RECEIPT: Hex Connectors reached a 5.32-billion-token store at 50 tokens per answer

Built 2026-09-17T23:13:10Z by an AI session (Claude Fable 5.1). Read-only on every store. 0 model calls. $0. Rebuild any time with the receipt’s own build script (available to reviewers on request).

What was proven, in plain language

An AI given a record’s Hex Connector address got the exact record back, 67 times out of 67, from a store of 5,316,092,497 tokens in 660,682 files, for 50 tokens each. That store is about 5,062 times the largest flagship context window (1,050,000 tokens). The cost per answer was the same at 1,000 files as at the whole store. It was measured twice on the whole store, on 2026-09-16, and the 67 lookups were re-run live for this receipt.

“Exact” means the sha256 of the bytes returned equals the target sha256 frozen before the full runs. Nothing was judged by eye and no model was asked.

The line that is safe to publish:

The best AI holds about 1 million tokens at once. Through Hex Connectors, an AI reached a store of 5.32 billion tokens, about 5,000 times that, and got the exact record 67 times out of 67 for 50 tokens each, when it had the record’s address.

Recount, from the stored per-record rows (not from the summaries)

The one “NO” under “same 67 needles” is the 100,000-file validation re-run: by then the live memory index yielded a 68th needle (a record filed after the first measurement). It also came back exact (68 of 68). It is used in no other tier, so every other tier is the same 67, as the full-store run report says.

Full store, with the AI assistant’s own memory notes about each ruling removed from the keyword ranking (run 2 only): first 22 of 67, top five 43 of 67. The unmasked figure is in the table.

Store arithmetic (per-store new files and tokens must add up to the totals): run 1 OK, run 2 OK.

Addresses indexed in run 2: 661,620 (duplicate files are aliases of one content row). Run 2 ran 2026-09-16T22:59:14Z to 2026-09-16T23:23:49Z. Tokenizer: tiktoken cl100k_base 0.14.0.

Live re-run for this receipt

Ran 2026-09-17T23:13:10Z. For each of the 67 frozen needles the connector was computed from its address, the record was read, and its sha256 was compared with the sha256 frozen on 2026-09-16.

• Found: 67 of 67. Byte-exact against the frozen sha256: 67 of 67.
• Pointer tokens: at most 50. Full record: mean 324.9, max 897. Lookup plus read: 0.048 ms mean.
• Distinct connectors: 67 of 67. Model calls 0. Network calls 0.
• This live step re-proves the ADDRESSED LOOKUP. It does not re-count the 5.32 billion tokens; that count is the two full runs above (14 and 25 minutes of CPU). The stores grow every day, so a new count would be larger, not the same.

What this does and does not show

1. It shows reach, not a bigger model window. The AI holds a 50-token pointer in its context and the pointer resolves to the record. It is not a 5-billion-token model input. Say “reached”, never “context window of”.
2. It worked when the address was known. With no address, plain keyword search put the right record first for 49 of 67 at 1,000 files and for as few as 3 of 67 across the whole store (22 of 67 with the AI assistant’s own memory notes masked).
3. That caveat is the reason to make addresses, not a weakness in the claim. Every record saved through MnemosyneC gets its address and its provenance at the moment it is saved, while the person simply uses AI the way they already do. Each person maps their own part of the world’s data; what they choose to share becomes addressable for everyone.
4. Work since then on the no-address case, each with its own report: Shiver/Frenzy boot card 22 to 35 of 67 first (top ten 54); rule B (answer only when keyword and card agree, else abstain) 23 right, 4 wrong, 24 abstain on a fresh 51 (session log, DELTA 80).
5. No model answered anything. This measures delivery and retrieval, not reasoning.
6. The store is the Founder’s working estate on M0. A new installation starts with an empty personal store.
7. The largest flagship window is taken as 1,050,000 tokens, the figure used in the earlier measurements. If a vendor ships a larger window, the multiple changes and the token count does not.
8. The token count is of stored text, de-duplicated by exact bytes only. About 4.96 of the 5.32 billion tokens are in the vault eblet store, which holds eblet copies of records that also exist elsewhere. A copy with a different wrapper is a different file by sha256 and is counted. So 5.32 billion is the measured size of the text an address can reach; it is NOT a measure of how much distinct knowledge that is. Nobody has measured that. It does not affect the 67 of 67 or the 50 tokens.
9. Token counts use one tokenizer (cl100k_base). Another tokenizer would give a different count for the same bytes; that difference has not been measured here.

Evidence (sha256 computed at receipt build)

A machine-readable copy, with all 67 live rows (connector, sha256, tokens, milliseconds), is available to reviewers on request.