Add a paid design-partner pilot

.github/ISSUE_TEMPLATE/design-partner-pilot.yml

@@ -0,0 +1,105 @@
+name: Paid design-partner pilot
+description: Check whether one private OpenAI Agents workflow fits the 7-day compatibility pilot
+title: "[Pilot]: "
+labels:
+  - question
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Read the [complete fixed scope and data boundary](https://github.com/FU-max-boop/agentrunproof/blob/main/docs/design-partner-pilot.md) first. This is a public intake for the USD 20, 7-day design-partner pilot. Do not paste private code, credentials, prompts, customer data, production traces, or certificates. An issue is not an order; scope and payment are confirmed separately before implementation work begins.
+  - type: input
+    id: organization
+    attributes:
+      label: Team or public project
+      description: Name the team or link a public project. Write "private team" if neither can be public.
+    validations:
+      required: true
+  - type: dropdown
+    id: sdk-version
+    attributes:
+      label: Current OpenAI Agents SDK version
+      options:
+        - openai-agents 0.20.0
+        - openai-agents 0.21.0
+        - openai-agents 0.22.0
+    validations:
+      required: true
+  - type: dropdown
+    id: python-version
+    attributes:
+      label: Python version
+      options:
+        - Python 3.10
+        - Python 3.11
+        - Python 3.12
+        - Python 3.13
+        - Python 3.14
+    validations:
+      required: true
+  - type: dropdown
+    id: workflow-boundary
+    attributes:
+      label: Primary runtime boundary
+      options:
+        - Session replay or persistence
+        - Function-tool linkage or exactly-once effects
+        - Approval, interruption, or RunState resume
+        - Guardrail durability
+        - Streaming and non-streaming parity
+        - Another public Runner boundary
+    validations:
+      required: true
+  - type: textarea
+    id: invariant
+    attributes:
+      label: Costly invariant
+      description: In one or two synthetic sentences, what behavior must remain true after an SDK upgrade?
+      placeholder: After resuming one approved interruption, the protected synthetic tool runs exactly once and the untouched sibling remains pending.
+    validations:
+      required: true
+  - type: checkboxes
+    id: execution-platform
+    attributes:
+      label: Execution platform
+      options:
+        - label: The workflow can run on customer-owned Ubuntu 24.04 GitHub Actions.
+          required: true
+        - label: The workflow uses the standard text Runner, not Realtime, Voice, Sandbox, or MCP wire protocols.
+          required: true
+  - type: dropdown
+    id: candidate-upgrade
+    attributes:
+      label: Candidate SDK upgrade
+      description: Choose a different exact version from the current SDK version.
+      options:
+        - openai-agents 0.20.0
+        - openai-agents 0.21.0
+        - openai-agents 0.22.0
+    validations:
+      required: true
+  - type: input
+    id: safe-contact
+    attributes:
+      label: Safe contact route
+      description: A public business email, website contact page, or GitHub handle. Do not provide credentials or private links.
+      placeholder: team@example.com or @github-handle
+    validations:
+      required: true
+  - type: checkboxes
+    id: commercial-fit
+    attributes:
+      label: Scope and commercial fit
+      options:
+        - label: If the written scope fits, the team can prepay USD 20 for the 7-day pilot.
+          required: true
+        - label: The pilot covers one repository, one sanitized workflow mapped to an existing contract shape, two preverified exact SDK baselines, and at most 90 engineering minutes.
+          required: true
+        - label: The current and candidate SDK selections are different and are one of the preverified 0.20.0, 0.21.0, or 0.22.0 baselines.
+          required: true
+        - label: Delivery is a customer-applied patch; no private repository content will be provided to the maintainer or external AI tools.
+          required: true
+        - label: I understand this is not an SLA, security audit, hosted dashboard, or guarantee that every defect will be detected.
+          required: true
+        - label: I removed private code, credentials, prompts, customer data, production traces, and certificates from this public issue.
+          required: true

PLAN.md

@@ -58,6 +58,13 @@ Deliver a standalone, installable tool that deterministically checks OpenAI Agen
   redacted replay without matching replacement prose. The full hosted Python 3.10–3.14 by SDK
   0.20.0/0.21.0/0.22.0 matrix passed, the fresh release-bound v3 bundle passed independent audit,
   and the immutable GitHub Release and PyPI wheel/sdist are byte-identical.
+- [ ] (2026-08-26) Run a bounded paid-design-partner experiment without adding a hosted backend:
+  offer three 7-day compatibility checks at USD 20 each for one sanitized OpenAI Agents workflow
+  in customer-owned GitHub Actions, capped at 90 engineering minutes. The paid-signal gate is one
+  unaffiliated team paying before work begins, completing the check, and then prepaying at least
+  one USD 20 continuation month. This intentionally tests low-friction willingness to pay, not
+  sustainable unit economics. Stars, clones, package downloads, interviews, free trials, and
+  verbal interest do not satisfy the gate.
 
 ## Milestones
 
@@ -169,6 +176,19 @@ The milestone is complete when the charter's Gate 4 is satisfied.
   final and immutable, and PyPI serves the byte-identical audited wheel and sdist. The v3 bundle
   records the 0.20 causal comparison; SDK 0.22 support is established separately by the exact
   packaged-wheel matrix.
+- **2026-08-26 — Commercial-validation boundary:** Do not build a dashboard, scheduler, hosted
+  runner, multi-tenant storage, billing integration, or broader framework support before a paid
+  pilot. The initial service uses the released package for one sanitized standard-text workflow
+  and two different preverified SDK baselines from 0.20.0, 0.21.0, and 0.22.0 in customer-owned
+  GitHub Actions. Delivery is a customer-applied patch based only on public or explicitly approved
+  synthetic material; private repository content is not provided to the maintainer or external AI
+  tools. The founding pilot is USD 20 prepaid for 7 days and at most 90 engineering minutes. It
+  maps the workflow to an existing supported contract, runs two fixed SDK baselines, and returns
+  copyable CI commands plus a short report; bespoke adapters, defect fixes, and open-ended triage
+  are excluded. An optional continuation is USD 20 per month and at most 30 engineering minutes
+  for one rerun of the same frozen contract plus a short result note. There is no automatic
+  renewal. Completion plus one prepaid continuation is a paid-demand signal only; this cohort does
+  not validate sustainable pricing or labor economics.
 
 ## Outcomes & Retrospective
 

README.md

@@ -60,6 +60,30 @@ unverifiable evidence. See the
 and the
 [OpenAI Agents integration guide](https://github.com/FU-max-boop/agentrunproof/blob/main/docs/openai-agents.md).
 
+## Paid design-partner pilot
+
+AgentRunProof is testing a small managed service for teams that depend on a private standard-text
+OpenAI Agents `Runner` workflow: a 7-day, fixed-scope compatibility check in customer-owned GitHub
+Actions. It maps one sanitized workflow to an already-supported deterministic contract, exercises
+two different preverified SDK baselines from 0.20.0, 0.21.0, and 0.22.0, and delivers copyable CI
+commands plus a short result report. It does not include bespoke adapters, a defect fix, or an
+open-ended regression investigation.
+
+The founding-pilot price is **USD 20 paid before work begins** for up to 90 minutes of engineering,
+with no automatic renewal and three slots available. The default delivery uses only public or
+customer-approved synthetic material and does not require private repository access. Private
+repository content, production prompts, API keys, raw customer data, and private certificates are
+not shared with the maintainer or external AI tools. This is a low-friction paid-demand experiment,
+not proof of sustainable pricing, a hosted AgentRunProof feature, an uptime SLA, a security audit,
+or a guarantee that every SDK defect will be found.
+
+Read the exact
+[scope, deliverables, and data boundary](https://github.com/FU-max-boop/agentrunproof/blob/main/docs/design-partner-pilot.md),
+then use the
+[design-partner intake](https://github.com/FU-max-boop/agentrunproof/issues/new?template=design-partner-pilot.yml)
+if the fixed pilot fits. Do not put private code, credentials, prompts, or production traces in the
+public issue.
+
 Maintaining a downstream library? The
 [isolated, test-only CI guide](https://github.com/FU-max-boop/agentrunproof/blob/main/docs/ci-adoption.md)
 provides a copyable real-`Runner` contract test and an ephemeral `uv` matrix. The published

docs/README.md

@@ -15,6 +15,8 @@ content-addressed JSON certificate.
   real-Runner contract.
 - [Isolated CI adoption guide](ci-adoption.md) provides a test-only downstream integration that
   does not add AgentRunProof to runtime metadata or the project lockfile.
+- [Paid design-partner pilot](design-partner-pilot.md) defines the fixed commercial experiment,
+  customer-owned CI boundary, price, exclusions, and success criteria.
 - [RunState case study](case-study-runstate.md) follows a released failure through upstream fixes
   and independently checkable evidence.
 - [Examples](../examples/README.md) contains copyable local and CI-oriented samples.

docs/design-partner-pilot.md

@@ -0,0 +1,115 @@
+# Paid design-partner pilot
+
+AgentRunProof is testing whether teams will pay for a bounded service that detects regressions in
+one deterministic OpenAI Agents `Runner` contract across two exact SDK versions. This is a fixed
+managed-service experiment built around the released open-source package. It is not a new hosted
+product and does not guarantee compatibility.
+
+## Founding-pilot offer
+
+- **Price:** USD 20, paid before work begins.
+- **Duration:** 7 calendar days after the written scope is accepted, payment clears, and the
+  agreed customer prerequisites are available.
+- **Capacity:** three founding design partners.
+- **Engineering cap:** 90 minutes across the pilot.
+- **Scope:** one repository, one sanitized standard-text `Runner` workflow, and two different
+  preverified SDK baselines chosen from 0.20.0, 0.21.0, and 0.22.0. The workflow must map to an
+  existing AgentRunProof contract shape; bespoke adapters are outside this pilot.
+- **Execution:** an Ubuntu 24.04 GitHub Actions job owned and paid for by the customer. The default
+  delivery is copyable commands and a workflow snippet that the customer applies, so no private
+  repository access is required.
+- **Renewal:** none automatically. After the pilot, the team may choose a USD 20/month continuation
+  capped at 30 engineering minutes for one rerun of the same frozen contract and a short result
+  note. It does not add a new scenario, support a new framework, or include defect triage or fixes.
+
+Payment proves demand only after both sides accept the exact workflow and access boundary. An
+intake issue or call is not an order, and no repository access or implementation work starts
+before payment.
+
+The pre-payment scope review is limited to public or customer-approved synthetic facts. If the
+agreed invariant cannot be mapped to an existing deterministic contract within the first 20 paid
+engineering minutes, the customer may choose a full refund and no artifact, or approve a revised
+written scope. Time blocked on missing customer fixtures or CI availability pauses the 7-day
+calendar and does not expand the 90-minute cap.
+
+## What the pilot delivers
+
+1. A written contract for one observable workflow boundary, such as session replay, function-tool
+   linkage, approval/resume, guardrail durability, or stream/non-stream parity.
+2. A provider-free synthetic scenario, selected from an existing supported contract shape, that
+   exercises the real `Runner` without production model calls or customer payloads.
+3. A baseline result for two different preverified SDK baselines, including the exact environment
+   and a content-addressed certificate when the scenario fits certificate v1.
+4. Copyable commands and a manual-dispatch GitHub Actions snippet for the customer to apply. The
+   team owns the job, logs, artifacts, and retention policy.
+5. A closing report that records what was run, what passed or failed, limitations, and the exact
+   commands needed to keep or remove the check.
+
+The pilot begins with a short scope review, then maps the sanitized workflow to an existing
+contract shape, runs the two-pin matrix, and returns the commands and report within 7 days. A
+failure is recorded and bounded; root-cause analysis, an application patch, or an upstream fix
+requires a separate scope. If the workflow cannot be represented safely or deterministically, the
+pilot is declined before payment.
+
+## Data and access boundary
+
+- Do not put private code, credentials, prompts, production traces, customer data, or certificates
+  in a public intake issue.
+- Delivery is a customer-applied patch and written commands. Private repository access is not part
+  of the founding pilot.
+- Work uses only public material or a synthetic reproducer that the customer explicitly confirms
+  is safe to process with the maintainer's development tools, including external AI tools. Private
+  repository content, production prompts, customer data, credentials, traces, and private
+  certificates are not provided for processing.
+- No customer finding or artifact is reported upstream or published without separate written
+  approval.
+- Synthetic local tools replace production side effects. The pilot does not execute payments,
+  messages, destructive tools, or other production actions.
+- Private certificates remain in customer-owned CI unless the customer reviews and deliberately
+  publishes a synthetic record.
+- AgentRunProof's private profile reduces raw payload serialization but is not encryption. Its
+  deterministic hashes may reveal equality and may be guessed for low-entropy values.
+- The open-source package and its dependencies still execute as third-party code. Customers keep
+  their normal dependency, action, and secret-review controls.
+
+## Not included
+
+- model-quality evaluation, prompt optimization, an LLM judge, or token-level scoring;
+- a hosted dashboard, multi-tenant storage, uptime SLA, or 24/7 incident response;
+- a security, privacy, compliance, or penetration-testing audit;
+- production interception or sandboxing of arbitrary tool side effects;
+- generic handoff contracts, Realtime, Voice, Sandbox, MCP wire-protocol, or non-OpenAI
+  agent-framework compatibility;
+- SDK versions other than the preverified 0.20.0, 0.21.0, and 0.22.0 baselines;
+- bespoke scenario or framework adapters, root-cause investigations, application patches, or
+  upstream defect fixes;
+- a guarantee that every SDK or application defect will be detected or fixed upstream.
+
+## Good fit
+
+The pilot is intended for a team that already has a standard text `Runner` workflow on Python
+3.10–3.14, can run Ubuntu 24.04 GitHub Actions, and can name one costly runtime invariant. Examples
+include a tool that must run exactly once after resume, a guardrail whose rejected result must not
+reappear, or equivalent post-run observations across streaming and non-streaming execution.
+
+It is not a fit for a team seeking general AI evaluation, a new agent implementation, a production
+observability platform, or free exploratory consulting.
+
+## Apply without sharing secrets
+
+Use the public
+[design-partner intake](https://github.com/FU-max-boop/agentrunproof/issues/new?template=design-partner-pilot.yml)
+to provide only public or synthetic facts: SDK version, CI platform, the invariant in one sentence,
+and a safe contact route. Do not attach private artifacts. If the public scope appears suitable,
+payment and delivery coordination are arranged privately through the safe contact route. Private
+repository access is not part of this offer.
+
+## Experiment success criterion
+
+This page is an offer, not proof of a market. The paid-signal experiment passes only when an
+unaffiliated team pays the fixed pilot price before work, completes the delivery, and prepays at
+least one month of the USD 20 continuation. A completed one-off pilot without a paid continuation
+validates only a low-friction project-service sale, not recurring demand. Even pilot plus
+continuation does not establish sustainable pricing or labor economics at this founding price.
+Stars, repository traffic, downloads, interviews, free pilots, and verbal interest are useful
+diagnostics but not the primary result.