Agastya AGI v01 Native Brain Evidence Release R1

Rabbit Industries Pvt Ltd

PUBLIC PROVENANCE RELEASE • R1 • 2026-09-30

Agastya AGI v01 — Native Brain Evidence Release R1

Rabbit Industries' externally citable provenance package for the Agastya from-scratch native foundation-model lineage. This release publishes cryptographic identifiers and validation evidence without redistributing the proprietary raw checkpoint, corpus or full tokenizer artifact.

EXTERNALLY CITABLE EVIDENCE PACKAGERANDOM-INIT EVIDENCE: PASSTHIRD-PARTY WEIGHT ANCESTRY: FALSEINDEPENDENT EXTERNAL VERIFICATION: PENDING

Important: “Externally citable” means this evidence is now published at a stable public URL with machine-readable manifests and SHA-256 identifiers. It does not mean an independent third party has already verified every claim. Independent verification remains a separate next gate.

Canonical from-scratch lineage

Agastya corpus → Agastya tokenizer → Agastya architecture → RANDOM INITIALIZATION → Agastya pretraining → Agastya post-training → Agastya Native Brain

Evidence summary

Parameters1,264,715,7761.265B unique parameters

Tokenizer98,320 vocabBPE • NFC • ByteLevel

InitializationRandomValidation: PASS

Teacher ancestryZERONo teacher weight ancestry at step 0

Third-party weightsFalseRecorded in provenance + validation

Third-party tokenizer importFalseRecorded in step-0 provenance

Step-0 checkpoint2,529,519,595 bytesRaw checkpoint retained privately

Clean-P0 acceptancePASScompleted_zero_third_party: true

Cryptographic anchors

Tokenizer artifacte8695c48f54bebac4c5e3a2ea01f743a56ce7268f73832ef9324eec5facf7257MATCHED

Architecture source6ce4ae3dfe7af1bbc34ebc5f4a4cde0c63877837c59bc9285c2776a9046fe2a4MATCHED

Initialization source54fc83ff58f40f68e3ab7db3518e3d012c0aa6dbb6dedc6fffec8a95918e8c72PUBLISHED HASH

Step-0 checkpointffd77d9c9d0f544c7ca8ca508ac8c75d837cecf985de7c422769a4dcfbcd9cf1PROVENANCE MANIFEST

Validation manifestda2f0bdfc01bbca6f2e39a2cf5f1d8394d74ab85caecd5b2a00d900e4a6b4ab5PASS

Final acceptance manifest7d37806ecb85153c70621c5603fa2f94accc0698713cce3aa6861c3e9fac4ec8PASS

Validation claims backed by the published manifests

Check	Result	Evidence source
Fresh random initialization	TRUE	Checkpoint validation R1
Third-party weight ancestry	FALSE	Checkpoint provenance + validation
Teacher weight ancestry	ZERO	Checkpoint provenance + validation
Third-party tokenizer artifact import	FALSE	Checkpoint provenance R1
Unique parameter count exact	TRUE	Checkpoint validation R1
Clean-P0 completed zero-third-party gate	TRUE	Final acceptance R1
Independent external replay	PENDING	Next external verification gate

Download / cite the evidence

Evidence manifest R1 (JSON) Native weight lineage R1 (JSON) Tokenizer provenance R1 (JSON) Checkpoint validation R1 (JSON) Clean-P0 acceptance R1 (JSON) Source hashes R1 (JSON) Initialization evidence excerpt Package checksums Citation metadata (CFF) Evidence README Complete R1 evidence package (ZIP) Public benchmark profile

Evidence boundaries

This R1 release deliberately does not publish the proprietary raw training corpus, full tokenizer vocabulary, architecture source code, or the 2.53 GB checkpoint. Those artifacts remain privately retained and are identified through cryptographic hashes. A qualified independent evaluator can be provided controlled access for hash and lineage verification under an appropriate review agreement.

Next gate: independent evaluator replay → signed external verification report → DOI/preprint/model-card publication → same-suite benchmark acceptance.