Chronicler · human view
Chain C — the classifier the plan skipped
PLAN-004 jumped over the structure classifier. After the keyword B null, what is Chain C actually, and what remains host-narrow or unconsumed?
- Dispositionlive, maintained and deliberately un-promoted
- arc
library/work/chain-c/README.md
What the account is
A live, maintained, deliberately un-promoted library arc. One file with no local imports — b_structure_classifier.py — plus contract, tests, and optional retrieval flags. It labels already-derived chronology text into five classes. It does not write history or replace conversation-core. unsure is never silently dropped.
Why it exists
PLAN-004 sequenced review → readers → reconcile and skipped Chain C. Keyword B-signal ran instead; REQ-132 found term-scored B added ledger echo, not new reader claims. FIND-137 closed that chain on 2026-08-14. FIND-138 corrected the over-reading the same day: the chronology classifier was not finished. Chain C is the different thing that was skipped — structure, not another keyword contest.
Standing, three readings
Disposition live under REQ-178. Delivery: REQ-178 and REQ-197 fulfilled, both CHANGEs complete. Evidence sound on Grok, host-narrow elsewhere, and still unconsumed — FIND-264 named that nothing consumes the labels at scale, and that leftover still earns keep.
Host-narrowness, measured
FIND-1047: across eight sessions per family, the file-read guard fired on Grok 578/2244 (25.8%), Claude 0/1265, Codex 0/303. Six of fifty Claude cat -n reads land in repeated_ledger — the label the guard exists to prevent. Codex is nearly inert (95% unsure). Match on what the house writes, not on how a host wraps it.
Session shape
A standing-office chat is not a uniform blob. Head and tail grow schema (enter ceremony, retirement ceremony); the middle is the day. Bookends are envelope, not history. Newer Corpus imports are a fairer teacher than the six old blocks.
Frontier
Two edges: detection is host-narrow (FIND-1047), and the FIND-264 gap is still open — retrieval flags are the only consumer and went cold with the corpus on 2026-08-29. Reopening wants a newer archive, a named question, and a measurement that separates fewer errors from better error descriptions. Promotion stays behind a later gate.
Surveyor · skip, live labeller, host-narrow
Structure drawn from the export's movement table, standing readings, and FIND-1047 host table. No wall-clock bars — host counts are stage evidence, not suite timings.
- 1PLAN-004 skipped Chain Creview → readers → reconcile ran; keyword B-signal instead
- 2Keyword B nullREQ-132 / FIND-137: term-scored B added ledger echo, not reader claims
- 3FIND-138 correctionClassifier not finished; PROP-050 and REQ-178 keep Chain C live
- 4Live labellerREQ-178 maintained; 21 tests passing; five labels; unsure never dropped
- 5Optional retrieval flagsREQ-197 --chain-c-prefer / --chain-c-drop; only consumer
- 6Unconsumed labelsFIND-264 gap still open; tool went cold with corpus 2026-08-29
- Grok2244 tool bodies · file-read guard 578 (25.8%) · FIND-1047
- Claude1265 tool bodies · guard fired 0 · some cat -n → repeated_ledger
- Codex303 tool bodies · guard fired 0 · nearly inert (95% unsure)
Mermaid source
flowchart LR
Skip["PLAN-004<br/>skipped C"] --> KW["Keyword B null<br/>FIND-137"]
KW --> Corr["FIND-138<br/>correction"]
Corr --> Live["REQ-178<br/>live labeller"]
Live --> Hard["v0 harden"]
Hard --> Flags["REQ-197 flags"]
Flags --> Gap["FIND-264<br/>unconsumed"]
subgraph hosts ["Host-narrow FIND-1047"]
Grok["Grok 578/2244"]
Claude["Claude 0/1265"]
Codex["Codex 0/303"]
end
Live -.-> hostsSurveyor · wall-clock figures
Bars use only numbers verified in the exported markdown for this arc. Each row shows value and source commit SHA. Gaps stay gaps.
No verified wall-clock figures in this export — structure above carries the weight.