Prime Radiant receipts · All receipts

PR-5: How much text is reachable by address

Built 2026-10-05T02:02:19Z by the proof-harness worker (Claude, PROOF_HARNESS). Read-only on every store. Token figures marked "chars/4" are estimates, not a tokenizer; figures marked cl100k use tiktoken cl100k_base. Rebuild: python build_receipts.py.

The statement (each line with its condition)

(a) By address. Every one of the 286,538 unique files in the Prime Radiant (266,732,428,557 bytes, junction content excluded) is reachable by its address for a pointer of about 50 to 64 tokens, and 600 of 600 random fetches came back byte-exact (PR-2). Text in that store: 366,528,777 tokens counted exactly (cl100k, the prose text the search reads; 367 times a 1,000,000-token window), and about 9.1 billion tokens estimated across all text-bearing files (prose full length about 0.94 billion, other text and code about 8.18 billion; about 9,123 times a 1,000,000-token window). Safe line: "By address, an AI reached a store of about 9 billion tokens of the Founder's own text (at least 367 million counted exactly), and got the exact file 600 times out of 600 for a 50 to 64 token pointer."

(b) Without an address. Only the searched prose universe (114,396 items, 366,528,777 cl100k tokens) has a measured rate, and that rate is LOW: right file first 29 of 100 sealed hard questions (95 percent interval 21 to 39 percent), in the top five 45 of 100 (PR-3). No effective context without an address is claimed at a reliable standard. The honest line: "Without an address, across 367 million tokens, the fixed search found the right file first about 3 times in 10 on deliberately hard questions."

(c) Earlier receipt, for comparison. RECEIPT_BP223_HEX_CONNECTOR_5_32B_TOKENS_AT_50_TOKENS.md: 5,316,092,497 cl100k tokens in 660,682 files (a different store: pheromone shards, ledger eblets, Conduit, vault eblets, memory), 67 of 67 exact by address at 50 tokens; without address 3 of 67 first at full scale. The two stores overlap partly and are NOT to be added together.

How the count was made (recount from the stored run file)

Files and bytes by class: text 75,444 files, 22,088,181,960 bytes; prose_universe_A 102,016 files, 3,586,677,114 bytes; binary 36,205 files, 191,972,452,744 bytes; doc 170 files, 254,398,518 bytes; image 4,009 files, 2,262,012,543 bytes; code 68,310 files, 1,525,977,533 bytes; archive 214 files, 41,687,598,980 bytes; audio 152 files, 503,418,815 bytes; video 18 files, 2,851,710,350 bytes.

Live re-run

The count is a full pass over the store (126.6 s); it is the run named below, made 2026-10-05. Re-running it gives a larger number as the estate grows; the exact A1 figure is fixed by corpus.db (built 2026-10-04T23:19:29Z).

What this does and does not show

1. "Effective context" here means reach: a pointer in the AI's window resolves to the exact bytes. It is not a model input of billions of tokens.
2. The 9.1 billion figure is an ESTIMATE dominated by B (22 GB of text-class files, mostly session transcripts, json and logs, many of them near-copies of each other). De-duplication is by exact bytes only; it is NOT a measure of distinct knowledge. B may double count Founder turns that also appear in A.
3. B's ratio comes from a file-uniform sample read up to 2 MB per file; very large files may tokenize differently. The exact floor is A1.
4. One tokenizer (cl100k_base), not Claude's; another tokenizer gives a different count for the same bytes, not measured here.
5. The 1,000,000-token comparison uses a round figure; the earlier receipt used 1,050,000 for the largest flagship window.

In a 50-turn session: context stays flat (added 2026-10-06)

The sections above measure reach: a pointer resolves to the exact bytes. This section measures what that does for a working session. Fifty questions were sealed (sha256 95202442c30a5b5f95f5c7365ac4cb1f02cb00508b47bb2da0e8064bdbc4f1bb) before any run and checked before each run. The 50 mix lookups in a 34-file folder (493 KB), follow-ups that ask about an earlier answer, and new notes typed in the message. Turns 1 to 25 ran on Claude Sonnet 5.5 and turns 26 to 50 on Claude Haiku 4.5, so 6 follow-ups cross a model swap. Code scored every answer against the sealed key. Every answer had to carry a QUOTE line, which code then looked for word for word in the sources; when nothing was found, the AI was required to say "I don't know".

The released app (0.3.9 Ask). Ask answers only from files and saved cards, never from what was typed earlier in the conversation. So 15 of the 50 cannot be answered there by design: the 10 typed notes and the 5 follow-ups that ask for them back. On the 35 whose answer is in a file: 29 right, 0 wrong shown, 6 "I don't know". The AI gave 2 wrong answers; the quote check caught both and they were shown as "I don't know". It also turned away 1 right answer because its quote did not match the source byte for byte.

What this does and does not show

1. One run each, one 34-file folder, 50 sealed questions. It is not a measure across other folders or other kinds of work.
2. The plain session got all 50 right. Conduit traded 3 answers for a context that stays flat; in a plain session the context grows every turn until it hits the window.
3. Cost in dollars was not lower at 50 turns: the plain session cost 67% of what run 3 cost (0.71 / 1.06; notional costs from the Claude command line on a subscription login, not an API bill). The flat line is about context size, not price at this length.
4. The 47 of 50 runs used Conduit's chat path, with the session's earlier messages recalled by address. The released 0.3.9 Ask does not have that path yet; its own figure is the one in the last column.
5. The 3 misses in run 3: one figure rounded ("about $3,300" against 3,320.87); one time converted to US Central under a house rule while the key held the source time; one "I don't know" with the answer in the passage.

Evidence (sha256 computed at receipt build)

Filed. Not done. The Founder decides.