Run the economic debugger commercial pilot

Historical validation plan (superseded 2026-08-31): retained as research evidence, not the active go-to-market plan. The owner subsequently selected a self-serve recurring SaaS centered on economic change control, with bounded model-change backtests as the first-value event. Do not use the closed-ledger pilot or its $500–$1,500 pricing experiment as the current product offer.

Bottom line

Use the free self-hosted diagnostic to find production teams with an organization-level reconciliation problem. The paid-product hypothesis is one plain promise: close the AI spend ledger by tying provider-billed dollars to the team, workflow, and outcome that incurred them—or marking the variance as unreconciled. The repository now contains a normalized provider-bill import and allocation report, but it has not closed a real customer export and the managed organization boundary is not implemented. Do not sell the debugger report—or this unvalidated primitive—as a finished hosted product.

Offer boundary

Layer	User receives	Commercial role
Free, self-hosted	Evidence ingestion, cost per successful outcome, timeline, cohort analysis, failed-run exposure, one bounded diagnostic	Trust, activation, and proof that the data can answer an operating question
Paid hypothesis	Provider-invoice reconciliation, cross-team ownership and rollups, chargeback/showback exports, retention, organization access controls, and signed change evidence	Organization economics control

The existing reconciliation screen calibrates ground-truth costs against estimates. It is not provider-invoice reconciliation. The new provider-bill API imports immutable billed-cost buckets and allocates them through measured workflow/outcome evidence while preserving variance. It is the paid-product prototype, not proof of willingness to pay. The existing backtest supports a bounded model-swap case; it does not yet prove savings for structural workflow changes.

Low-touch activation loop

1. A developer can run zeroth-econ demo to inspect the product's bounded output without a server or UI. It does not count as a first diagnostic, production validation, or commercial evidence.
2. The developer installs the backend without the UI and instruments one real workflow with InstrumentationClient.authenticated. The setup check uses confirmed execution and outcome delivery so rejected evidence cannot look like activation.
3. An Admin defines success for the exact workflow version; the definition is immutable and content-digested.
4. The team collects 7–14 days of representative execution and outcome evidence.
5. The developer runs zeroth-econ diagnose and keeps the artifact locally.
6. For OpenAI, an Admin converts one complete local Costs API page with zeroth-econ normalize-openai-costs, then runs zeroth-econ reconcile. Other providers use the normalized contract until a real export justifies an adapter. Neither the provider data nor the report is uploaded.
7. If the closure report exposes a current operating problem, the team opens the public design-partner request with aggregate bands and problem descriptions only.
8. Zeroth qualifies the request asynchronously. A pilot advances only when the organization names a current provider-bill closure problem, has usable cost evidence, and identifies a plausible budget owner.

Do not request raw prompts, responses, traces, credentials, subject_id values, customer names, proprietary workflow names, or unredacted reports in a public issue. Move any later confidential exchange to an approved private channel with explicit retention and deletion terms.

Qualification experiment

The following are hypotheses to test, not validated market facts or product requirements:

• a real provider export or Costs API result is available;
• a recurring manual, disputed, or missing reconciliation process exists;
• the primary gap is reconciliation, allocation, outcome attribution, or evidence trust—not “no current provider-bill closure problem”;
• at least 100 resolved runs in the selected window;
• at least 90% outcome coverage;
• no undefined workflow versions in the report window;
• a closure report shows variance, unresolved outcomes, or an ownership gap; and
• an explicit monthly budget band and organization owner exist.

Do not reject a strong buyer solely for missing a numerical threshold. The purpose is to distinguish a real recurring operating decision from curiosity about a dashboard.

Track this funnel without collecting production content:

demo → real instrumentation → first event → first resolved outcome → first diagnostic
        → first provider-bill closure report → design-partner request → paid pilot

Record counts and elapsed time between stages. The decisive conversion is not repository interest or report generation; it is an organization asking to pay to close a reconciliation or governance gap.

Pricing and payment gate

Use the issue form's budget bands to learn willingness to pay before publishing a permanent SKU. A reasonable founding-pilot test is $500–$1,500 per month, month to month, for one organization and a deliberately limited integration scope. That range is a pricing experiment, not evidence of market acceptance. Do not accept payment or promise a managed SLA until all of these exist:

• tenant isolation verified for the actual hosted topology;
• encrypted backups plus a witnessed restore procedure;
• retention and deletion controls;
• authentication and organization authorization appropriate to the buyer;
• service terms, privacy terms, support boundary, and incident contact; and
• at least one real provider-billing import with a reconciliation audit trail.

No bespoke provider or workflow integration belongs in the founding offer unless it is reusable and separately scoped. The initial paid deliverable must remain the closed-ledger result, not general consulting.

Build trigger and stop criteria

The normalized reconciliation primitive was implemented to test the hypothesis. Do not build provider credential storage, multiple bespoke adapters, or the entire organization shell until at least one qualified organization requests this closure result without being led to the answer, or three qualified organizations independently rank it as their first paid need.

Pause or change the wedge if either experiment fails:

• 20 inbound requests produce fewer than two real diagnostics or closure reports; activation is the problem, so simplify instrumentation and the diagnostic path. Do not substitute PyPI downloads for this explicit evidence.
• 20 real closure reports produce no qualified paid request; the report is useful but not a buying trigger, so test managed hosting plus SSO/RBAC/retention as the simpler paid offer.

Use the launch experiment for the single-channel copy, attribution fields, release gate, and 14-day decision rule.

Adversarial review

The strongest objection is that provider clouds, gateways, and observability vendors can add cost allocation. Zeroth only has a defensible position if it reconciles billed money through workflow execution to business outcome and preserves evidence quality and ownership boundaries. Cost charts alone are not a moat.

The key unknowns are whether teams have reliable outcome signals, whether finance will trust developer-generated evidence, and whether invoice variance is painful enough to fund another system. The safer fallback is to keep the debugger open source and sell managed hosting, retention, and organization access controls rather than claiming unique savings intelligence.