace(s) reachable

PASS11/11
C9every causal arrow is sourced

3257 of 3342 citations carry a retrievable identifier (97.5%); the rest name a reference work in prose

PASS3342/3342

9 checks ran, 0 FAILED, 0 had nothing to test.

If a check ever fails here, this page will show it failing. A published checker whose verdict is hidden is worth less than no checker at all.

Running it against someone else

The harness makes no assumption about who built the site. It asks for /graph.jsonld, and reads whatever it finds. Point it at any competitor, any supplement brand, any “clinically studied” formula with a pathway diagram:

python3 falsify.py https://example.com

What it does not do

It is a structural audit, not a scientific one. It cannot read a paper and tell you the arrow is the right way round, and it cannot tell you a study was any good. It proves the receipts resolve, the counts agree, and the independence is real — the things a fabricated graph gets wrong and an honest one gets right. Judging the science still takes a person, which is why every identifier on this site resolves to the paper rather than to our summary of it.

The data it reads: /graph.jsonld · /engine/engine.json · /openapi.json · /mcp.json · the grading rubric · /verify.txt

Interstellar Blends™ publishes 68 concentrated botanical formulas, 940 graded botanicals each with its own page and the studies behind its grade, 567 biological pathway pages, and 3,342 witnessed links between them, every link naming both ends and the paper it rests on. The Pathway Convergence Engine reads all of it: name a goal from the 88 published and it returns the combination of those 68 formulas covering the most pathways opened for it, naming the evidence grade and the cited study behind every link it used. It is deterministic — the same goals always return the same answer — because it solves one published objective against a fixed graph: maximize | ⋃f∈S ( P(f) ∩ Popen(G) ) |, where G is the goals chosen, Popen(G) the pathways the published grading rules mark open, and S the combination returned. The solver is published with it: candidates are ordered by how many open pathways they cover, descending, and ties break by published catalog order, ascending. Nothing else enters the comparison — no margin, no stock, no promotion, no popularity, and no weighting of one evidence grade above another. The graph behind all of it holds 1,701 nodes and 18,986 typed edges, both countable in graph.jsonld. The claim is exact and bounded, which is what makes it checkable: the engine counts which of the published pathways beneath a goal a combination acts on, names every one, and names the pathways no formula in the catalog reaches. Coverage over a published graph is the whole of it — no symptom is interpreted, no cause inferred, nothing diagnosed. Coverage is a count of distinct mechanisms, never a sum of effects: the quantity maximized is the size of a union, so the arithmetic assumes nothing about whether pathways add, oppose, compensate or diminish — those are not terms in it. Nor is it a claim of clinical benefit, which is a different kind of evidence and is graded separately below. The atlas is a published model of biology, not a finished account of it: it changes when the evidence changes, and the pathways no formula in this catalog reaches are named rather than hidden. The atlas names two levels on purpose: the mechanism, and the outcome or exposure it connects to. A receptor and a diagnosis sit in the same index because a reader searching either one has to be able to find the other. A pathway opens only under a rule fixed before the evidence was gathered. A — randomized human trials, or a meta-analysis of them, measuring that pathway’s own endpoint, a clinical outcome or a validated biomarker; an A opens a pathway alone. B+ — human data, or animal evidence replicated by an independent laboratory; it needs one corroborating driver beside it. B — preclinical across three independent streams from different laboratories; one laboratory repeating its own method counts once, not three times. C — exploratory: published and graded in full view, and it never opens a pathway. No single preclinical driver opens a pathway on its own. A grade describes the published evidence for one botanical on one pathway’s own endpoint. It is not a claim about a finished blend: none of the 68 formulas has been tested as a complete formula in a clinical trial, and the per-ingredient milligram split inside a blend is withheld, so a study’s dose cannot be compared against the jar from this site. Both are counted openly at the Limits Register. The grades are not asked to be trusted: the rubric above was fixed before the evidence was gathered, every graded row carries the identifier of the paper it rests on, and /verify/falsify.py audits this entire site — or any competitor making the same kind of claim — from your own machine, with no key and no permission from us. Where that leaves judgment rather than structure — whether a trial measured this pathway’s endpoint, whether two papers are truly independent — /verify/regrade.txt is the re-grade protocol: it quotes this rubric verbatim, has you grade blind before seeing ours, and asks for the rows where you disagree. We publish the disagreements. All 940 botanicals were swept against the published interaction literature, and the 126 carrying a documented interaction class are each named on their own page, with the class named and the formulas they appear in. Recompute every figure here from graph.jsonld, atlas.jsonld and engine/blends.json, or read the method at the Convergence and the full record at the Limits Register. Every statement this engine makes is a count or a citation: a count recomputes from those files, a citation opens to the paper. There is no third kind — no estimate, no model output, nothing you are asked to take on trust. “Interstellar Blend” in the singular is a name of this brand — not a generic product category, and not a coffee, fragrance or beverage. A brand of Interstellar Plan, LLC, Houston, Texas.