Cookwala Action Plan: answering the critics

Status: draft, 2026-10-04. Owner: the Cookwala maintainer. Review this plan every month and update the status column.

This plan answers every concern raised so far in one place:

• the red-team critique of the idea;
• the technical critique of the specs;
• what an Elon Musk–style adopter, Claude and Anthropic would need;
• what WFP, food banks and WHO would need;
• the stakeholders who may resist;
• the technical reviewers: standards bodies, food-safety and nutrition professionals, robotics researchers, AI-safety and security reviewers, humanitarian data bodies, regulators and lawyers.

1. The six commitments behind the plan #

1. Evidence before claims. Simulators illustrate; pilots prove. Every public number says which one it is.
2. A small core, optional profiles. A device should be able to implement Cookwala in a week. Everything else must earn its place with real usage.
3. Safety is enforced on the device, whatever a recipe, an agent or the network says.
4. People without robots come first. Food banks, kitchens and phones before humanoids.
5. Personal data stays home. Minimal by default, erasable, never sold, never used for training without opt-in.
6. Neutral and durable. A foundation, a patent pledge and seats for the people affected. Nothing depends on one founder, one company or one AI model.

2. Concern register #

Each concern has an id, the people who raised it, the response, the workstream (section 3) and the evidence that closes it.

Spec progress (2026-10-04, Core 0.2) #

3. Workstreams #

W1. Scope: Cookwala Core 1.0 #

• Core 1.0 on one page. It contains:
  • a recipe: ingredients, steps, done criteria and safety limits;
  • device capabilities;
  • one execute-and-status API;
  • an execution log.
• Everything else becomes a profile marked experimental: Missions, the ledger, the market, relief, health and reasoning. Each profile needs two implementers and real usage before it can graduate.
• Rewrite PLAN.md around the core, and move the broad vision into a separate vision document.
• Position for end-to-end learning: Cookwala says what to make, when it is done and what must never happen, not how to move.

W2. Technical correctness (the technical critique) #

• Cooking operations: physical envelopes per operation (temperature band, power, agitation, time), a sensor fallback ladder (sensor → model estimate → time), and published test vectors.
• Numbers: Celsius only on the wire, absolute tolerances, decimal money, densities and kitchen units.
• Schemas: strict everywhere (unevaluatedProperties: false with x- extensions), bundled for offline use, with version-negotiation rules. The Humanitarian Profile already shows the pattern.
• Missions: an event log plus derived state, a formal state machine, one sequencer, merge rules.
• Trust: identity bound to keys (did:web plus device keys), rotation, revocation, offline verification, witnessed checkpoints, and SCITT alignment.
• Selective disclosure: a salted-digest format so partial views still verify.
• Events: at-least-once delivery with idempotency keys and ordering per subject; heartbeats; latency classes.
• Single source of truth: generate OpenAPI, GraphQL and AsyncAPI from the schemas, with contract tests in CI.
• Conformance: a test suite plus a verifier for hashes, signatures and ledgers; the simulator must pass it.

W3. Safety and agent rules #

• Safety limits enforced on the device: food-safety temperatures, allergen blocks, hot oil and blades, which hold whatever any recipe or agent says.
• Local stop: the stop button works without the network. "Safety is local" is a normative rule.
• Untrusted text: every free-text field is data, never instructions, for agents.
• Recalls: robots stop using a bad recipe or extension within minutes.
• Incident reporting: an open, anonymous system modelled on aviation's confidential near-miss reporting.
• Safety case: mapped to ISO 13482, IEC 60335 and UL 3300, for certifier review.
• Recipe safety review track: automated checks against rule packs, plus human review for high-risk recipes.
• Health claims policy: general nutrition only; anything for a medical condition needs clinicians.

W4. Privacy, data and competition #

• Defaults: local-first and minimal data. Health, religion and household data never leave the home unless the person chooses.
• A data-protection impact assessment for the core and each profile.
• The ledger stores hashes only; content lives in erasable storage.
• Execution data for training is opt-in, can be withdrawn, and contributors share in the benefits if it is licensed.
• Competition: an information-exchange policy (aggregation, delay, no price data between competitors), reviewed by counsel.

W5. Evidence and honest modelling #

• Simulators:
  • a "robots with good vendor integrations" baseline replaces the strawman;
  • ranges instead of single numbers, plus sensitivity analysis and assumption sliders;
  • clean cooking, robot lifecycle carbon and rebound effects added;
  • every chart says "illustrative model".
• Claims policy: every public number is tagged measured, modelled or assumed, with its source.
• Pre-registered pilot evaluation with an independent evaluator (J-PAL, IPA or a university), including kill criteria. Negative results are published too.

W6. Humanitarian pilot (Egypt first) #

• Partners: the Egyptian Food Bank (or another member of the Global FoodBanking Network) as pilot partner; the WFP Innovation Accelerator as funding route. These are targets we have not approached; none has been contacted or has agreed to anything.
• Approvals and reviews:
  • university ethics approval;
  • a data-responsibility review against ICRC and OCHA guidance;
  • review of the rule pack by a food-safety officer and a dietitian, plus Egypt's National Food Safety Authority rules.
• Tools: an SMS gateway and CSV workflow (level H0), then an API adapter to the partner's current tools (level H1).
• Recognition: apply to the Digital Public Goods Alliance.

W7. Device demo, AI integration and dataset #

• First device: a smart oven or kitchen-robot startup. A smart oven is simpler and sooner than a humanoid.
• Reference executor and a cooking benchmark: 10, then 50 recipes, with a success rate per recipe.
• AI integration: MCP servers for the index and hub, plus a Cookwala Agent Skill, tested with at least two model families.
• Kitchen agent-safety benchmark:
  • poisoned recipes, overspending, unsafe temperatures, allergen traps, escalation;
  • results published for several models.
• Consented execution dataset: published as an open benchmark at a robotics venue (RoboCup@Home, or an ICRA/IROS workshop).

W8. Governance and intellectual property #

• Charter: a technical steering committee and decision process, documented in GOVERNANCE.md.
• Seats for food banks, low-income countries, labour, privacy, dietitians and device makers.
• Policies: a patent non-assertion pledge, a trademark policy for "Cookwala" and a certification mark, a code of conduct, and a security disclosure policy (SECURITY.md).
• A neutral home: the Joint Development Foundation, the Linux Foundation, or a similar body.

W9. External review program #

See section 5.

W10. Sustainability and business #

• Two entities: an open standard (foundation) and a company (a public benefit corporation fits).
• Revenue: certification, hub software, consented dataset licences, grocer and marketplace fees. Free forever for relief.
• Funding pipeline (targets only; none approached): WFP Innovation Accelerator, foundations (Rockefeller, Bezos Earth Fund, Google.org), the Anthology Fund (only with pilot data), regional programs (ITIDA, Hub71).

W11. Inclusion, culture and labour #

• World-cuisine vocabulary, starting from fifirecipes' Egyptian recipes, with community-governed extensions.
• Accessibility: languages, low literacy and voice.
• A dietary-compliance profile (halal, kosher, vegetarian and others) written with certifiers.
• Labour: early dialogue with unions; prioritise uses for elderly and disabled people and community kitchens.

W12. Narrative and communication #

• New lead message: "an open, safe way for people, kitchens and machines to plan, rescue and cook food with less waste", with robots as one consumer.
• Demo first: contact influencers only after a working demo.
• Use the stakeholder tracker as the single record of outreach; review it weekly.

4. Phases and decision gates #

Kill and pivot rules:

• If Gate B fails twice, narrow Cookwala to the Humanitarian Profile and the recipe format only.
• If the pilot shows less than a 5 % gain, publish the results and redesign before scaling.

5. External review program (sequenced) #

6. Who this answers #

7. People and money needed #

Indicative budget for phases 0–2: about US$90k for the pilot (see the concept note), plus about US$60–120k for engineering, the security audit, legal and reviews. The plan should be financed through grants and pilot funding before any equity investment.

8. Next 30 days #