Threads Formula 30-Day Test, Week 2: Samples Arrived, All Four Below the Median
Table of Contents
Part 5 of the "work system" series. The opening post (in Chinese) promised weekly progress data, including the ugly numbers. This is the second delivery, and the headline comes first: the sample-accumulation problem we said we would fix last week is fixed, and this week produced 4 new formula posts; but all 4 landed below the baseline median, and 3 of them had their data contaminated by a bug in our own scheduler. Three things to account for this week: the results, the bug, and a correction to Week 1's data.
The Raw Numbers: Four New Formula Posts
| Account | Published | Views | Likes | Reposts | vs. baseline median | Percentile |
|---|---|---|---|---|---|---|
| @universe_signal_tw | Aug 14, 21:32 | 160 | 1 | 0 | −56% | 18th |
| @risk.clock.tw | Aug 13, 12:43 | 55 | 0 | 0 | −55% | 23rd |
| @risk.clock.tw | Aug 13, 21:33 | 45 | 0 | 0 | −63% | 14th |
| @risk.clock.tw | Aug 14, 12:48 | 59 | 0 | 0 | −51% | 27th |
The bar we set in the opening post: the formula only has teaching value if it consistently beats the account's everyday posts by 50% or more. This week's four posts average roughly −56%. That is not missing the bar; that is the opposite direction.
The numbers go on the table as always, but not all of them deserve equal weight, for the following reason.
The Three risk Posts Come With a Discount: Our Scheduler Had a Bug
The fix we promised last week shipped late on Aug 12: formula templates now get priority in rotation, so the moment cooldown expires the next formula post gets queued instead of being starved by everyday content. The fix itself works. @universe_signal_tw is the proof: after publishing one formula post, the two remaining ones in its queue sat properly through the 4-day cooldown with no back-to-back posting.
@risk.clock.tw stepped into a trap built from two mechanisms stacking. First: the content-restock mechanism counts formula posts toward its stock level, so the moment 3 formula posts entered the queue it read as "fully stocked" and stopped adding everyday content, leaving a formula-only queue. Second: when the scheduler finds every queued item in cooldown, it publishes the first one anyway. That fallback exists so an account never goes silent, and in a normal mixed queue it never fires. Against a formula-only queue it turned the 4-day cooldown into decoration: 3 posts from the same template within 25 hours, until all three were spent and the restock mechanism refilled everyday content, at which point the cadence returned to normal. We walked the publish sequence against the code path item by item; they match exactly.
Publishing the same template to the same audience three times in a row invites both audience fatigue and algorithmic downranking of repetitive content, and both can push numbers down. That is our leading hypothesis for why these three ran so low; there is no control group, so it cannot be proven, which is why the treatment is tagging them as "bug-condition samples" in a separate segment, never pooled with normal-cadence samples, and never promoting the hypothesis to a conclusion. The fix is decided: the fallback changes to skipping the slot when a queue holds nothing but cooling formula posts, and the restock mechanism stops counting formula posts as stock. The ship date will be noted in the next update.
The @universe_signal_tw post at 160 views is the clean sample: normal rotation, normal cooldown, published at 21:32 in the account's strongest slot. Non-formula posts in that same slot on the surrounding three days pulled 675, 311, and 551 views. The clean sample lost with even less excuse.
The Most Important Observation This Week: Reposts Showed Up, Just Not on Formula Posts
Week 1 floated a hypothesis: accounts this small may simply lack repost fuel, because the absolute number of natural resharers inside a few hundred views rounds to zero.
This week partially refuted that. @universe_signal_tw's non-formula posts (everyday science content) collected 4 reposts this week, including one post that reached 896 views. The fuel exists.
The 6 formula posts published to date still have zero reposts, all of them.
For the formula, that is a worse signal than "no fuel." The fuel is there; the fire did not start. The stance formula's core hypothesis is that a stance triggers reposts and reposts drive reach. Right now the same account with the same audience reposts science content about colliding galaxies and does not repost stance declarations. Four reposts is itself a tiny sample, so this stays an observation under watch, not a verdict.
A Correction to Week 1's Data
Week 1 described @risk.clock.tw's control group as "just 10 posts in the 30 days before launch, median 102." This week we rewrote the data-pull script to be rerunnable and checked it into version control, and the complete pull found the same time window actually contains 22 posts with a median of 121. The Week 1 pull was incomplete and missed posts from early July.
Recomputed against the fuller baseline, Week 1's 276-view post (277 on this week's re-pull; views accrue naturally, the same 489-to-492 effect Week 1 already explained) lands at the 86th percentile, slightly higher than the 80th we published. The directional conclusions do not move: the baseline is weak and the account restart is a confounder, both caveats stand.
@universe_signal_tw's baseline also updates from 119 posts to 121, and its median from 348 to 366. The difference has two sources: the complete pull found 2 more posts inside the window, and insights are alive, with month-old posts still slowly accruing views today, so the baseline itself drifts. From this week on, percentiles are computed against the completed baseline, snapshotted once per report.
Get it wrong, fix it, say so. That is the rule the opening post set, and it applies to our own numbers too.
Account-Level Numbers
Followers: risk 1,513 → 1,509 (−4), universe 280 → 282 (+2). Single-digit weekly moves; we attribute nothing to them.
Non-formula medians in the current window are 267 for risk and 390 for universe, with both accounts publishing every day on schedule and everyday content performing normally. The formula posts' deficit is not an account-wide slump; on that point the data is clean.
How the Numbers Are Computed
Same rules as Week 1: control group = every post from the same account in the 30 days before launch, insights pulled post by post; medians and percentiles, never means. Two additions this week:
- The baseline pull now runs from a fixed script under version control, rerunnable and auditable. This week's correction is exactly what that produced.
- Segmentation rule: the rotation fix that shipped late on Aug 12 is the segment boundary, and post-fix samples are further tagged "normal cadence" vs. "bug condition." Contaminated samples are never pooled.
Next Update
The next update is planned for around August 24. Three things to watch: normal-cadence samples after the scheduler-bug fix ships (universe still has 2 formula posts queued, expected to rotate out after cooldown expires on Aug 18), clean risk samples (its formula posts are all spent, and a new batch will be queued before the next update, running through normal rotation), and whether any formula post finally records its first nonzero repost. The full experiment design and formula text are in the opening post.
The formula templates and fatigue-cooldown mechanics used in this test are built into the MindThread formula marketplace. Data updates weekly as promised in the opening post: if it replicates we say so, if the fuel is missing we say so, and if the bug is ours we say that too.