Agastya Native Brain Independent Verification Protocol R1

Purpose: allow an independent evaluator to determine whether the supplied Agastya checkpoint satisfies the published from-scratch native-lineage claims without requiring public release of proprietary weights or corpus.

Inputs to evaluator

1. Frozen step-0 checkpoint or controlled hash-verification access.
2. ANTR2-R1 tokenizer artifact.
3. Frozen architecture source/configuration.
4. Initialization source and provenance manifest.
5. Clean-P0 training/acceptance manifests.
6. Selected corpus-manifest hashes and training-token accounting evidence.
7. Frozen post-training checkpoint selected for benchmark evaluation.

Verification gates

V1 — Artifact identity

• Recompute SHA-256 of tokenizer, architecture source, step-0 checkpoint and supplied manifests.
• Match against Public Evidence R1.

V2 — Random initialization

• Inspect step-0 checkpoint and initialization source.
• Verify deterministic seed derivation.
• Confirm optimizer state is absent at step 0.
• Confirm production training is false at step 0.

V3 — Weight ancestry

• Verify provenance declaration: third_party_weight_ancestry=false.
• Verify teacher_weight_ancestry=ZERO.
• Inspect state initialization path for absence of third-party checkpoint loading before step 0.

V4 — Tokenizer ancestry

• Verify tokenizer SHA-256 and tokenizer structure.
• Review tokenizer creation records sufficient to determine whether vocabulary/merges were trained as an Agastya artifact rather than imported from another model.
• Record any unresolved provenance gaps explicitly.

V5 — Parameter count and architecture

• Independently count unique model parameters.
• Verify architecture configuration and tied-weight treatment.
• Compare against published 1,264,715,776 unique parameters.

V6 — Training lineage continuity

• Verify checkpoint chain from step 0 to selected trained checkpoint using immutable hashes/manifests.
• Verify no unapproved external weight insertion occurs in the chain.

V7 — Benchmark replay

• Freeze model version, tokenizer, prompts, scoring, tool permissions and environment.
• Run agreed public suites.
• Publish exact harness/version and confidence/variance where applicable.

Required evaluator output

A signed report containing:

• evaluator identity/organization
• date
• artifact hashes
• PASS/FAIL/INCONCLUSIVE for V1–V7
• methods
• deviations
• conflicts or unresolved questions
• benchmark results, if executed
• signed report hash

Public status rules

• Do not use “independently verified” until V1–V6 are PASS.
• Do not publish an external performance rank until V7 uses directly comparable frozen suites.