Info

Method, statistics, controls, data handling, sources.

1. Protocol
2. Verdict classes
3. Statistics
4. Controls
5. Cross-source score
6. Rig
7. Data handling
8. Limits
9. Sources

1. Protocol

Scoring rule written to docs/agents/*.md, committed alone, sha256 in the commit message.
Source read once, polite rate (3 s gap, one backoff on 429 or 5xx, then stop). Identifiers keyed and hashed in memory.
Rule applied unchanged. Any rule written after seeing data is labelled post-hoc.
Compared with a control. Verdict, numbers and rule sha appended to the decisions inbox.
Person reads the page before publication.

2. Verdict classes

Class	Condition
Confirmed	frozen rule passed, every control beaten, independent record agrees
Open lead	at least one frozen rule passed and one control beaten; not confirmed
Found, not covert	real pattern, ordinary explanation (one owner, open program, tagged)
Null / closed	rule fails, or pattern disappears under a control
Incomplete / unobservable	coverage or schema insufficient; not a null

3. Statistics

Symbol	Definition
S	score of a day or group under the frozen rule
B	baseline score, ordinary-day median
K	count of independent sources flagging the same UTC day; HIT bar K≥3 and p≤0.01
p_ord	share of ordinary days with score ≥ observed
p_shuf	share of shuffle draws (2,000 unless stated) with score ≥ observed; minimum about 0.0005
p_ordK	share of days with K ≥ observed K (2 of 136 = 0.0147)
p_global	probability that max K over all days reaches the observed value under circular shift of each source
circular-shift p	share of random rotations of one source timeline with ≥ observed shared days
L_hat, r	lead-lag offset in hours of the cross-correlation maximum, and the correlation at that offset
AUC	rank statistic, 0.5 chance; bootstrap 95% CI where stated
ARI	adjusted Rand index of two partitions, 0 chance
Spearman ρ	rank correlation of two series

4. Controls

Control	Procedure	Used for
Ordinary day	score days with no known event by the same rule	daily-count bursts
Shuffle	permute event times within day or hour, 1,000 to 10,000 draws	time-clustered groups
Circular shift	rotate one source timeline by a random offset	two sources flagging the same day
Activity-matched	draw comparison owners or accounts with equal volume	owner or account groups
Degree-preserving	rewire reply edges, node degrees fixed	reply graphs
Rarity-matched	draw names from the empirical background multiset, not a within-window shuffle	same-name multi-owner tests

5. Cross-source score (29 Sep)

Source	S	B	Rank	Note
PyPI	33 first releases in 164.8 min			one hashed owner; 32 one-version
Wayback CDX echo hosts	0.83	0.09		23 captures
npm	37		37 / 62	not flagged
HF datasets	102		23 / 63	not flagged

K=2. p_global=0.1356. p_ordK=0.0147 (2 of 136 days). Pair circular shift: httpbin+ntfy p=0.0082; httpbin+PyPI p=0.092 (shared n=21). Lead-lag L_hat=-11, r=0.252, p=0.597. Status: open lead.

6. Rig

Language	Python 3.9, standard library plus requests; static HTML with inline SVG
Hash	sha256 for rule files; keyed hashes of identifiers, key dropped after the run
Parallelism	independent runs in separate git worktrees and branches; results merged by a person
Local data	about 150 GB gzip (AI Village HF dataset), inside gitignored folders

7. Data handling

Kept: counts, dates, hashed identifiers, scoring code.
Not kept: page text, post text, usernames, addresses. Network addresses reduced to operator counts.
Datasets released under terms stay local and are used within those terms.
No scanning or probing of hosts, no login or paywall bypass, no contact with the sites measured.
No attribution to any person, lab or company.

8. Limits

Public feeds drop history; an empty old date can mean the record is gone.
Single-day and single-source results are leads, not findings.
Code and drafts were produced with AI assistants; rules, numbers and wording are checked by a person before publication.
One reader checks everything; corrections welcome.

9. Sources

collusion.wiki and its dataset (wiki incident, documented first); recovered here independently from the wiki public change log.
AI Village public chat and dataset (sanctioned coordination example).
Public records named per row in the ledger. Full URL list: /observatory, section 11.

Corrections and data requests: contact details on maramasaeva.com.