# Live replication protocol

The published result is one five-request sequence per configuration. Offline verification reproduces the recorded arithmetic exactly; fresh live results can differ because provider cache state, model aliases and account context change.

## Pinned systems

| System | Version / runtime |
|---|---|
| AGNT | 0.6.6; source SHA-256 identifiers in artifacts/environment/versions.json |
| Claude Code | 2.1.263 |
| Codex CLI | 0.153.4 |
| OMP | 18.1.14; Bun 1.3.14 |
| OpenClaw | 2026.9.2; Node 24.20.0 |
| Hermes | 0.21.1, release v2026.9.7; Python 3.13 |

Install pinned releases into a disposable directory, not over a working installation. The npm package-lock.json and Hermes uv.lock are included as dependency records. The lock files do not contain credentials. Native binaries and the operating system are not redistributed.

## Authentication

Use authorized subscription credentials. The drivers remove funded API-key environment variables from their child environments. Never print or package tokens.

- Claude Code: an authorized `CLAUDE_CODE_OAUTH_TOKEN` or native supported subscription login.
- OMP / Claude: `ANTHROPIC_OAUTH_TOKEN`.
- OpenClaw / Claude: `ANTHROPIC_OAUTH_TOKEN` on the measured native subscription path.
- Hermes / Claude: `ANTHROPIC_TOKEN`.
- Codex CLI / OpenClaw Codex: an isolated, authenticated `CODEX_HOME` prepared through the supported login flow.
- OMP / Codex: `OPENAI_CODEX_OAUTH_TOKEN`.
- Hermes / Codex: `CODEX_ACCESS_TOKEN`.
- AGNT: a separately provisioned isolated AGNT host with its own data directory and authorized subscription connections; supply `AGNT_AUTH_TOKEN` and `--base-url`. The driver does not start or modify AGNT.

Successful authentication alone does not establish permission for every client, product or deployment. Respect applicable provider terms and account limits.

## Commands

Run the driver with a NEW output directory. `--dry-run` validates the local workload and command construction without inference or authentication requests. CLI paths below are placeholders for your isolated installation.

    python artifacts/scripts/live_protocol.py --harness agnt --track claude --base-url http://127.0.0.1:PORT --output NEW_AGNT_CLAUDE
    python artifacts/scripts/live_protocol.py --harness agnt --track codex --base-url http://127.0.0.1:PORT --output NEW_AGNT_CODEX
    python artifacts/scripts/live_protocol.py --harness claude --track claude --exe PATH_TO_CLAUDE --output NEW_CLAUDE
    python artifacts/scripts/live_protocol.py --harness codex --track codex --exe PATH_TO_NODE --entry PATH_TO_CODEX_JS --output NEW_CODEX
    python artifacts/scripts/live_protocol.py --harness omp --track claude --exe PATH_TO_BUN --entry PATH_TO_OMP_CLI --output NEW_OMP
    python artifacts/scripts/live_protocol.py --harness omp --track claude --retention long --exe PATH_TO_BUN --entry PATH_TO_OMP_CLI --output NEW_OMP_LONG
    python artifacts/scripts/live_protocol.py --harness omp --track codex --exe PATH_TO_BUN --entry PATH_TO_OMP_CLI --output NEW_OMP_CODEX
    python artifacts/scripts/live_protocol.py --harness openclaw --track claude --exe PATH_TO_NODE --entry PATH_TO_OPENCLAW_MJS --output NEW_OPENCLAW
    python artifacts/scripts/live_protocol.py --harness openclaw --track claude --retention long --exe PATH_TO_NODE --entry PATH_TO_OPENCLAW_MJS --output NEW_OPENCLAW_LONG
    python artifacts/scripts/live_protocol.py --harness openclaw --track codex --exe PATH_TO_NODE --entry PATH_TO_OPENCLAW_MJS --output NEW_OPENCLAW_CODEX
    python artifacts/scripts/live_protocol.py --harness hermes --track claude --output NEW_HERMES
    python artifacts/scripts/live_protocol.py --harness hermes --track claude --retention long --output NEW_HERMES_LONG
    python artifacts/scripts/live_protocol.py --harness hermes --track codex --output NEW_HERMES_CODEX

For Hermes, execute with the Python environment containing the pinned Hermes installation. It calls the real AIAgent loop. The original experiment reconstructed that object from saved history at the idle phase; the portable driver retains the object through idle. This is a replication procedure, not a claim of byte-identical historical requests.

The driver sends four rapid turns, waits at least 390 seconds after request 4 completes, then sends request 5. It saves response counters and payload digests per request. It does not fabricate replies or retry failed sequences. An inference timeout may leave a provider request in flight: classify that entire sequence invalid rather than resending into a warmed cache.

## Configuration controls

The driver uses the same fixed-output workload files, explicit model, no intended tool execution, isolated CLI state, and documented retention settings. OMP disables tools/LSP/skills/rules/extensions/title generation. Claude Code disables tools, external MCP configuration and normal settings sources. OpenClaw denies tools and explicitly selects its embedded Codex runtime. Hermes disables tools, memory, context files and background review.

AGNT uses normal text-history request shape and its own orchestrator/adapter pipeline. Account-specific frozen prompt sections are not distributed; their omission means third-party runs need not match AGNT's absolute input denominator. The publication reports the actual measured input counts rather than asserting equal system prompts.

## Offline reconstruction

    python artifacts/scripts/verify.py
    python artifacts/scripts/build.py --output NEW_REBUILD_DIRECTORY

Then, from the new rebuild directory:

    pdflatex -interaction=nonstopmode -halt-on-error cache-wars-2.tex
    pdflatex -interaction=nonstopmode -halt-on-error cache-wars-2.tex

Required chart dependency: matplotlib 3.10.3. LaTeX uses article, geometry, amsmath, amssymb, booktabs, array, longtable, graphicx, times, hyperref, and caption. The SVG and figure PDF files are vector assets. HTML contains all thirteen SVG charts inline and has no external runtime dependencies.

`SHA256SUMS.txt` contains a checksum for every released file except itself. The externally delivered archive digest provides an additional reference. No signing key is created; checksums are not digital signatures. PDF bytes may differ across TeX installations or metadata generation even when content is unchanged; the numerical verifier and generated HTML/vector figures provide separate reproducibility checks.

The common-content Codex model is implemented in codex_analysis.py; verify_codex.py re-derives every projected request using exact fractions and Decimal prices. codex-scenario.json includes the per-request price tier. The live measurements are distinct from these analytical scenarios.