
When an analyst drops an untrusted CSV into an AI coding assistant to generate a chart, three failure modes routinely occur: 1. Silent Hardcoded Mock Arrays (P1_NO_HARDCODED_MOCK): Dirty currency formatting ("$12,400", "NT$380,000", "N/A") causes naive parsers to choke and silently return hardcoded literal arrays so the chart still renders. 2. Subtotal Double-Counting & Mixed-Currency Summing (P2_SUBTOTAL_AND_CURRENCY_CONSERVATION): Summary rows (Total, Subtotal, Grand Total) are summed alongside leaf categories (inflating totals by 2x, e.g., $840,000 instead of $420,000), or TWD integers are added directly to USD amounts. 3. Untrusted CSV Prompt Injection & Sandbox Escape (P3_BLAST_RADIUS_ZERO): Malicious CSV notes columns attempt to read process.env.VULTR_API_KEY, exfiltrate secrets over HTTPS, overwrite host files, or trigger infinite CPU/memory loops. ChartTruth solves all three on Vultr Cloud Compute and Vultr Serverless Inference (https://api.vultrinference.com/v1): - Stage 1 (nemotron-3.5-content-safety): Pre-scans CSV headers, leaf cells, and summary rows for prompt-injection and exfiltration vectors. - Stage 2 (glm-5.3-flash): Synthesizes the candidate data-transformation script in the control plane while keeping VULTR_API_KEY strictly out of the execution container. - Stage 3 (Zero-Secret Ephemeral Sandbox + k×3 Shadow-CSV Probe): Executes the script inside an isolated container (--rm --network=none --memory=256m --cpus=0.5 --read-only --pids-limit=64 --cap-drop=ALL) twice—once on the original CSV and once on a k=3 scaled shadow CSV. - Stage 4 (Independent Accounting Oracle + clef): Verifies P1 (every plotted bar scales by exactly 3x), P2 (leaf-only USD conservation), and P3 (0 Outbound Bytes, 0 Secrets Exposed, 0 B Host Disk Delta), while clef computes calibrated decision probabilities. - Stage 5 (Bounded 1-Pass Self-Heal): Feeds the exact Oracle counterexample back to glm-5.3-flash to repair and re-verify the chart in a fresh Container #2.
8 Nov 2026