01
Direct serve-once
The API output is the result. A shuffle, a weighted choice, a UUID, a 16-byte pulse. Ask forsign: true and bind a nonce you already published. A third party checks the Ed25519 receipt offline.
§00Southernmost physical entropy
Use
True
Random.
Touch the unwritten physical world.
Wideband noise from proprietary solid-state avalanche sources, sampled in southern South America, conditioned with HMAC-SHA-256, and served once. The request ends at this API.
Reading
- Control
- ~298×
- Assessed
- 4.313228
- Credit
- 2 bit/sample
- Raw kept
- 0
4.313228 bit/sample is an SP 800-90B non-IID initial assessment (full floor 4.313227856926572). CP2 credit stays at 2 bit/sample. Neither figure is a certificate.
16BSandbox
Live avalanche pulse
Awaiting trigger
Nothing has been read. This control does not preview a sample.
Sandbox ceiling 64 bytes. This pulse draws 16 and asks for a signed receipt. A full key raises the ceiling.
§01Black-box architecture
From physical noise to a byte that leaves once.
Six stages, one direction. Select a stage for the guarantee it actually makes. The circuit itself is not on this page.
Functional flow. Hardware topology is unpublished. One direction: a failed gate never reaches the pool.
01Solid-state avalanche sources
Physical avalanche breakdown is the sole declared entropy source. The server never mixes vendor on-chip RNGs, operating-system CSPRNGs, or deterministic DRBGs into conditioned output.
- Sources
- Proprietary solid-state avalanche noise sources. Circuit topology, parts, and operating points stay unpublished.
- Control
- Every architecture is checked against a matched negative control: amplification and digitization stay powered while the avalanche stage is isolated.
- Margin
- In the reference campaign, isolating the avalanche stage collapsed variance by about 298× and left strong temporal memory.
- Credit
- A new hardware profile does not inherit a credit policy. It needs its own non-IID assessment first.
§02Zero-bias compute
A seed repeats. A remainder leans. A physical byte leaves.
A seeded generator is a function. Give it the same state and it walks the same path. That is what makes it useful inside a program, and what makes it the wrong object when the draw itself has to be unrehearsed.
rand % N fails a smaller test. If the generator's range is not a multiple of N, some residues come up once more than the others. The lean is quiet, and a lively seed does not remove it. Integer draws here use bitmask rejection. Bytes that fail the predicate are discarded. Putting them back would let a later caller condition on the rejection.
An Ed25519 receipt is an origin attestation over physical output. It is not a VRF. A VRF recomputes its output from a key and an input. This output is the noise.
02
Entropy-in-depth
Mix a physical sample into an operating-system CSPRNG with HKDF-SHA-256. SaltAS-HKDF-v1, info analogseed-csprng-reseed. If the pool answers 503, keep the operating-system generator and record that the physical sample was absent. The fallback is not avalanche.
03
MCP
analogseed-mcp gives Cursor, Claude Desktop, and other agents the same primitives: integers, shuffles, weighted choices, raw bytes, and the coarse status. There is no second entropy story for the model.
| Question | Seeded PRNG | rand % N | Direct serve-once | Entropy-in-depth |
|---|---|---|---|---|
| Repeats | Same seed, same stream. | Same generator, plus a lean. | A served byte does not return. | The mix is not the seed's stream. |
| Bias | None, used as specified. | Residues are not equiprobable when the modulus does not divide the range. | Bitmask rejection. Leftovers are discarded. | The mix does not repair a biased draw you already published. |
| Source down | Still answers. | Still answers. | 503 empty. No partial body. | Keep the OS CSPRNG and label the miss. |
| A third party | Recomputes, given the seed. | Recomputes, given the seed. | Checks the receipt. Does not recompute the noise. | The receipt covers the physical sample, not the mixed key. |
§03Metrology
The numbers are measurements, not a certificate.
~298×
Variance ratio, on / off
σ 249.21 counts with the avalanche stage engaged, against σ 14.44 counts with that stage isolated and the rest of the chain left powered. Flatness 0.9956 against 0.6428.
4.313
bit/sample assessed · credit 2
SP 800-90B non-IID initial assessment on the 8-LSB representation. The floor is 4.313227856926572 bit/sample. Operational CP2 credit stays at 2, a 2× margin into HMAC-SHA-256.
0
Production raw retained
Production blocks are conditioned in memory, then the raw samples are destroyed. Audit blocks are kept for metrology and are not served.