{
 "arms": {"plain": "Claude Code talks to Anthropic direct.", "recall": "Claude Code talks to Anthropic through Context Mode Gateway, Recall on."},
 "client": "Claude Code 2.1.283, claude-opus-5-5, the client's real 1M window, thinking off",
 "copied_as_is": "Every JSON file of both runs, as committed. Only these changes: every tenant value is \"withheld (disposable test tenant)\"; prereg names the commit that holds the pre-registration; internal file and route names inside free-text fields are replaced by the plain words listed in scrub; renamed fields are listed in rename.",
 "cost_basis": "US dollars at API list price: the higher of the gateway's ledger and the client's own figure, every call priced at list (see usd_source in each pair file).",
 "date": "2026-10-03",
 "rename": [
  {"to": "ledger_usd_sum", "what": "the gateway's sum of its ledger rows, renamed from its internal field name"}
 ],
 "runs": [
  {
   "derived": {
    "recall_net_usd": {"from": "gateway.recall_stats.saved_usd_recall in each pair-<k>-recall.json", "high": -0.217194, "low": -0.368532, "mean": -0.324899, "n": 4}
   },
   "fields_changed": ["checker","ledger_usd_sum (renamed)","prereg","tenant","usd_source"],
   "files": ["/benchmarks/long-session/ab.json","/benchmarks/long-session/dry-plain.json","/benchmarks/long-session/dry-recall-b4k.json","/benchmarks/long-session/dry-recall.json","/benchmarks/long-session/dry.json","/benchmarks/long-session/dry2-recall-b4k.json","/benchmarks/long-session/dry2-recall.json","/benchmarks/long-session/dry2.json","/benchmarks/long-session/dry3-plain.json","/benchmarks/long-session/dry3.json","/benchmarks/long-session/pair-1-plain.json","/benchmarks/long-session/pair-1-recall.json","/benchmarks/long-session/pair-1.json","/benchmarks/long-session/pair-2-plain.json","/benchmarks/long-session/pair-2-recall.json","/benchmarks/long-session/pair-2.json","/benchmarks/long-session/pair-3-plain.json","/benchmarks/long-session/pair-3-recall.json","/benchmarks/long-session/pair-3.json","/benchmarks/long-session/pair-4-plain.json","/benchmarks/long-session/pair-4-recall.json","/benchmarks/long-session/pair-4.json","/benchmarks/long-session/spend.json"],
   "id": "final",
   "prereg": "pre-registered in gateway commit f2b5c567 (2026-10-03T07:01:17+03:00), before the first pair",
   "result_file": "/benchmarks/long-session/ab.json",
   "source_commit": "9fbdf3c0",
   "title": "Long agentic coding session, Opus 5.5, 1M window",
   "what": "Plain Claude Code against Claude Code through Context Mode Gateway (Recall on). 30 steps: 20 work steps, then 10 memory questions with every tool forbidden."
  },
  {
   "derived": {
    "recall_net_usd": {"from": "gateway.recall_stats.saved_usd_recall in each pair-<k>-recall.json", "high": -0.176683, "low": -0.399376, "mean": -0.288259, "n": 4}
   },
   "fields_changed": ["checker","ledger_usd_sum (renamed)","prereg","source","tenant","usd_source"],
   "files": ["/benchmarks/long-session/amendment-4/ab.json","/benchmarks/long-session/amendment-4/dry4-plain.json","/benchmarks/long-session/amendment-4/dry4-recall-b4k.json","/benchmarks/long-session/amendment-4/dry4.json","/benchmarks/long-session/amendment-4/pair-1-plain.json","/benchmarks/long-session/amendment-4/pair-1-recall.json","/benchmarks/long-session/amendment-4/pair-1.json","/benchmarks/long-session/amendment-4/pair-2-plain.json","/benchmarks/long-session/amendment-4/pair-2-recall.json","/benchmarks/long-session/amendment-4/pair-2.json","/benchmarks/long-session/amendment-4/pair-3-plain.json","/benchmarks/long-session/amendment-4/pair-3-recall.json","/benchmarks/long-session/amendment-4/pair-3.json","/benchmarks/long-session/amendment-4/pair-4-plain.json","/benchmarks/long-session/amendment-4/pair-4-recall.json","/benchmarks/long-session/amendment-4/pair-4.json","/benchmarks/long-session/amendment-4/spend.json"],
   "id": "amendment-4",
   "prereg": "pre-registered in gateway commit 5e8b5c85 (2026-10-03T15:07:40+03:00), before the first pair",
   "result_file": "/benchmarks/long-session/amendment-4/ab.json",
   "source_commit": "9fbdf3c0",
   "title": "Amendment 4: the same session, archive search allowed",
   "what": "As above, but on the gateway arm's 10 memory questions the one allowed tool is the archive search (context-mode-search). Plain is unchanged."
  }
 ],
 "schema": "context-mode/long-session-evidence/v1",
 "scrub": [
  {"from_kind": "internal file name", "to": "the gateway's model price table"},
  {"from_kind": "internal file name", "to": "the harness's exact-match checker"},
  {"from_kind": "internal admin route", "to": "the gateway's exec log, mode gateway"},
  {"from_kind": "internal admin route", "to": "the gateway's exec log"},
  {"from_kind": "internal admin route", "to": "a whole-doc archive read"},
  {"from_kind": "internal admin route", "to": "archive search"},
  {"from_kind": "internal admin route", "to": "Recall's ledger stats"}
 ]
}