The Audit Came Home: 1.8 Billion Tokens, 95% Re-read — the Cure Was Forgetting

Episode 4 ended with a bill that stung: the plant’s only AI call paid ~97% door fee, and was rebuilt the same day to a twenty-fifth of the price. One question survived the rebuild. The audit had covered the plant — a single call, frugally scoped. But the audit itself had been run from a desk where AI is used the way water is: interactive sessions all day, a chat gateway answering a messaging bot, research relays that run for days. What did the operator’s own desk burn?

The answer rearranged the furniture. Thirty days of the desk’s interactive coding-agent sessions — 85 of them — came to ~1.8 billion tokens, ~95% of it cache re-reads. The plant’s entire audited AI life, the 2,239,178 tokens that shocked Episode 4, is about one part in a thousand of that book (computed). The whale was never in the factory. It was in the chair.

The other bill

The mechanism was already familiar — Episode 4 called it door fee. A coding-agent session re-reads its own system prompt, tool catalog and house rules on entry; an interactive session then keeps re-reading everything already said, on every message, forever. The plant’s lane paid door fee per call. The desk pays it per sentence.

Book, as of 2026-09-12	Size	Nature
Desk: interactive sessions, 30 days	85 sessions · ~1.8B tokens	human exploration and debug; ~95% cache re-reads
Chat gateway’s own ledger, since 07-25	2,564 calls · 263M tokens	a messaging bot’s brain; details below
Plant: the AI-summary lane, whole life	44 calls · 2,239,178 tokens	the audited lane of Episode 4 — already thin

The desk book concentrated exactly where the plant book had been frugal: the top five sessions were 39% of it, the top ten 64% — all multi-day relay jobs, hundreds of turns each. The heaviest single session ran 667 turns and 197,060,000 tokens.

Falsify first, then treat

The audit’s discipline was the one Episode 2 taught the plant: before changing anything, attack every proposed fix with the real numbers. Three verdicts came back in one sitting.

Compaction — confirmed. Counterfactual simulation on that heaviest session: press the history into a summary at turn 300 (~35K of minutes), and the back half’s re-reads fall from 144,730,000 to 12,810,000 — −91% on everything after the fold, and the back half had been 73% of the session’s whole cost. The mathematical optimum lands just past a session’s midpoint; the house trigger line was set at 300 turns or 250K of context.

Background slimming — falsified. The full standing background — house rules plus the memory index, the files every session carries — is ~6% of consumption. A plausible fix, killed by data in under an hour and demoted.

Model tiering — a trap. Switching to a cheaper model mid-session sounds like thrift. One receipt from the stress test: 2,777 fresh-input tokens against 154,944 served from cache — ~98% of the context arriving at the cache price, and cache bills at ~18.5% of fresh input. Switch the model and the cache no longer matches: the re-read re-prices to full fare, a multiplier of more than five (computed). The rule flipped from “downgrade when easy” to pick the tier at the door, never mid-meeting.

The rules that landed — and the one that never had

Four changes went into the house that day. The house rules gained a token-discipline section: three cache bans (no mid-session model switch, no mid-session edits to standing background, oversized tool output always truncated), the compaction trigger lines, and an anchoring rule — before any compaction, write the live conclusions and open questions into a file, so the minutes cannot orphan the work. The memory network was re-woven, 40 lines down to 38. The plant project moved out of the family living room: its sessions — most of the desk’s heavy relay work — had been hauling the household’s entire memory index on every plant job, so the project got its own rulebook and its own nine memories, and a door of its own to enter by. Episode 2’s map, it turned out, had been violated most thoroughly by its own drafting desk: every job its smallest seat — except the operator’s. Episode 5’s staircase had drawn the same rule for the plant — every job stands as low as it can — and this was the desk’s own confession: its heavy relay jobs had been living on the top step.

And the verification pass caught a real one. A red line the operator had ordered on September 8 — no auto-push, ever — turned out to have never been indexed into the memory every session loads. Written into a file, four days earlier; never on the wall. A house rule that never reached the wall was never written. Re-indexed on the spot.

The method becomes a tool, and the tool finds a patient

The audit method itself was pressed into a reusable instruction — an audit prompt any machine in the fleet can be handed, with the discipline written in blood: no number without its measurement basis; debug traffic and steady state reported separately; the auditor’s own consumption watched, because an audit that becomes the largest consumer of the day has failed twice. Its checklist reaches the places token audits usually miss — text injected into every prompt by hooks, scheduled jobs paying appearance fees, big documents read whole when a grep would do.

Its first patient was the chat gateway — the self-hosted agent behind a messaging bot. Its ledger, read from the gateway’s own store: 2,564 metered calls since late July, 263M tokens, cost-equivalent to ~91.5M fresh-input tokens at the cache rate. Same disease as the desk — 70–97% of it re-reads — plus two lesions all its own.

A heartbeat tax. The gateway polls for work every 30 minutes, and each poll carries the session’s full history — ~159K tokens — to the cloud model, which almost always answers with a stock four-token silence. The audit counted 92 of these low-value polls hauling 16,600,000 tokens. Every half hour: the entire family silver taken out of the vault to ask whether anyone needs a spoon, and the answer, 92 times running, was no.

An immortal session. The gateway’s main session had been born on September 3 from a one-line test message — reply with ok and nothing more — and then never ended. Nine days, 118M tokens, 11.19M of them on the morning of the audit itself.

The treatment, verified at the wallet

A stress test of four questions — what identifies a session, what the operation path touches, what it costs, which of the candidate cures survives — chose compaction over a hard reset: keep minutes, not amnesia. The main session went from 160,000 of a 205,000-token window (78%) to 41,000.

One trap on the way, classic distributed-systems fare: the command-line call timed out at 120 seconds — and meant nothing. The gateway was still compacting on its side. A blind retry could have double-pressed the history; instead the treatment was watched until it landed. Then the proof that matters: the wait for the first real heartbeat after compaction, and its receipt — 41K in, 4 out. The wallet, not the console, closes a treatment.

The closing review asked whether every job was actually done. It was not: three more sessions sat over the line, one of them at 216,000 — 106% of the window, already over the edge. All three pressed in turn: 216,000→23,000, 120,000→23,000, 115,000→20,000. Final count across the gateway: 42 sessions, none over the 100,000 line; the four pressed sessions each ended at a quarter or less of their former size (computed).

The wallet-level pass also surfaced something unrelated, which is what wallet-level passes are for: a monitoring job had failed to deliver 346 messages — every send going bare to a long-blocked external endpoint. Data collected, reports composed, nothing delivered, for weeks. A proxy was restored in three places; the endpoint answered 200; and the first scheduled run after the fix, that same evening, completed with zero send failures — the streak ended the night it was found. One stumble worth keeping: a hand-retyped credential failed auth, and the credential read back verbatim from the script passed. Never retype what you can re-read.

The soul question

Then the operator asked the question this episode exists to carry. You compressed something to save tokens. What exactly — the memory, or the context? Will something be lost?

┌─ model knowledge     the brain        nobody can press it
├─ long-term memory    the notebook     files compaction never touches
└─ conversation        the recording    this is what gets pressed

Compaction is the meeting-minute move: the ever-growing recording is replaced by one page of minutes, and the meeting continues with the minutes plus the last stretch of verbatim transcript. What is lost is real and is stated plainly, because a savings story that hides its price is marketing: exact phrasing, dead ends, days-old asides — anything the minutes didn’t keep. What survives: conclusions, decisions, recent messages, the notebook — and the originals themselves, archived with checksums, 58 entries so far, recoverable in principle. The recovery flow has never been rehearsed; this record does not claim it has.

Nothing about the operator’s own interactive sessions changed that day. The group bot’s grip on days-old chat detail softened — the price, paid where the tokens were actually going.

Which is the principle the whole day distilled into: transcripts are consumables; memory is the asset. Anything that matters gets written down — into manuals, worklogs, rulebooks — and nothing important is ever left depending on an AI remembering a conversation.

What the day settled, and what it cost

The ledger, estimated at current intensity and labeled as estimates: the gateway’s four compactions are worth ~500M tokens a month if the periodic discipline holds — group sessions rebound; the desk-side compaction discipline, law passed and first battle unfought, 300–700M at half-credit; the plant lane was already fixed in Episode 4. Together, roughly 0.8–1.2B tokens a month — about half of a ~2B monthly burn. The catch is the same sentence twice: these are estimates, and laws without enforcement are decoration.

And the cost side, stated as loudly as the savings: four sessions’ verbatim history became a summary. Every future compaction will pay the same. The trade, in one line — you trade what can be written down for what no longer needs re-reading — and the desk now writes things down.

Sources and method

First-party: the operator’s field records of 2026-09-12 — the 30-day session aggregation computed from per-session usage fields (session count and concentration shares verbatim from the aggregation output), the counterfactual compaction simulation’s own printed results, the gateway’s local ledger store read read-only (call count, token totals, heartbeat poll counts, the archive table with its 58 checksummed entries), the house-rule and project-rulebook files as committed, and the before/after session watermarks. Estimates are marked as estimates; derived ratios are marked computed. Assembled into English with AI assistance under human editorial direction; facts and numbers unchanged from the records. Deliberately absent, per the series’ disclosure policy: anything identifying the plant, the operator, the fleet’s machines and network, the gateway and model products used, messaging-platform identifiers, addresses and credentials. The measured numbers are registered in the /data/ ledger.

All episodes — Machines Keep the Watch, a field-record series. The map: the anchor · Episode 1: the dress rehearsal · Episode 2: the second without AI · Episode 3: the blank-paper exam · Episode 4: the bill · Episode 5: the staircase.