How TitrateLab turns documents into evidence

From document to record

The archive total counts published COA report entries, including separate assays and reports whose compound or measurements have not been extracted. A certificate can contribute more than one entry. Compound summaries, measurement charts and vendor rankings show their own included populations.

1. Collect the source

The certificate corpus combines reports from Freedom Diagnostics, Janoshik Analytical, Finnrick, public forum attachments, public vendor channels, and buyer submissions. Freedom Diagnostics is the largest source in the current corpus. Janoshik records usually link to the lab’s public verification page. Finnrick publishes panel results under its own coded lab labels, so we can name the panel operator but not always the bench lab.

We also collect observed supplier prices, regulatory records, public enforcement actions, research papers, and market events. Those sources answer different questions and are stored separately. A price observation never becomes proof of product quality.

2. Read and preserve it

When a document has a usable text layer, fields are parsed directly. Images and scans are read with Gemini 2.5 Flash. We retain the original document or its verification link where available so the structured record can be checked against its source.

Automatic extraction can be wrong. A number that changes the meaning of a record should be checked against the document itself. Records can also become harder to verify when a lab removes an older public image.

3. Resolve the identity

Supplier names are inconsistent. One business may use a legal name, storefront name, short code, and several spelling variants. We compare declared names, known aliases, domains, source channels, and community co-mentions. Low-confidence matches go to a review queue instead of creating a new vendor automatically.

This is still an imperfect layer. Two storefronts may share an upstream manufacturer, and one manufacturer may use several identities. Unless a relationship is documented, vendor-level results should be read as results for the named identity, not proof about an entire factory.

What the measurements mean

Purity is the percentage reported by the testing lab, usually from HPLC or UHPLC.

Quantity deviation compares the measured amount with the amount on the label. A negative value means the tested vial was underfilled.

Endotoxin is shown only when the record contains an endotoxin result. No result means unknown, not passed.

In spec is a TitrateLab rule for summarizing records. A result counts as in spec when purity is at least 90% and, when a label comparison exists, quantity is no more than 10% below label. Overfill is not counted as a failure. This is a one-sided rule, so it is more permissive than requiring every result to fall within plus or minus 10% of label.

How the vendor ranking works

The numbered ranking on /vendors is the Most-reviewed peptide vendors list. It answers one narrow question: which eligible vendor identities have the most unique, evidence-backed lab reports in the published corpus? It is not a recommendation, quality score, price judgment, or claim that the vendor made the tested product.

Eligibility is deliberately fail closed. A vendor enters the numbered list only when an operator has manually marked its canonical identity as not a U.S. reseller. Identities marked as U.S. resellers and identities that have not yet been reviewed receive no rank. They keep their vendor dossier and can still appear in buyer-report search. Merged, suppressed, flagged, unresolved, or conflicting identities do not rank.

Each unique lab report counts once. A report qualifies when TitrateLab retains an artifact or has a direct report or certificate link. Duplicate archive copies and multiple analyte rows from one certificate collapse to one report. If a strong report identity is assigned to more than one vendor, it is credited to neither until the attribution is resolved.

Vendors are ordered by:

1. unique qualifying report count;
2. artifact-backed report count when totals tie;
3. newest dated qualifying report when both counts tie; and
4. canonical slug as the final deterministic tie-break.

The same numbered ranking powers /vendors and /dashboard/vendors. The dashboard adds filters, vendor files, and the analytical Trust Grade, but those secondary measurements do not determine rank.

How the Trust Grade works

The Trust Grade is a relative summary of the lab evidence attached to a vendor identity. Six components feed the score:

For the purity and quantity components, a record loses half its row weight every 180 days, with a small floor so older history does not disappear. Those weighted rows also determine the evidence-depth component.

The consistency component uses pooled variation within each peptide. Endotoxin uses reported pass and fail counts with a prior. The public-evidence component currently uses records explicitly classified as public in the structured corpus; it does not count every record that happens to carry a public URL. Those three components are not recency-weighted. Purity, quantity, and consistency are pulled toward market priors when evidence is thin; two favorable reports cannot produce an extreme score by themselves.

The letter is graded on a curve against the tracked vendor population:

An A means stronger documented evidence than most tracked identities. It does not mean the vendor meets a government or pharmaceutical quality standard. Fewer than three purity measurements produces Insufficient purity data, not a letter.

Serious failures can cap the letter even when the average looks good. An independent-lab record at 0% purity or 40% or more under label can trigger a cap, as can a reported endotoxin failure or major batch inconsistency. Repeated critical failures can force an F. The underlying score remains visible so readers can see both the average and the cap.

The public vendor letter uses lab evidence. Subscriber comparisons may also show community incidents and other context, so those views can differ from the letter on a vendor profile.

Buyer submissions

Anyone can submit a certificate. Automated screening checks whether it appears to be a COA, whether it is readable, and whether it contains signs of manipulation or prompt injection. Buyer-uploaded records retain their source and review status. A buyer upload currently gets the same numerical weight as another record of the same age. We disclose that limitation here instead of implying the model applies a discount that it does not.

Soft warnings do not silently turn into passes. Possible duplicates, unusual values, unknown vendors, and other review flags remain attached for operators to inspect.

Price and market records

Prices are observations, not quotes or manufacturing-cost claims. We record what a supplier asked, the product and size, the delivery lane when known, and when the offer was seen. Shipping, discounts, minimum orders, and stock changes can make a delivered price different.

Pricing does not affect a Trust Grade. A high price is not evidence of quality, and a low price is not evidence of failure.

Community discussion helps identify events worth checking, but it does not enter the public Trust Grade. Public-forum material may retain the attribution already visible on that forum. Private-community material is handled differently.

Discord anonymization

Discord author handles and server names are anonymized before material is published. The system is meant to surface claims and source documents, not expose private-community identities. Community claims remain claims until supported by a lab report, public filing, or other checkable record.

Scoring and commerce

TitrateLab earns revenue from subscriptions and from a margin on some customer orders it routes. That is a real commercial interest and should be visible on the page where readers evaluate our independence.

The separation is mechanical: supplier prices, subscriptions, routed-order volume, and commercial relationships are not inputs to the public Trust Grade. There are no paid grade changes, sponsored positions, or affiliate boosts on graded vendor surfaces. A score changes only when the underlying record, attribution, or published scoring method changes.

Limits, biases, and open problems

• Selection bias. Tested batches are not a random sample of everything a vendor ships. Buyers and vendors choose what gets tested and published.
• Source concentration. A large share of the corpus comes from Freedom Diagnostics. Any source-specific bias can affect market-wide summaries.
• Lab disagreement. Different labs and methods can report different quantity or purity results. We preserve the source instead of pretending every result is interchangeable.
• Identity uncertainty. Storefront, supplier, and upstream-manufacturer relationships are not always public. A vendor page can merge or separate identities incorrectly.
• Uneven coverage. Popular vendors and compounds are tested more often. Thinly tested vendors may look average because the model correctly has little confidence.
• Freshness. A recent observation is not a live inventory guarantee. The measured date and source matter more than the page’s publication date.
• Missing safety fields. Most purity reports do not answer sterility or endotoxin questions. Purity alone cannot establish that a product is safe for use.

What we claim, and what we do not

TitrateLab claims that its displayed figures are traceable to a document, observation, or disclosed calculation. We publish the source, date, and evidence gaps where the source permits it.

The figures are not medical advice, a therapeutic recommendation, a safety clearance, or a guarantee about an individual vial. See the medical disclaimer. Vendor names identify parties in the documentary record; inclusion is not endorsement and exclusion is not condemnation.

Corrections and right of reply

If a source is misattributed, a calculation cannot be reproduced, or a material fact is wrong, email [email protected] with the page, record, and supporting source. We correct the record and date material changes. We do not remove an accurate lab result solely because its subject dislikes it.

Change log

2026-09-04. Replaced the /vendors and /dashboard/vendors composite-score ordering with the Most-reviewed ranking: manually reviewed non-U.S.-resellers ordered by unique evidence-backed lab reports, with duplicate and conflicting report identities handled fail closed.

2026-08-19. Corrected the live corpus measurements and clarified their populations. Replaced two inaccurate score descriptions with the behavior the current model actually implements, removed repeated introductory copy, and replaced named forum references with the generic public-forum description.

2026-08-13. Rewrote the page in plain language, removed stale snapshot counts from prose, clarified the difference between source records and derived scores, and added a direct disclosure of TitrateLab’s order-routing margin and its separation from vendor scoring.

2026-07-25. Audited the published methodology against the live code and database. Corrected source, OCR, scoring, freshness, and pipeline claims.

2026-04-01. First published.