Sarvam Code / OpenAI Codex technical provenance analysis

This repository contains a reproducible static analysis of Sarvam Code v0.38.0 and the Apache-2.0-licensed OpenAI Codex Rust codebase.

Read the full technical report. For a shorter explanation without reverse-engineering terminology, see Sarvam Code provenance explained.

Main result

The analyzed Sarvam Code binary is derived from OpenAI Codex with very high technical confidence. This conclusion rests on 996 exact upstream source literals, 48 systematic branding substitutions, matching command and architecture surfaces, retained OpenAI references, and targeted implementation context in the MCP OAuth subsystem.

Derivation is not the same as theft. Apache License 2.0 permits modification and redistribution subject to conditions. The report does not conclude that Sarvam infringed copyright or violated the license.

Repository map

• REPORT.md — findings, methodology, caveats, and references
• evidence/source-literal-matches.tsv — all exact and branding-substitution matches
• evidence/version-fingerprint-scores.tsv — comparison across 37 Codex release tags
• evidence/cli-help-normalized.diff — normalized CLI comparison
• evidence/targeted-control-flow.txt — targeted Ghidra export
• scripts/source_fingerprint.py — reproducible scanner
• evidence/PUBLIC-SHA256SUMS — checksums for the published exhibits

Identified target

• SHA-256: 6a23109a27d702e6ef5245de157150174445d1f407896cee44b6831dadf55c68
• ELF build ID: 932f528bbba83ac425182c07b45871f3
• Reported release channel version: 0.38.0

The lean target used for reproducible scans has SHA-256 07110e0631cfc55cb579e9e3cd099f2b87ac634b48c60cfa4596ccf732d78092.

Upstream baseline

• Repository: https://github.com/openai/codex
• Commit: b545c94041017d000e2c8b2f6272705d21b85dfb
• License: Apache License 2.0

Evidence standard

The report distinguishes:

1. exact byte and string correspondence;
2. near-verbatim correspondence with defined branding substitutions;
3. structural and targeted implementation correspondence;
4. common third-party dependency behavior;
5. unsupported inference.

Ghidra output is reconstructed C-like pseudocode from a stripped, optimized Rust executable. It is not the original Rust source.

Why binaries are not committed

This repository does not redistribute the Sarvam executable, OpenAI release binary, Ghidra distribution, or multi-gigabyte analysis databases. Their cryptographic hashes identify the exact inputs. Obtain software from its publisher, verify the hash, and run the published scanner locally.