The Methodology Reference

Wave 4 corpus expansion — findings

Additions (4 Mns): cassandra, valkey, word-processor, xmpp. Rationale: test singleton-decomposition hypotheses (each addition predicted to resolve or weaken a specific singleton). Also address user-flagged backend-software corpus bias by adding office productivity (word-processor) and federated protocols (xmpp).

Hypotheses tested

Wave 4 MnPredicted nearest-neighborResult
cassandramongodb (test "non-relational NoSQL" cluster)CONFIRMED: cassandra ↔ mongodb cos_n=0.983 (tech-scope pair); spine_1 N=6 cluster forms
valkeyredis (test "substrate-twin separates from ecosystem" hypothesis)CONFIRMED: valkey ↔ redis cos_n=1.000 (perfect substrate twin); spine_9 N=2 at 11/11
word-processorspreadsheets (test "office-apps same-modulo-X" hypothesis)PARTIAL: pairwise cos_n=0.974 confirmed, but didn't form spine at 9/12 threshold
xmppsmtp-email (test "federated-protocol-pair" hypothesis)PARTIAL: pairwise cos_n=0.983 confirmed, but didn't form spine at 9/12 threshold

Detailed empirical results

Cassandra → consensus-kv-substrate ABSORPTION (unexpected breadth)

The consensus-kv-substrate anchor (previously consul + couchdb + etcd + zookeeper at N=4) ABSORBED cassandra AND mongodb post-Wave-4 to become a 6-Mn cluster at 10-12/12 strict threshold:

spine_1 (N=6, avg=11.1, min=10/12): cassandra, consul, couchdb, etcd, mongodb, zookeeper

This is a structural finding that the existing anchor name no longer fits: "consensus-kv-substrate" implied coordination systems (etcd/zookeeper) plus replicated document storage (couchdb). With mongodb + cassandra now joining at 10-12/12, the unifying structural property is broader: peer-replicated data systems with tunable consistency. Possible reinterpretations:

The post-Wave-4 anchor inventory needs revision.

Valkey ↔ redis — pristine substrate-twin demonstration

valkey was authored as a structurally-identical substrate twin to redis. Result:

Methodological value: valkey is the cleanest empirical demonstration that the framework separates substrate from ecosystem. Same Mn-substrate at different ecosystem-tier positions yields predictable pair-behavior. Paper 6 cite-worthy.

Word-processor near-twin to spreadsheets — confirmed user hypothesis

User observation: "word-processor and spreadsheets look the same — the difference is probably the execute."

Empirical confirmation:

The user hypothesis is essentially correct (cos_n=0.974 is very strong) but the broader cluster of "interactive content creation tools" (vscode, notion, wikipedia, spreadsheets, word-processor, obsidian, figma) is even more cohesive. Office-apps-pair is real but it lives INSIDE a larger interactive-content-tools cluster.

At threshold 9/12 the office-app pair didn't form a spine because the broader cluster pulls each member toward different K-means partitions across configs.

Xmpp ↔ smtp-email — confirmed at pair level, lost at threshold

Pairwise smtp-email ↔ xmpp cos_n=0.983 (very strong). But neither appears in any spine_cluster at threshold 9/12. Both classify as singletons.

Interpretation: smtp-email and xmpp ARE structural near-twins (federated-messaging-protocol substrates) but the corpus partitioner places them in different K=6 clusters in MORE than 3-of-12 configs. The federated-protocol cluster is a genuine 2-Mn pair that needs a 3rd member to cross threshold robustly.

Candidate 3rd members: irc, sip, gemini-protocol, gopher. None currently in corpus.

Other Wave-4 cluster shifts (not from new Mns directly)

Adding 4 Mns reshuffled K-means cluster boundaries elsewhere:

Methodological findings

1. Anchor membership is corpus-incremental AND threshold-sensitive

Adding 4 Mns reshuffled cluster boundaries enough that 5 of the 10 inductive anchors had their membership change. Specifically:

This is consistent with the framing established earlier ("anchors aren't final") but the magnitude of shift is greater than expected. Two implications:

2. Threshold-9 catches robust co-clustering, misses strong-but-config-fragile pairs

Pairs at high pairwise cos_n (0.974-0.983) can fail to surface as spine clusters because K=6 partitions place them in different clusters under enough configs. This is a separate methodological signal: "strong-but-fragile" pairs need either (a) corpus expansion to add a 3rd member, or (b) lower-threshold reporting (7/12 or 8/12).

Recommendation: report TWO threshold levels in cluster outputs — strict (9-10/12) for robust anchor candidates, looser (7-8/12) for "near-anchor pairs" worth tracking.

3. Wave 4 partially confirmed corpus bias hypothesis

Two of four new Mns address backend bias:

The corpus is still backend-heavy. Adding presentations, emacs, irc, gemini-protocol would further test bias-correction.

Recommendations for next moves

  1. Anchor inventory revision. Update existing anchor JSONs with post-Wave-4 cluster boundaries OR explicitly note them as corpus-state-dependent. The consensus-kv-substrate anchor in particular needs name/description revision.

  2. Lower-threshold anchor candidate tracking. Add a "near-anchor pair" category at threshold 7-8/12 for pairs that are structurally close but don't form spines at strict thresholds. Examples: smtp-email + xmpp; spreadsheets + word-processor; sqlite + datomic.

  3. Paper 6 corpus-state disclosure. Add explicit corpus-N + corpus-bias acknowledgment to Paper 6's methodology section.

  4. Anchor-space topology analysis (next move from session pickup): now that the post-Wave-4 cluster topology has shifted significantly, computing inter-anchor centroid distances becomes more interesting. Some anchors may be MUCH closer than they look from membership alone.

Status