5byfive
GLP-1 Enthusiast
This is a repost of a topic from several months ago. I had posted in it Public Square and @ZippityDooDah said it got nuked due to source discussion issues and that is should have been posted here. I still see people referring to it so it seemed worth reposting. Fortunately, I kept the post in a doc so I can recreate the original post.
I plan to do a version 2 in a couple months once we have enough new data. I have found one more large source to pull from (it will mostly add new vendors) and I also plan to go through and remove all the data from Krause labs. There is also some data comparing cap colors across vendors I want to look at. Looking at it briefly, I see some issues but I'll let Claude see what he can make of it. I also want to thank everyone for all the positive feedback y'all have given me; I really appreciate it. Without further ado, here is the original post:
I have been working on a project for a few weeks now trying to determine vendor supply chains. The main reason for doing this was trying to figure out who to trust. A handful of COAs are good, but with the variation in results, being able to group a batch of results across a group of vendors provides a lot more information.
The analysis uses the term manufacture although it’s not really not an accurate description. We are really talking about common supply chain. My thought, and this is obviously conjecture, is that there are probably only a couple manufactures and a larger (but still small) number of finishers that actually take the bulk chemicals and package them. This is really where all the variability seems to be. It seems like manufacturing quality is actually very consistent with most of the variability being in finishing.
The process:
My initial data set was around 6200 COAs scrapped from the the scammers manipulating test results for profit at Finnrisk website. From that list, I removed anything that sounded like an American reseller on the basis that they were just further down the food chain and that they may purchase from more than one Chinese vendor. That left me with around 4200 COAs. This was my initial run.
Then I manually pulled 42 COAs from the Jeep listings after it came out that they failed for Endotoxins. This actually helped refine the groups a little bit. That was the data set from my first post in the Endotoxin thread.
For this set, I added about 1000 COAs from the STG group and I added a bias towards more recent results to reflect possible changes in manufactures after some were closed last year and also the possibility of vendors changing finishers.
There are some obvious limitations. The analysis assumes that every vendor is using a single source for every product. I would think this is likely and it wouldn’t be possible to really analyze the data without this assumption, but it is a possible weakness. It also assumes that every vendor has one source and that they are not always going with whoever is cheapest at the moment. Again, I think it is reasonable to assume stable relationships but there is no guarantee that some or all of them do not do this at least some of the time. The recency bias offsets this risk some.
The how:
I used Claude to analyze the data. I started with ChatGPT but found it to be very unreliable. I caught it in several lies. It seems to have a tendency to make stuff up both to conserve resources and to save face if it can’t do something you ask it to. I didn’t see any of this behavior with Claude.
Here was the initial command set:
The vendors in the attached file likely get their product from one of 6 to 8 manufactures. Please analyze the data in the "Difference %", "Purity %", "Product" columns to identify and group vendors with common manufactures.
Assume that each manufacture makes most of the available products and that this is not a good criteria for grouping them. Very similar scores in purity of the same product (ie similar scores in Tirzepatide purity and similar scores in Tesamorelin purity) around the same time frame would be a better indicator of shared manufacturing. Over all, the most similar scores of the same products around the same time is the best indicator. Please take your time and analyze this closely to group vendors.
Please take what you have calculated and factor in "Difference %" to determine your groupings. Take your time.
Then I added the following assumptions:
Each vendor only sources from one manufacturer.
Purity % results are only accurate within .25%.
There are likely around 6-8 manufactures but it is ok if there are more or less based on the data.
A single product is one where "Product" and "label claim" are the same ie Tirzepatide 10mg.
The Difference % of a product is likely a better indicator than the Purity % but use both.
How many different products a vendor has listed is not an indicator.
When comparing we want to compare products where the test date is in the same time frame.
On the newest run, the only change I made in how it was calculated was to add weight to recent testing. Someone pointed out that there were some labs closed down last year that would have shaken things up. Claude moved 6 vendors into a different group based on the increased emphasis on recent tests suggesting that they may have been impacted by the factory closings. I also added the STG data.
Here are the notes about what changed between my last version and this one:
Dataset grew significantly. 5,181 records vs 4,138 before (+25%), with 11 new vendors. The new vendors split as: 7 into Group 1, 1 into Group 3 (SPB), and 2 with insufficient cross-product data (ACR, PTB).
Recency weighting changed who matters. With a 180-day half-life from March 2026, tests from October 2025 onward carry roughly 4–8× more weight than tests from early 2025. The signatures shown for each group are based on this recent window where possible. Group 6 (Alpha-Gen, Loti Labs, Suaway Lab Research) has zero data after September 2025 — their prior pattern is preserved but flagged as stale.
The three clearest discriminating signatures, in order:
Who would I tend to buy from? For that I asked the following question:
Which group has high quality while having good consistency in difference% that also tends towards overfull rather than under fill. When trying to figure out the actual fill to calculate dosage, I would rather have the error be in a consistent direction (plus extra product doesn’t hurt).
The result from Claude:
Group 4 is the answer.
Here's why it wins on all three criteria simultaneously:
Consistency — Group 4 has the lowest Diff% standard deviation of any group at ±15.7%. That's meaningfully tighter than G1 (±23.4%), G2 (±22.5%), and G3 (±17.7%). G5 is disqualified immediately with ±51.1%.
Overfill tendency — Average Diff% of +3.6% and 59.5% of tests coming in positive. Crucially, 45.8% of tests land between 0% and +15% — a moderate, controlled overfill rather than a random mix of large positives and large negatives averaging out to something positive.
Quality — Purity at 99.43%, which is solidly high, though G2 (99.52%) is slightly better. However G2's Diff% standard deviation is 6.8 points worse than G4's, meaning you're trading off consistency for marginal purity gain.
G2 is the close runner-up — it has slightly better purity and the highest rate of positive tests (63.2%), but the fill is noticeably less consistent than G4. If fill consistency is the priority, G4 is the clear choice. If you weight purity slightly more and can tolerate more variability in fill, G2 is a reasonable alternative.
G3 and G5 are out — G3 has the lowest purity (99.22%) and G5 has wildly inconsistent fill (±51.1%).
So my answer would be either group 2 or 4 although I have dealt with several group 2 vendors and that is the one I’ll probably stick with,
I hope this is helpful or at least interesting. Let me know if y’all have any questions or any suggestions on how I can further refine the data. Thanks.
I plan to do a version 2 in a couple months once we have enough new data. I have found one more large source to pull from (it will mostly add new vendors) and I also plan to go through and remove all the data from Krause labs. There is also some data comparing cap colors across vendors I want to look at. Looking at it briefly, I see some issues but I'll let Claude see what he can make of it. I also want to thank everyone for all the positive feedback y'all have given me; I really appreciate it. Without further ado, here is the original post:
I have been working on a project for a few weeks now trying to determine vendor supply chains. The main reason for doing this was trying to figure out who to trust. A handful of COAs are good, but with the variation in results, being able to group a batch of results across a group of vendors provides a lot more information.
The analysis uses the term manufacture although it’s not really not an accurate description. We are really talking about common supply chain. My thought, and this is obviously conjecture, is that there are probably only a couple manufactures and a larger (but still small) number of finishers that actually take the bulk chemicals and package them. This is really where all the variability seems to be. It seems like manufacturing quality is actually very consistent with most of the variability being in finishing.
The process:
My initial data set was around 6200 COAs scrapped from the the scammers manipulating test results for profit at Finnrisk website. From that list, I removed anything that sounded like an American reseller on the basis that they were just further down the food chain and that they may purchase from more than one Chinese vendor. That left me with around 4200 COAs. This was my initial run.
Then I manually pulled 42 COAs from the Jeep listings after it came out that they failed for Endotoxins. This actually helped refine the groups a little bit. That was the data set from my first post in the Endotoxin thread.
For this set, I added about 1000 COAs from the STG group and I added a bias towards more recent results to reflect possible changes in manufactures after some were closed last year and also the possibility of vendors changing finishers.
There are some obvious limitations. The analysis assumes that every vendor is using a single source for every product. I would think this is likely and it wouldn’t be possible to really analyze the data without this assumption, but it is a possible weakness. It also assumes that every vendor has one source and that they are not always going with whoever is cheapest at the moment. Again, I think it is reasonable to assume stable relationships but there is no guarantee that some or all of them do not do this at least some of the time. The recency bias offsets this risk some.
The how:
I used Claude to analyze the data. I started with ChatGPT but found it to be very unreliable. I caught it in several lies. It seems to have a tendency to make stuff up both to conserve resources and to save face if it can’t do something you ask it to. I didn’t see any of this behavior with Claude.
Here was the initial command set:
The vendors in the attached file likely get their product from one of 6 to 8 manufactures. Please analyze the data in the "Difference %", "Purity %", "Product" columns to identify and group vendors with common manufactures.
Assume that each manufacture makes most of the available products and that this is not a good criteria for grouping them. Very similar scores in purity of the same product (ie similar scores in Tirzepatide purity and similar scores in Tesamorelin purity) around the same time frame would be a better indicator of shared manufacturing. Over all, the most similar scores of the same products around the same time is the best indicator. Please take your time and analyze this closely to group vendors.
Please take what you have calculated and factor in "Difference %" to determine your groupings. Take your time.
Then I added the following assumptions:
Each vendor only sources from one manufacturer.
Purity % results are only accurate within .25%.
There are likely around 6-8 manufactures but it is ok if there are more or less based on the data.
A single product is one where "Product" and "label claim" are the same ie Tirzepatide 10mg.
The Difference % of a product is likely a better indicator than the Purity % but use both.
How many different products a vendor has listed is not an indicator.
When comparing we want to compare products where the test date is in the same time frame.
On the newest run, the only change I made in how it was calculated was to add weight to recent testing. Someone pointed out that there were some labs closed down last year that would have shaken things up. Claude moved 6 vendors into a different group based on the increased emphasis on recent tests suggesting that they may have been impacted by the factory closings. I also added the STG data.
Here are the notes about what changed between my last version and this one:
Dataset grew significantly. 5,181 records vs 4,138 before (+25%), with 11 new vendors. The new vendors split as: 7 into Group 1, 1 into Group 3 (SPB), and 2 with insufficient cross-product data (ACR, PTB).
Recency weighting changed who matters. With a 180-day half-life from March 2026, tests from October 2025 onward carry roughly 4–8× more weight than tests from early 2025. The signatures shown for each group are based on this recent window where possible. Group 6 (Alpha-Gen, Loti Labs, Suaway Lab Research) has zero data after September 2025 — their prior pattern is preserved but flagged as stale.
The three clearest discriminating signatures, in order:
- GHK-Cu 50mg Diff% — Group 1 is at +17.0%, everyone else is between -8.8% and +3.5%. A 17-point gap is unmistakable.
- Tesamorelin 10mg purity — ranges from 93.0% (Group 3) to 98.8% (Group 2), with Groups 1, 4, and 5 in between. Still the strongest single purity signal.
- Tirzepatide 60mg — Group 3 runs negative (-3.8%) and with lower purity (98.52%) while all other groups are positive and above 99.4%. This is a new clean separator that wasn't as visible in the prior dataset.
Who would I tend to buy from? For that I asked the following question:
Which group has high quality while having good consistency in difference% that also tends towards overfull rather than under fill. When trying to figure out the actual fill to calculate dosage, I would rather have the error be in a consistent direction (plus extra product doesn’t hurt).
The result from Claude:
Group 4 is the answer.
Here's why it wins on all three criteria simultaneously:
Consistency — Group 4 has the lowest Diff% standard deviation of any group at ±15.7%. That's meaningfully tighter than G1 (±23.4%), G2 (±22.5%), and G3 (±17.7%). G5 is disqualified immediately with ±51.1%.
Overfill tendency — Average Diff% of +3.6% and 59.5% of tests coming in positive. Crucially, 45.8% of tests land between 0% and +15% — a moderate, controlled overfill rather than a random mix of large positives and large negatives averaging out to something positive.
Quality — Purity at 99.43%, which is solidly high, though G2 (99.52%) is slightly better. However G2's Diff% standard deviation is 6.8 points worse than G4's, meaning you're trading off consistency for marginal purity gain.
G2 is the close runner-up — it has slightly better purity and the highest rate of positive tests (63.2%), but the fill is noticeably less consistent than G4. If fill consistency is the priority, G4 is the clear choice. If you weight purity slightly more and can tolerate more variability in fill, G2 is a reasonable alternative.
G3 and G5 are out — G3 has the lowest purity (99.22%) and G5 has wildly inconsistent fill (±51.1%).
So my answer would be either group 2 or 4 although I have dealt with several group 2 vendors and that is the one I’ll probably stick with,
I hope this is helpful or at least interesting. Let me know if y’all have any questions or any suggestions on how I can further refine the data. Thanks.
