Agastya AGI v01 — Native Brain Technical Evidence Report R1

Abstract

Rabbit Industries Pvt Ltd publishes this R1 evidence report to make the Agastya AGI v01 from-scratch native-model lineage externally citable. The release exposes cryptographic artifact identifiers and validation/acceptance records while retaining proprietary raw weights, corpus and source code.

Core claim under test

Agastya's canonical Native Brain lineage is defined as:

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

The R1 evidence manifests record a fresh step-0 checkpoint, zero teacher-weight ancestry, false third-party weight ancestry, false third-party tokenizer artifact import, and a PASS checkpoint validation.

Architecture and scale

• Unique parameters: 1,264,715,776
• Dense architecture
• Canonical context: 4,096 tokens
• Tokenizer vocabulary: 98,320
• Tokenizer model: BPE
• Normalization: NFC
• Pre-tokenization: ByteLevel

Step-0 provenance

The published weight-lineage manifest exposes:

• architecture source hash
• architecture hash
• tokenizer hash
• deterministic random seed
• seed-derivation hash
• record-selection hash
• token-cache hash
• step-0 checkpoint hash and byte length

The public package does not redistribute the raw 2.53 GB checkpoint.

Validation

The checkpoint-validation manifest records PASS for random initialization, zero teacher ancestry, false third-party weight ancestry, tokenizer identity and exact unique parameter count.

The Clean-P0 R2 final acceptance manifest records PASS and completed_zero_third_party=true.

Evidence scope

This R1 report is company-generated evidence. It is externally citable because the package is publicly accessible with stable URLs, checksums, machine-readable manifests and citation metadata.

Independent external verification remains pending and must be completed before describing Agastya as independently verified.

Benchmark scope

Comparable frontier benchmark scores are not asserted until an identical frozen evaluation harness is run and independently replayed. Internal diagnostics are not substituted for public same-suite results.

Next publication gates

1. Independent evaluator replay.
2. Signed verification report.
3. DOI-backed archival deposit.
4. Public model card mirror.
5. Research preprint.
6. Same-suite benchmark submission / external leaderboard evaluation where applicable.