#ReTrace · 可复算的 AI 数据报告

0.2.0 is a working prototype for the September 2026 MoonBit Hackathon. All included datasets are synthetic. It does not certify data authenticity, causal conclusions or engineering compliance.

#Package

#Try it

git clone https://github.com/FidollarinLA/retrace.git
cd retrace
npm run build
npm start
# Open http://127.0.0.1:4173

moon run cmd/retrace

#What you can do

• Read quoted CSV with Unicode, BOM, CRLF, embedded commas/newlines and escaped quotes. Reject duplicate headers and ragged records.
• Execute a strict version-1 analysis plan: AND filters, sequential derived columns, grouping, sum / mean / median / min / max / sample standard deviation / valid numeric count.
• Trace results to one-based source data-record IDs, included values, explicitly excluded missing values and filtered rows.
• Open the original records behind a chart, with included / missing / filtered states, derived cells, pagination and original CSV download. Parsing and record classification run in MoonBit.
• Turn a natural-language question into a proposed Plan v1 using your chosen model. Inspect the metric, filters, missing policy, units and trial execution before explicitly applying it.
• Inspect three scenarios using the same engine: regional sales, laboratory measurements and motor temperature rise.
• Export standalone HTML or a JSON bundle containing the CSV, plan, engine version, calculated results and a SHA-256 input fingerprint.
• Import a bundle and recompute; flag changed inputs, changed saved results and unsupported versions.
• Recompute existing 0.1.1 bundles under 0.2.0 and export an upgraded bundle. Every saved result field is checked, with only the declared engine version normalized.
• Check AI claims for exact numeric equality, evidence reference and unit. The browser includes a correct fixture and deliberately corrupted fixture, both clearly labeled as tests.

#Command line

node scripts/cli.mjs analyze examples/sales.csv examples/sales.plan.json report.json
node scripts/cli.mjs replay report.json
node scripts/cli.mjs html report.json report.html
node scripts/cli.mjs prompt report.json
node scripts/cli.mjs verify report.json draft.json

#Optional AI integration

node scripts/plan.mjs examples/sales.csv '按区域求销售额合计，单位元' proposal.json
# Review proposal.json before running analysis with it.

export RETRACE_AI_URL='http://localhost:11434/v1/chat/completions'
export RETRACE_AI_MODEL='your-installed-model'
# RETRACE_AI_KEY is optional; set it only for a provider requiring authentication.
node scripts/ai.mjs report.json draft.json

#Verification

moon check --deny-warn
moon test --target js --deny-warn
moon test --target wasm-gc --deny-warn
npm run build
npm test
npm run reference
moon info && moon fmt

#Architecture

CSV + Plan v1
    │
    ▼
MoonBit engine: parse → validate → derive → filter → group → aggregate
    │                                                    │
    ├── result + source record IDs + exclusions ──────────┤
    ▼                                                    ▼
browser / CLI → HTML + replayable JSON bundle       AI claims
                       │                                │
                       └── replay with MoonBit ── exact-value verifier

• engine/: reusable MoonBit parser, plan validator/executor, numerical aggregates and claim checker.
• bridge/: the same engine exported as an ES module for browsers/Node.
• cmd/retrace/: an executable MoonBit example.
• dist/: authored interface plus generated engine.js; no chart calculations are reimplemented in UI code.
• scripts/: build, CLI, optional AI adapter, local server and independent validation.
• docs/: schema, limitations, competition proposal and roadmap.

#Scope and limitations

• CSV only; no arbitrary Python/JavaScript execution, SQL, spreadsheet formulas, XLSX or PDF extraction.
• Units are explicit labels; dimensional analysis is not implemented.
• Missing numeric values are errors unless missing: "skip" is explicit. Non-numeric cells, NaN/Infinity and arithmetic overflow are always errors on evaluated numeric paths.
• count counts valid numeric observations, not all records. stddev uses denominator n−1 and returns null below two observations. Empty groups return null, never a fabricated zero.
• Claims require the full computed IEEE-754 value. Rounded natural-language statements need to be converted to structured exact claims; arbitrary prose is not verified.
• A checksum proves consistency with the stored input, not its authenticity. A bundle's hash can be recomputed by anyone who modifies it.
• CSV computation does not upload source records or use analytics. Explicit model requests share the displayed prompt context with the chosen provider. Exported bundles contain the raw CSV: share them intentionally.

#Originality and license