What each public proof does and does not prove.
The old partial 10-row Codex API artifact contains a mixed-arm raw meter delta, not verified CostLayers-attributed savings. Provider-native caching, CostLayers provider-call avoidance, local context estimates, and usage stretch are separate evidence streams.
Public proof map
Read these as four separate lanes. A number from one lane should not be promoted into a claim for another lane.
Corrected boundary for the old 10-row artifact
The legacy artifact aggregated meter differences from plain, CodeGraph, codebase-memory, and CostLayers arms. Its field named verified_invoice_saved_cost_usd verifies neither mechanism attribution nor CostLayers causation, so none of that aggregate is counted here as verified CostLayers-attributed savings.
old_partial_artifact: legacy_verified_field_accepted=false; verified_costlayers_attributed_saved_cost_usd=$0.000000; provider_avoided=0; best-in-world=false
What it proves: 10 completed API-mode rows and 10/10 recorded quality gates. What it does not prove: which dollars came from provider-native caching, any CostLayers-attributed savings, no quality regression across balanced arms, full live-matrix completion, customer invoice savings, subscription savings, or rank.
Repeated-cache provider-call avoidance proof
This is a separate direct gateway smoke artifact for repeated requests. It proves CostLayers response-cache provider-call avoidance can happen; it is not part of the old mixed-arm meter delta.
separate_repeat_cache_smoke: verified_costlayers_attributed_saved_cost_usd=$0.000000; aggregate_meter_excluded=true; provider_avoided=11
What it proves: repeated gateway response-cache avoidance on repeated API requests. What it does not prove: receipt-reconciled CostLayers-attributed dollars, attribution for the old 10-row artifact, the full autonomous benchmark, or customer invoice savings.
Local context matrix
The public local matrix is command/context evidence. It uses redacted receipts and quality gates, but it does not attach live API invoice counters to those rows.
What it proves: local command/context behavior across public repos with quality-gate receipts. What it does not prove: provider-dollar savings, autonomous code-edit success, full live task coverage, or customer invoice reduction.
Legacy 10-row quarantine descriptor
This is a historical summary, not a current proof receipt. The registry hash-binds the unchanged raw JSON and JSONL and prohibits promotion.
classification: legacy mixed-arm raw artifact
current_verified_proof: false
promotion_status: prohibited
rows: 10 completed; stopped for spend control
recorded quality labels: 10/10 (historical only)
provider calls: 695
legacy mixed-arm meter field: excluded from verified savings
verified CostLayers-attributed savings counted: $0.000000
provider-native savings portion: not isolated
provider_avoided: 0
best-in-world: false
mechanism: not attributable from this legacy artifact
raw file SHA-256: d77731d4adcd2a3afa2e9ddd1481a447343a1a2a4baa51e6c7df14a66c885dea
internal legacy certificate: ba7314f027889d119ba9ffff71a25773918765fa6d5761c6f3e38f1d7cf47d57
claim boundary: raw mixed-arm meter evidence; not CostLayers savings proofReproducible demo
No separate public sample repository is attached in this tree. The reproducible path checks current labels plus the hash-bound quarantine registry without promoting legacy raw files.
# From the repo root
node benchmarks/world-best/claim-consistency-smoke.js
node benchmarks/world-best/quality-safe-proof-fixture-smoke.js
node -e 'const r=require("./proof_artifacts/legacy-artifact-registry.json"); console.log(r.registry_status, r.policy.promotion_status, r.artifacts.length)'Expected output labels include claim consistency smoke passed, quality-safe proof fixture smoke passed, active_quarantine, and prohibited. The consistency smoke recomputes every registered raw-file and screenshot hash.
Missing-proof checklist
Real customer invoice export: missing. This is external proof and is not possible without customers supplying invoice data.
Full live task matrix: missing. The public API proof is 10 rows, not the full live matrix.
Third-party or raw-enough review: missing. Current public artifacts are redacted and hash-based; an independent reviewer still needs enough raw evidence under NDA or another controlled process.
What we show publicly
We do not publish raw prompts, source files, provider responses, private dashboard URLs, or detailed internal receipts on the public website.
How to verify your own savings
0.8.61 is deployed at the immutable hosted URL but is not published to npm. Run the local audit in one repo, run your own quality comparison, and use API mode only for a controlled workflow where you want provider-dollar evidence.
npx -y https://costlayers.com/costlayers-0.8.61.tgz audit --monthly-spend 1000
npx -y https://costlayers.com/costlayers-0.8.61.tgz quality compare -- npm test
Hosted package checksum: costlayers-0.8.61.sha256. It verifies downloaded bytes, not source ownership or a signature.
Start with the install command