Prime Radiant receipts · All receipts
PR-2: Exact recall by address
In plain language: With the address, 600 of 600 random files came back byte-exact, for a pointer of about 50 to 64 tokens and a sub-millisecond lookup.
The full receipt follows, as written: method, every limit and incident, and the SHA-256 of every input and output file. File locations are shown by file name only.
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.
What was proven, in plain language
Given a file's address (its sha256 or the 12-character short form), the Prime Radiant returned that exact file 600 times out of 600 over two random draws of 300 from 295,636 unique files, re-hashed from disk; the short address resolved to exactly one file every time. The pointer an AI holds costs a median of about 48 to 64 tokens (chars/4; max 88 in these draws), and the lookup takes under a millisecond at the median. Founder's Plan items by id: 200 of 200 turns and 96 of 96 topic sections exact.
The line that is safe to publish: with the address, 600 of 600 random files came back byte-exact, for a pointer of about 50 to 64 tokens and a sub-millisecond lookup.
Recount, from the stored run files
| run | seed | byte-exact sha256 | short address unique | failures | pointer tokens median per subtree bucket (chars/4) | lookup median | Founder's Plan |
|---|---|---|---|---|---|---|---|
| first 2026-10-04T23:00:38Z | 20261004 | 300 of 300 | 300 of 300 | 0 | small_subtrees: 49.0 (max 66), medium_subtrees: 49.5 (max 69), large_subtrees: 64.0 (max 80) | small_subtrees: 0.2911 ms, medium_subtrees: 0.3426 ms, large_subtrees: 0.5266 ms | turns 100 of 100, sections 48 of 48 |
| live re-run 2026-10-05T01:35:27Z | 20261005 | 300 of 300 | 300 of 300 | 0 | small_subtrees: 48.0 (max 73), medium_subtrees: 50.0 (max 71), large_subtrees: 64.0 (max 88) | small_subtrees: 0.348 ms, medium_subtrees: 0.3531 ms, large_subtrees: 0.467 ms | turns 100 of 100, sections 48 of 48 |
The claim 2 script stores per-bucket aggregates and the failure list, not one row per fetch; the independent check is the live re-run with a new seed.
Live re-run for this receipt
Re-run 2026-10-05 with seed 20261005 (new random draw of 300 files plus 100 turns and 48 sections): the second row.
What this does and does not show
- It shows reach by address: the AI holds a pointer of about 50 to 64 tokens and the pointer resolves to the exact bytes. It is not a bigger model window. Say "reached", never "context window of".
- Token figures are chars/4 estimates of the pointer text (path length varies), not a tokenizer count; a conservative hex-weighted estimate is 64 to 80 at the median (max 112).
- File read plus re-hash time (about 5 to 7 ms median) depends on file size and OS cache and is not part of the pointer cost.
- It says nothing about finding a file WITHOUT its address (PR-3).
Evidence (sha256 computed at receipt build)
| path | bytes | sha256 | what it shows |
|---|---|---|---|
| claim2_20261004T230038Z.json | 2987 | b3ed987703a79659b3166bad4da073aa9e22fe49740066d22ac7af4f639ed1dc | first run |
| claim2_20261005T013527Z.json | 2988 | 44dbb71b8e88c3ef09b1f8894c11a67b9b10427aa3e024f24e87b2dd503418e2 | live re-run, new seed |
| common.py | 2163 | 127eb220e276f5c48df501247ba717b8e2003acd56d3f784025326362c7ce596 | the code that produced or checks these numbers |
| claim2_address_recall.py | 8749 | 2641ecc3e65573b39f013b3362d755d84afad3aa288c9e8ef2ee540ee5335ae2 | the code that produced or checks these numbers |
Filed. Not done. The Founder decides.