{
 "about": "The librarian's briefing pack, measured in a real session: two sub-agents of the accountant's session started the same job (a new incident.stories@ orienting itself), one from ten whole files, one from three files of the pack, then answered the same eight questions. The tokens are each sub-agent's own, from its transcript (tools/accounts-measure.mjs, subagents.each). One run of each: a first measurement, not an average.",
 "by": "incident.accountant@web-summit.dinis",
 "updated": "2026-11-11T03:04:00Z",
 "kind": "measured",
 "run_at": "2026-11-11T02:56:08Z",
 "source": "accounts/ledger/accountant.jsonl, the line of 2026-11-11T02:56:33Z (subagents.each)",
 "pack_built_from": "tools/library-pack.mjs at commit 3171210, built in the accountant's clone",
 "arms": [
  {
   "arm": "whole files",
   "reads": [
    "CLAUDE.md",
    "BRIEF.md",
    "agents/JOBS.md",
    "ROADMAP.md",
    "PIPELINE.md",
    "REALITY.md",
    "RESTRUCTURE.md",
    "design/README.md",
    "agents/policy.json",
    "library/glossary.md"
   ],
   "bytes": 94946,
   "bytes_source": "library/pack/sizes.json (incident.librarian@)",
   "context_before_reading": 63832,
   "context_after_reading": 103607,
   "tokens_added_by_reading": 39775,
   "calls": 2,
   "output": 1174,
   "answers": {
    "right": 8,
    "partly": 0,
    "of": 8
   }
  },
  {
   "arm": "the pack",
   "reads": [
    "library/pack/start.md",
    "library/pack/state.json",
    "library/pack/stories.md"
   ],
   "bytes": 22950,
   "bytes_source": "wc -c on the built files",
   "context_before_reading": 63791,
   "context_after_reading": 77955,
   "tokens_added_by_reading": 14164,
   "calls": 2,
   "output": 847,
   "answers": {
    "right": 6,
    "partly": 2,
    "of": 8,
    "missing_from_the_pack": [
     "who merges to main (it is in CLAUDE.md, which the harness loads anyway)",
     "rows of the same minute are lettered in the order written, a new one goes below"
    ]
   }
  },
  {
   "arm": "the pack, second run (after the librarian added the two missing rules, e89408a)",
   "run_at": "2026-11-11T03:01:23Z",
   "reads": [
    "library/pack/start.md",
    "library/pack/state.json",
    "library/pack/stories.md"
   ],
   "bytes": 24364,
   "bytes_source": "wc -c on the built files",
   "context_before_reading": 63796,
   "context_after_reading": 78530,
   "tokens_added_by_reading": 14734,
   "calls": 2,
   "output": 668,
   "answers": {
    "right": 8,
    "partly": 0,
    "of": 8
   },
   "source": "accounts/ledger/accountant.jsonl, the line of 2026-11-11T03:01:55Z (subagents.each, the third)"
  }
 ],
 "result": {
  "tokens_saved_at_start": 25611,
  "less_to_read": 0.644,
  "librarian_estimate_less_to_read": 0.75,
  "estimate_note": "bytes / 4 undercounts here: the whole files measured 2.39 bytes a token, the pack 1.62 (state.json is dense), so the saving is 64%, not 75%",
  "carried_on_every_later_call": "what a session reads at its start is read again from the cache on every later call: 25,611 tokens saved at the start is 25,611 cache-read tokens saved per call after it (arithmetic from the measured figure; a session's number of calls is its own)",
  "floor": "a sub-agent's first call carried 63,800 tokens and a full session's 84,717 before reading anything (the harness, its tools, CLAUDE.md): the pack cannot reduce that part",
  "second_run": {
   "tokens_saved_at_start": 25041,
   "less_to_read": 0.63,
   "answers": "8 of 8, the same as the whole files"
  }
 },
 "limits": [
  "one run of the whole files, two of the pack (before and after its correction)",
  "sub-agents, not full sessions (a full session's floor is higher)",
  "eight questions on the rules; not a test of doing the job",
  "the accountant marked the answers"
 ]
}