ExitCanary

Before you enter, prove you can leave.

Live demo · Demo video · Devpost · EV1 Labs

Built by EV1 Labs as part of the Build Week 2026 collection.

ExitCanary is a bounded SaaS exit-readiness prototype. It places a known, synthetic CRM dataset inside a trial account, examines the vendor's exported CSV, JSON, or ZIP packet, and produces a reproducible verdict about what survived the exit.

GPT-5.6 Sol handles the part that benefits from semantic reasoning: mapping an unfamiliar vendor export back to the known canary fields. Deterministic code — not the model — evaluates the confirmed mapping and decides EXIT_READY, NOT_EXIT_READY, or NEEDS_REVIEW.

Why this exists

Teams often test whether a SaaS product can accept their data, while export fidelity can remain untested until a migration is already urgent. An export button is not enough if it silently drops relationships, activity history, attachments, custom fields, timestamps, or Unicode values.

ExitCanary turns that vague procurement question into a concrete pre-purchase test:

1. Start with a versioned synthetic CRM canary pack.
2. Adapt its vendor-neutral tables to the SaaS import format and import them into the trial being evaluated.
3. Request the vendor's advertised full export.
4. Upload the returned CSV, JSON, or ZIP packet.
5. Review the proposed semantic mapping.
6. Receive a deterministic exit-readiness verdict and receipt.

Competition scope

The Build Week prototype deliberately stays narrow:

• one bounded synthetic CRM profile;
• CSV and JSON evidence, optionally grouped inside ZIP archives;
• one source record table per canonical CRM entity in the current real-upload normalizer;
• manual import/export instead of live vendor integrations;
• semantic field mapping with GPT-5.6 Sol;
• deterministic checks over records, values, identifiers, relations, timestamps, Unicode, custom fields, history, and attachment evidence;
• bundled flawed and complete demo exports.

This is not a universal SaaS portability auditor, legal compliance guarantee, source attestation, or proof that a vendor generated an export honestly.

Quick start

Prerequisites

• Node.js >=22.13.0
• pnpm 10.33.2 or a compatible pnpm 10 release
• an OpenAI API key with access to gpt-5.6-sol for live semantic mapping

Install and run

pnpm install --frozen-lockfile
cp .env.example .env.local
pnpm dev

Open http://localhost:3000.

Set the server-only key in .env.local when testing the live model path:

OPENAI_API_KEY=your_key_here
EXITCANARY_LIVE_MAPPING_ENABLED=true
EXITCANARY_PUBLIC_ORIGIN=http://localhost:3000

EXITCANARY_QUOTA_SALT is optional for local demo use. When configured, use a private random value of at least 16 characters and never commit it. The current process-local rate limit is a demo guardrail, not a production quota.

Set EXITCANARY_PUBLIC_ORIGIN to the exact canonical HTTPS origin on a public DNS hostname in production; the request boundary does not trust forwarded headers. Live mapping runs only when EXITCANARY_LIVE_MAPPING_ENABLED is exactly true; every other value forces a labeled deterministic fallback even if a key is inherited. Do not expose a public live key behind only the process-local limiter: durable identity/quota, abuse monitoring, and an operational kill switch are pre-public gates.

The published Build Week judge URL is https://exitcanary.vercel.app. It is intentionally keyless and fallback-only. Its fail-closed environment check and signed-out verification procedure are in docs/PUBLIC-DEMO-DEPLOYMENT.md. Run pnpm preflight:public-fallback with the exact production environment before every future public release.

Never expose the key through a NEXT_PUBLIC_ variable. The no-key path for a real upload defaults to local deterministic exact-alias mapping unless the user explicitly consents to the GPT path. The bundled judge demo is separate: it uses a disclosed, pre-mapped synthetic fixture and never calls GPT. The mapping is not confirmed until the user reviews it. Confirm all three modes in the final build before a public handoff.

Run the bundled demo

1. Open the app and start the bundled flawed CRM export.
2. Verify the interface labels its versioned mapping as bundled and pre-mapped; this repeatable path does not claim a GPT call.
3. Inspect the 33 mapped fields, select Mark all reviewed, then Verify confirmed mapping.
4. Observe six deterministic failures and NOT_EXIT_READY.
5. Use the simulated demo fix action. It swaps to a bundled complete fixture; it does not modify a vendor account or repository.
6. Re-run the evaluation and observe EXIT_READY with a new digest.

To exercise semantic mapping, choose Use my export and upload a supported CSV, JSON, or ZIP. That path parses locally. By default, it maps exact aliases in the browser and sends no export evidence to OpenAI. If the user selects the explicit GPT-5.6 consent control, bounded field evidence and up to five sample cell values per field are sent to /api/map. The result is labeled live or fallback. Unresolved targets can be assigned from bounded source candidates; every mapping must then be reviewed before evaluation.

The current normalizer rejects a mapping that spreads one canonical entity's required fields across multiple source tables. That bounded support limit is an explicit error, not a partial or successful verdict.

Public synthetic judge ZIPs are checked in under examples/exports and are also available from GET /api/demo-export?variant=flawed and GET /api/demo-export?variant=complete on a running instance.

See the recording script for the locked 93-second walkthrough and its claim gates.

Built with Codex and GPT-5.6

This repository was created in Codex during the Build Week submission period. Codex was the primary engineering workspace: it turned the product thesis into explicit invariants, split implementation into bounded core, parser/API, interface, and documentation workstreams, and was used for contract tests, integration review, UI verification, and claim red-teaming.

The entrant retained the key product and risk decisions:

• test a narrow synthetic CRM export instead of claiming universal SaaS support;
• keep vendor import/export manual in the prototype;
• use GPT-5.6 Sol only for evidence-referenced semantic mapping;
• fail closed when meaning is ambiguous and require confirmation of a complete mapping before evaluation;
• reserve EXIT_READY, NOT_EXIT_READY, and NEEDS_REVIEW for deterministic code;
• enter Work & Productivity rather than reuse a developer-tool framing.

One concrete Codex review already caught a parser-to-mapper size-boundary mismatch; the adapter and targeted test were tightened before integration. GPT-5.6 Sol's runtime contribution is different from Codex's build role: the model interprets unfamiliar export field names under a strict structured-output contract and has no verdict field.

The source is public at github.com/pakales/exitcanary, and the bounded public judge smoke passed against the fallback-only deployment. The 93-second demo is public at youtu.be/FWoRiuhg8sY. The Build Week entry is submitted at devpost.com/software/exitcanary with primary Codex session ID 019f813d-fdfb-7e93-8365-783b07ade86f; signed-out access and the Devpost Submitted state were verified on 2026-07-21.

Verification

pnpm lint
pnpm typecheck
pnpm test
pnpm test:public-preflight
pnpm test:public-smoke
pnpm build
pnpm audit:prod
pnpm verify

pnpm verify is the repository gate. Tested runtime source snapshot a178969062a631aa669dcdf664b9c05f4a297d28 passed lint, typecheck, 12 Vitest files / 86 tests, five public-preflight tests, seven public-smoke tests, the production build, and the production audit. The full manual and automated matrix is in docs/TESTING.md.

The bounded synthetic live-mapping smoke also passed on the current prose-free model contract: gpt-5.6-sol returned 33 proposals, zero unresolved targets, no warning, and no verdict field. This does not authorize a public paid endpoint or change the public deployment's separately verified fallback-only posture.

The bounded black-box verifier passed against the published canonical origin for the tested runtime snapshot. It sends only checked-in synthetic canary data, rejects redirects, and never sends credentials:

pnpm smoke:public-judge -- https://exitcanary.vercel.app

The command verified the public security headers, bounded canary and judge ZIPs, the exact 33-target registry, fallback-only mapper identity, foreign-origin rejection on both APIs, all three deterministic verdict states, injected-verdict rejection, digest stability, and digest separation. Browser accessibility, signed-out file upload, and provider environment remain separate manual gates.

How the decision boundary works

Layer	Responsibility	May set the verdict?
Browser parser	Builds a bounded normalized manifest from supplied files	No
GPT-5.6 Sol	Proposes semantic mappings with evidence references	No
Human review	Resolves bounded candidates and explicitly confirms every mapped field	No
Server evaluator	Compares confirmed evidence with the application-owned canary contract	Yes
Server receipt builder	Binds versions, packet, mapping, and result into a SHA-256 digest	No

The digest detects changes to the bounded receipt inputs. It is not a digital signature, trusted timestamp, tamper-proof seal, or proof of evidence origin.

Privacy and security posture

• OPENAI_API_KEY remains server-only.
• Uploaded values are untrusted data, never model instructions.
• OpenAI Responses API requests use store: false.
• The application is designed not to persist export packets or model output.
• After explicit GPT consent, /api/map receives field paths and up to five bounded sample values per field; /api/evaluate receives the full normalized packet and confirmed map.
• Model output is schema-constrained, bounded, and never rendered as raw HTML.
• A model failure cannot upgrade or downgrade the deterministic truth.
• Production responses add HSTS, frame denial, same-origin isolation, and a restrictive baseline content-security policy.

Review docs/THREAT-MODEL.md before changing parsing, model prompts, request handling, persistence, or receipt generation.

Documentation

• Product brief
• Architecture
• Evidence contract
• Threat model
• Testing
• Public judge deployment
• Demo script
• Build Week submission record
• Public sample data

Build Week

ExitCanary is submitted in the Work & Productivity category. The official challenge evaluates technological implementation, design, potential impact, and idea quality. The source repository, fallback-only judge URL, public YouTube demo, and public Devpost page are live. The exact Codex /feedback session ID and the verified submission evidence are recorded in docs/BUILD-WEEK-SUBMISSION.md.

License

MIT — see LICENSE.