The receipts
Every number Peddra says about itself lives here first. What is claimed, what was measured, and the line that cannot be crossed when it is said.
A number that is not on this page does not get said, however true it feels.
Peddra is built by three people; the corpus is the founders' own.
Every claim is checked against the measurement before it appears in public.
Dollar figures on this page are list-rate value consumed, not cash paid; for subscription use the public units are tokens, turns, and percent.
“I metered 21 billion tokens of my own Claude Code use.”
2026-08-24 run: cache read 21,688,523,492 tokens; total across all four token classes higher. 77 sessions, 82 distinct ids, 72,337 turns, 13 projects.
swift run -c release meter, 2026-08-24; an internal review of record banks 21 billion as the case-study scale word, never an account count.
That any of this is customer data, or anyone's outside the company. The case study speaks as “I” so the privacy line and this number never contradict: one account, measured by its owner.
“97 of every 100 were Claude re-reading what it had already read.” At the wall: 98 of every 100.
Wall window 720,380,476 deduped, 98.0 percent re-reads, verified to the cent 2026-08-24 before the 0.1.5 cut.
The launch-day record of 2026-08-24.
“About a third” or “nearly a third” of the spend didn't have to happen. The bare “a third” is retired while the measured share sits under 33.3 percent.
2026-08-24: $6,418.35 (list-rate value consumed, not cash paid) avoidable of $20,256.83 list-rate value consumed. that is 31.68 percent.
A third is 33.3 percent. The corpus measures 31.68. The 33 percent figure was banked when the corpus was smaller; it has grown under launch-week load and the share has fallen. The published claim now sits marginally above the measured value, which inverts the sandbag law.
Founder's words: “Nearly a third is good.” Later the same day: “Can we say about 1/3?” Ruled allowed. Tripwire: if a measured run drops below 30 percent, “about” dies and the claim is re-ruled. Re-hardening to a bare “a third” needs a run above 33.3 percent and a fresh ruling.
Nothing yet without a ruling.
PUBLISHABLE as the floor only: the published floor phrase is “a third more Claude.”
At 31.68 percent avoidable, 1 / (1 − share) gives 1.464x. The site copy prints “a third more Claude” as a deliberate floor, with the arithmetic note of record recording the ceiling. The founder ruled the floor stands (2026-08-24): recovery is never total, so the floor is the number that holds.
The ceiling (1.464x). It is a spec-internal note, never a public claim.
“Context growth past 100k was my biggest avoidable line by a lot, more than duplicate reads and oversized tool dumps put together.”
2026-08-24: context growth past 100k $2,795.42 (list-rate value consumed, not cash paid); duplicate file reads $116.58; oversized tool dumps $8.85; cache expiry $1,321.55.
The ledger notes only that the comparison holds by a factor of about 22 against the two named lines combined; no further claim is checked.
“Carrying them was more than 200x the one-time cost.”
2026-08-24: oversized tool dumps entry cost $8.85 (list-rate value consumed, not cash paid); carried dumps as cache reads $1,980.85. Ratio 223.8x, floored to 200x for publication.
224x. The floor is the claim.
“One project had the same file read 241 times.”
2026-08-24: dashboard.py, 241 reads, $5.30 (list-rate value consumed, not cash paid).
The project name, unless naming it serves the point. Name the count, not the project.
Withdrawn. This entry used to carry a measured claim about agent citations that traced back to nothing. Our own re-measurement could not stand behind the numbers as stated, so the claim came down before anything replaced it. It returns when a number survives its own audit.
Held back. The ledger's own rule for this number is “Available on request, not yet shaped into a claim.” Ask us and we show it.
“Around 114 adversarial cases, the rebuild agreed on 104, and of the ten divergences three were defects in the spec itself.”
The second foreman pass, commit 66d53ae (2026-08-25): independent reimplementation reproduced the gate across about 114 adversarial cases, agreed on 104; the ten divergences included three defects that were the spec author's.
Repo commit record, 66d53ae. Founder ruling 2026-08-25 on publishing repo-record numbers in social threads: “i'm not worried about publishing numbers in reddit, I know nothing we are looking at or using is outlandish.”
That the three spec defects were caught by a reader. The point of the number is that only the rebuild found them.
“Coming back used to cost 760K tokens. With Peddra, about 1K.”
One chat, resumed once, measured across the ten most expensive idle chats; the cold resume against the Peddra restart prompt. Banked into the site's v2.x master text before this ledger existed; carried forward 2026-08-25 as the homepage's sanctioned teach figure.
Site master text of record; the measurement predates this ledger and its worksheet should be re-located and linked here on the next measurement pass.
That every resume costs 760K; the figure is the measured cold-resume case across the ten most expensive idle chats, stated as “used to cost” for that case.
A 1,130,000-word chat, read end to end by Apple Intelligence. Chopped into 892 pieces. Tokens spent: 0.
Site master text of record; the measurement predates this ledger and its worksheet should be re-located and linked here on the next measurement pass.
A per-chat guarantee. It is one measured chat.
“Three people build Peddra.” Product of Maltby Ventures LLC, made in New York.
Founder-attested 2026-08-25: “The truth is we are 3 people.”
The founder's words in session, banked same day.
An org larger than three; never imply solo either.
Every number is a floor. When a measurement says 223.8x, the page says 200x. When the share measures 31.68 percent, the claim is “about a third,” with a tripwire that kills it if a run drops below 30. You discover more than we promised, never less.
Every token is counted once. The ledger splits what was spent into needed and avoidable: every token lands in exactly one pile, never both, never twice. A saving that can't be traced cleanly isn't shown.
Free means measured free. The app only calls a finish free when your Mac verifiably did every bit of the work. A cost it can't measure is never shown as zero.
The record never gates. The statement, the full account of a finished chat, stays readable whether or not you ever pay.
And you hold the instrument. The meter is free forever and reads the same disk these numbers came from. Run it over your own transcripts and compare what it finds to every claim on this page before a cent changes hands. If your numbers disagree with ours, believe yours, and tell us.
The unit law. Dollar figures in this ledger are list-rate value consumed, not cash paid. For subscription readers they are never the public unit: share, tokens, turns, and stretch are. Dollars appear in public only where real money changed hands, and are always labeled counterfactual otherwise.
The borrowing law. Every entry above is from measuring the founders' own Claude Code corpus. Nothing in this file came from another person's post, thread, or article. If a third party's number ever earns a place here, it gets its own entry with the source named in the claim itself, and it never gets spoken in first person.