abdelstark commited on
Commit
5309557
·
verified ·
1 Parent(s): 15c8c4a

Document Circle STARK as the default proof workflow

Browse files
Files changed (2) hide show
  1. README.md +21 -13
  2. manifest.json +2 -2
README.md CHANGED
@@ -137,19 +137,27 @@ recorded examples is not a global float/integer equivalence theorem.
137
  ## Proof integration
138
 
139
  These exact weights appear in Hallmark's 96-byte, three-classifier demo.
140
- `make e2e` trains/reproduces the frozen exports, creates fresh Groth16/BLS12-381
141
- setups and nine proofs, and checks them using separate py-ecc arithmetic.
142
- Model download/inference is independent of that proof workflow.
143
- The 292-byte proof encoding excludes keys, enrollment, statements and model data.
144
-
145
- A valid proof establishes the enrolled computation, not the usefulness of its
146
- policy. The caller must independently enroll the model/program/circuit/key,
147
- reconstruct the expected request and bind the public inputs. Current research
148
- setup is not an independently reviewed ceremony or a production qualification.
149
-
150
- [Exact classifier statement](https://github.com/AbdelStark/hallmark/blob/9b74e1c6b6e90de2d226556aa637936a61a0757b/docs/PROOFS.md) ·
151
- [Proof composition evidence](https://github.com/AbdelStark/hallmark/blob/9b74e1c6b6e90de2d226556aa637936a61a0757b/research/evidence/w28-trained-common-proofs-2026-10-22.json) ·
152
- [End-to-end demo](https://github.com/AbdelStark/hallmark/blob/9b74e1c6b6e90de2d226556aa637936a61a0757b/docs/DEMO.md)
 
 
 
 
 
 
 
 
153
 
154
  ## Licence, provenance and contact
155
 
 
137
  ## Proof integration
138
 
139
  These exact weights appear in Hallmark's 96-byte, three-classifier demo.
140
+ `make e2e` now defaults to **STWO Circle STARK**: it trains/reproduces the frozen
141
+ integer exports, pins public preprocessing and generates nine proofs with fresh
142
+ proof-only verification. No Groth16 setup or proofs run first. Model download
143
+ and native inference are independent of the proof workflow.
144
+
145
+ The recorded Circle proofs are 1.91–2.16 MB canonical JSON, excluding enrollment,
146
+ statements and model data. `make e2e BACKEND=gnark` preserves the explicit
147
+ Groth16/BLS12-381 comparison, with separate py-ecc verification and 292-byte
148
+ binary proofs. Encodings, memory and security profiles differ; proof sizes alone
149
+ do not rank proof systems.
150
+
151
+ A valid proof establishes the enrolled computation, not policy usefulness.
152
+ The caller independently enrolls the model/program/preprocessing, reconstructs
153
+ the expected request and binds public inputs. No hiding, hardware provenance
154
+ or independent security qualification is claimed. Weights and thresholds are
155
+ unchanged by this documentation update; original release revisions remain valid.
156
+
157
+ [Current classifier statement](https://github.com/AbdelStark/hallmark/blob/main/docs/PROOFS.md) ·
158
+ [Circle classifier guide](https://github.com/AbdelStark/hallmark/blob/main/docs/CIRCLE-CLASSIFIERS.md) ·
159
+ [Original Groth16 evidence](https://github.com/AbdelStark/hallmark/blob/9b74e1c6b6e90de2d226556aa637936a61a0757b/research/evidence/w28-trained-common-proofs-2026-10-22.json) ·
160
+ [Current end-to-end demo](https://github.com/AbdelStark/hallmark/blob/main/docs/DEMO.md)
161
 
162
  ## Licence, provenance and contact
163
 
manifest.json CHANGED
@@ -15,8 +15,8 @@
15
  "sha256": "b91065cc7a54ab9ec16b955a99de915e744e5086b7aad7010070c840c7f69392"
16
  },
17
  "README.md": {
18
- "bytes": 9095,
19
- "sha256": "4b2b7ee6186c192b2a17701a90102a1b57416306f4df2580a1fa7852d89699fb"
20
  },
21
  "arithmetic.cbor": {
22
  "bytes": 1059,
 
15
  "sha256": "b91065cc7a54ab9ec16b955a99de915e744e5086b7aad7010070c840c7f69392"
16
  },
17
  "README.md": {
18
+ "bytes": 9560,
19
+ "sha256": "2b9b71a2e359ccf64c4c57c5fef9619c54e2255fb559f92b5057969b35e46ac0"
20
  },
21
  "arithmetic.cbor": {
22
  "bytes": 1059,