ChangeProof

Turn every external requirement into a verified internal change.

ChangeProof is a governed B2B change-management agent. It converts a new regulation, contract clause, security requirement, audit finding, or internal policy decision into a cited and reviewable change package spanning policies, controls, procedures, software, tests, owners, and verification evidence.

The initial project is being built for the OpenAI Build Week 2026 — Work & Productivity category.

Product promise

Most compliance tools tell a team what changed. ChangeProof completes the change—and proves it.

external obligation
    -> cited structured requirement
    -> organizational impact graph
    -> coordinated document and code changes
    -> human approval
    -> tests and residual-risk checks
    -> immutable evidence package

Initial demonstration

A fictional Indian lender receives an RBI circular reducing its complaint-response deadline from 30 days to 21 days. ChangeProof identifies every affected policy, SOP, control, service constant, scheduler, UI label, and test; produces coordinated redlines and patches; incorporates reviewer feedback; runs verification; and generates a source-to-change proof manifest.

Documentation

• Research and recommendation
• Product specification
• Architecture
• Build Week execution plan
• Delivery checklist
• Capital adequacy reference scenario
• Submission checklist
• Prior-work boundary
• Self-compliance obligations register
• Security policy and privacy boundary

Status

The deterministic offline vertical slice is implemented. It loads the synthetic lender workspace, computes cited artifact impacts and governed proposal history, applies only approved revisions to an isolated workspace, runs a real regression check and residual scan, and writes a hash-addressed proof record. A local reviewer UI binds approval to the exact displayed eight-artifact change plan.

Impact discovery is runtime analysis, not fixture lookup: ChangeProof scans the workspace for obligation-relevant deadline spans and structural references, records file/line/column evidence per edge, and constructs the impact graph. The checked-in impacts.json is used only by tests as the expected-result oracle. Offline edits target those evidenced offsets; unrelated occurrences are untouched.

Judge path — three commands

git clone https://github.com/SalilBhasinOfficial/changeproof.git
cd changeproof
PYTHONPATH=src python3 -m changeproof review --workspace /tmp/changeproof-review --database /tmp/changeproof-review.sqlite3

Open http://127.0.0.1:8000/. No package installation, API key, BBB service or private data is required. Inspect the highlighted obligation and computed graph, reject one impact with rationale, inspect the revised/superseded plan, then approve the current plan to reach the proof certificate.

Run the CLI golden path

Requires Python 3.11 or newer. The demo runtime uses only the standard library. The workspace must be absent or empty.

PYTHONPATH=src python3 -m changeproof demo \
  --workspace /tmp/changeproof-demo \
  --database /tmp/changeproof-demo.sqlite3 \
  --approve-as "$USER"

The command exits non-zero if the executable check fails or an obsolete 30 reference remains in the isolated workspace. --approve-as is the explicit human authorization recorded before applying changes to that workspace; it does not authorize production changes. Evidence is written to /tmp/changeproof-demo/proof_manifest.json.

To run the reviewer flow from an existing checkout:

PYTHONPATH=src python3 -m changeproof review \
  --workspace /tmp/changeproof-review \
  --database /tmp/changeproof-review.sqlite3

Open http://127.0.0.1:8000/, inspect the cited obligation and eight diffs, then enter a reviewer identity to authorize that exact plan. This applies only to the isolated workspace and redirects to the resulting proof certificate.

Each impact can instead be rejected with reviewer identity and rationale. The review agent produces a revised proposal and new snapshot hash; the rejected snapshot remains viewable and is persisted as a rejected proposal revision. Only the latest exact snapshot can subsequently be approved.

Run the repository tests with:

python3 -m pytest -q

On a host whose system python3 is older than 3.11, create an isolated Python 3.12 environment with uv and run the same commands through it:

uv venv --python 3.12
uv sync --extra dev
uv run python -m pytest -q

The offline demo itself remains dependency-free. The optional AWS evidence sink uses the separately declared hosted extra:

uv sync --extra dev --extra hosted

ChangeProof on ChangeProof

Before any video, screenshot or hosted demo, ChangeProof scans this repository and its hosting configuration against the cited M-SC controls:

PYTHONPATH=src python3 -m changeproof self-check --repo .

The command writes an ignored, locally generated manifest to .changeproof/self_compliance_proof.json. It identifies frameworks checked, file-hash evidence for each automated control and disclosed exceptions. This is a self-assessment—not legal advice, certification, a penetration test, SOC 2 report or independent audit. Public hosting and live uploads remain blocked.

Build Week implementation boundary

Before July 13, the repository contained product research, architecture and submission-planning documentation. The runnable Python package, synthetic lender fixture, reviewer UI, governed approval flow, integrity checks, executable tests and proof manifest were created during Build Week. Pre-existing BBB concepts informed the design, but the demo neither imports nor requires BBB code or data.

Codex accelerated repository analysis, implementation, deterministic fixture and test creation, adversarial review, documentation and end-to-end verification. The entrant retained the product decisions: the single fictional-lender scope, human approval boundary, evidence requirements, local-first architecture and final acceptance of every change.

Multi-agent Codex build

The primary Sol session decomposed and reviewed the architecture and final integration. Terra implemented the exact-plan governance boundary; Luna built integrity validation and adversarial tamper coverage. The primary session then integrated, tested and accepted each increment. This mirrors ChangeProof itself: specialized proposals, explicit boundaries, independent verification and a human-controlled final decision. The dated record is in docs/build-week/PROVENANCE.md.

GPT-5.6 integration

When OPENAI_API_KEY is configured, ChangeProof uses the OpenAI Responses API and gpt-5.6-luna for schema-constrained obligation extraction. It rejects any citation that is not copied byte-for-byte from the source and any deadline not supported by that quote. Model, call time, response ID, prompt/schema/source and validated-response hashes are preserved in the proof lineage.

The proof manifest promotes extraction and drafting into first-class model evidence records. Each record binds the phase, actual model or deterministic provider, prompt version, schema version and validated response hash under its own canonical SHA-256; the manifest also binds the combined evidence chain.

GPT-5.6 Terra also drafts a complete replacement proposal for each of the eight artifacts. ChangeProof binds every proposal to a known fixture ID and path, requires exact coverage, validates JSON/Python syntax and the 30-to-21 change, then displays locally generated diffs. These remain proposals until the reviewer approves the exact snapshot hash.

Without a key—or when either model output fails validation—the same full golden path uses the checked-in deterministic obligation and proposal package and records the fallback reason. Model tiers and the local cache are configured through .env.example.

Cost-effective governance AI

ChangeProof treats model efficiency as proof evidence. Exact source-span SHA-256 keys root a local result cache, so an identical validated obligation can reuse its bound response without another API call. Prompts keep system policy, schema and static task text in a stable prefix and dynamic evidence at the end, with a stable prompt_cache_key. Provider-reported input/output, cached and cache-write tokens flow into the proof certificate alongside execution mode, estimated cost and cost avoided. Missing provider usage remains “not reported”; it is never invented.

The default tier is GPT-5.6 Luna for structured extraction and Terra for redline reasoning; Sol is reserved for an explicit hard-reasoning choice. For non-urgent multi-artifact and multi-framework work, prepare a Responses Batch JSONL with:

PYTHONPATH=src python3 -m changeproof batch-plan --repo .

This command does not submit or require a key. Explicit submission is a separate credentialed API action. As verified 2026-07-16, OpenAI prompt caching applies at 1,024+ input tokens, exposes cache reads/writes, and recommends stable prefixes and keys; Batch advertises 50% lower cost with a 24-hour completion window. GPT-5.6 cache writes are billed at 1.25× uncached input, so the recommendation surface also flags low reuse. Sources: prompt caching, Batch API, model guidance, and pricing.

Runtime dependencies: Python 3.11+ only; the optional OpenAI call uses the Python standard library. Development verification uses the dev extra; the optional AWS evidence sink uses the hosted extra. All fixture content is synthetic; no private services, credentials or third-party data are required.

License

Licensed under the Apache License 2.0.