# Build Token Value

> Tarski-honest metamethodology for estimating AI build tokens and three value layers
> (L1 cash / L2 API-list / L3 human). Multi-op fan-out across Claude Code, Codex, Grok Build.
> Product: Pharosiraptor.

## Site

- Dashboard: /
- Certificate seed: /data/estimate.json
- Methodology: /data/methodology.json · /methodology/SKILL.md
- Agents: /data/agents.json
- Prompts: /prompts/00-multi-op-compose.md … /prompts/10-gemini-cli.md
- Scripts: /scripts/*.py
- Bare git (dumb-HTTP): /git/build-token-value.git

## Clone

```bash
git clone https://build-token-value.pages.dev/git/build-token-value.git
```

## Billing (L1 cash)

- Template: /data/billing.config.json · repo config/billing.example.json
- Local only: config/billing.local.json (gitignored) + CSV env vars BTV_*_BILLING_CSV
- Collect: python scripts/collect_billing.py → fold via scripts/pipeline.py
- L1 never inferred from L2 tokens

## Quick run

```bash
py -3 scripts/pipeline.py --since 2025-11-01
py -3 scripts/verify_deploy.py
```

## Deep links

- Multi-op prompt: /#prompt/multi-op
- Per agent: /#prompt/claude-code · codex · grok-build · cursor · copilot · windsurf · aider · cline · devin · gemini-cli

## Invariants

1. Multi-op fan-out when ≥2 backends ready (OTC multiset — not exclusive cascade).
2. Claude stats-cache is primary measured source.
3. Grok/Codex session bytes are semi-decidable proxies.
4. Never collapse L1 / L2 / L3 / L4 market value.
5. Name Tarski residuals explicitly.
6. Public estimate.json is sanitized (no local filesystem paths).