Parent: Corpus results. Qualitative comparison of the filtered 14,554-atom pursuit against the original 65-atom pursuit.
The original nine examples plus one SHA-256-minimum-key selection from each of 16 named families: popit3, pwsafe, Z80-sudoku-challenge, delayrepay, ask, location-db, personal, proposals, astronomy, poems, uucp, ansible, music, hosts, fitness and ssl. Selection is reproducible in review-sample.json. This is an agent spot review of representative source passages, not blind human scoring or a held-out benchmark. Every key was processed numerically; only these 25 receive qualitative verdicts.
Judgements: {'Worse': 5, 'Improved': 13, 'Mixed': 7}. Treat these as examples of benefits and failure modes, not a corpus accuracy estimate.
| Source | Original gloss | Filtered expanded gloss | Comparison | Reason |
|---|---|---|---|---|
| PROGRAMMING_RULES/philosophy | correctness; philosophy; Python | correctness notes; philosophies; arbitrary python | Worse | Correctness notes and philosophies are relevant, but arbitrary python adds a misleading specificity absent from the original correctness/philosophy/Python gloss. |
| PROGRAMMING_RULES/exceptions | error handling; Python; exception | exception handling; perl python; general principles | Worse | Exception handling is precise, but perl python introduces Perl into a Python-only policy. Negation is still absent. The direct-cosine labels are redundant but more faithful here. |
| ideas/concept-factorisation | versioning; hypertext; matching pursuit | concept notes; indexing; concept factorisation hierarchical | Improved | Concept notes, indexing and concept factorisation hierarchical describe the purpose more directly than versioning/hypertext/matching pursuit, though the phrases are clumsy. |
| first-thoughts/representation-not-language | embedding; matching pursuit; reasoning | open questions weak; note embeddings; behavioural equivalence | Worse | Open questions weak is a heading fragment. Note embeddings is relevant but the representation-versus-language thesis is not captured. More phrases have introduced boilerplate competition. |
| gdata-server/notes/design | note; API; software design | gdata notes; notes client; python module | Improved | Gdata notes, notes client and python module identify the actual component more specifically than note/API/software design. |
| envoy/design-phases | workflow; versioning; summarising | implementation notes; processing state machine; envoyresponse schema | Improved | Processing state machine and envoyresponse schema are specific and relevant. Implementation notes remains generic; this is better topical coverage, not a complete summary. |
| README/orphaning | versioning; note; orphaning | avoiding orphaned notes; metadata and links; design documentation | Improved | Avoiding orphaned notes and metadata and links capture the purpose and even retain an avoidance framing that a single orphaning label cannot express. |
| TFG/exercise-demo-protocol | exercise; demonstration; coaching | functional exercise; review and handover; sample demonstrations | Worse | Functional exercise and sample demonstrations are related, but review and handover is misplaced. The original exercise/demonstration/coaching gloss is clearer. |
| using-gps-data | GPS; matching pursuit; versioning | checks gazetteer comparisons; project notes; answered nearest station | Mixed | Gazetteer comparisons and nearest-station language are more relevant than matching pursuit/versioning, but project notes is generic and the phrases read like fragments. The expanded direct-cosine baseline is cleaner. |
| popit3/ops | error handling; logging; orphaning | maildir test completion; computer crash occurred; pops | Worse | Maildir test completion and computer crash occurred misdescribe a reference dominated by a POP3 EOF/spinning-process bug and recovery. Broader vocabulary introduces a misleading test/crash story. |
| pwsafe/transport/plan | software design; deployment; versioning | pwsafe article revision; transport plugin; current phase | Mixed | Transport plugin and current phase are relevant improvements, but pwsafe article revision describes the wrong activity. |
| Z80-sudoku-challenge/code/deterministic-fill-loop | programming; matching pursuit; fastest | deterministic fill loop; address of cell; bit candidate mask | Improved | Deterministic fill loop, address of cell and bit candidate mask are directly relevant; the old programming/matching pursuit/fastest gloss was too generic and misleading. |
| delayrepay/session-2026-08-29-kombu-decommission | logging; orphaning; command line | firefox vnc; kombu; automated via ansible | Improved | Firefox VNC, kombu and automated via ansible identify the setup being dismantled. It still misses the crucial terminated/decommissioning status. |
| ask/jobs/schema | JSON; software design; versioning | json schema; cover letters; docx generation | Improved | JSON schema, cover letters and DOCX generation capture the document-generation purpose, substantially better than JSON/software design/versioning. |
| location-db/outcodes-1deg | GPS; JSON; timetable | bristol bath newport; postcodes; cardiff gloucester swindon | Improved | Postcodes and the covered cities are meaningful. The city-list fragments remain awkward, and the geographic bounding rule is missing. This long table is chunk-averaged. |
| personal/interests | GPS; software design; astronomy | interests hobbies; GPS; john has asked | Mixed | Interests hobbies is a better umbrella label, GPS is relevant, but john has asked is boilerplate and other interests disappear. |
| proposals/section-editing | versioning; JSON; hypertext | jsonhtl documentation improvements; completed section; semantic spec | Mixed | JSONHTL documentation improvements is relevant but generic; completed section and semantic spec do not clearly convey nested containers or heading-path addressing. |
| astronomy/eclipse-2026 | astronomy; GPS; timetable | path of totality; bristol london; solar eclipse | Improved | Eclipse-specific labels and the totality path are much more informative than astronomy/GPS/link. Later place/source labels still need checking against the intended topic. |
| poems/first-to-Kathy | poetry; journey; email | kathy; summer poem; forever | Mixed | Kathy and forever capture recipient and enduring affection. Summer poem adds a season that the poem does not establish. |
| uucp/setup-status | authentication; versioning; command line | taylor uucp; existing stunnel; protocol restrictions | Improved | Taylor UUCP and existing stunnel identify the mechanism; protocol restrictions is broadly relevant. This beats the old generic authentication/versioning/command line result. |
| ansible | versioning; API; note | git ansible; private notes store; hostnames | Improved | The Ansible-specific lead is more useful than versioning/API/note. Private notes store is not a safe summary of the warning that secrets must never go into notes; polarity still needs source reading. |
| music/exercise-folk-rock | fitness; music; matching pursuit | exercise music; functional exercise; shanties | Improved | Exercise music and shanties capture the joint subject. Functional exercise is broad, but the spurious matching pursuit label is gone. |
| hosts/gravlax | versioning; logging; hypertext | webdav notes; gravlax and kelp; freebsd host | Mixed | WebDAV and gravlax/kelp are relevant; freebsd host is a wrong-host/OS association. Broader labels improve recognition while introducing a false specific. |
| fitness/bodypump-attendance | fitness; timetable; JSON | bodypump; david lloyd schedule; bristol minima shown | Mixed | BODYPUMP and David Lloyd context are more specific, but the later label must not be mistaken for a faithful description of monthly attendance counts. |
| ssl/certhub-todo | API; versioning; hypertext | certificate configuration; certhub; future work | Improved | Certificate configuration, certhub and future work capture both the domain and the status of this short backlog. |