How Moonshot's recommendations work.

Founder contributions

When you run MedBay in Break Through or Reach, the intake includes an optional contribution step. Opt in and your numbers (win rate, deal value, CAC, etc.) join the dataset. Once an industry reaches the sample-size threshold, those medians blend into the engine math for every founder running MedBay afterward.

Composition rules per field

• •Conversion / close rate, industry growth rate: contribution median replaces the secondary-source baseline directly.
• •Ad efficiency floors: blended 50/50 with the baseline.
• •Ratio fields (price-raise %, repeat-revenue %, demand-headroom factor): blended 50/50 below 25 samples, 60/40 weighted toward founder data above 25 samples.
• •Contributions older than 365 days are excluded from aggregation — older self-reported data is less representative.
• •Contribution medians more than 3× outside the secondary-source baseline are rejected as likely outliers.

Outcome tracking

ActionItems generated by an in-OS MedBay save are tagged with the source run. When you mark them done, the OS prompts you to share what happened — revenue moved up, a specific metric changed, no measurable change yet, or made things worse. Skip is always available.

• •Captured outcomes feed back into your next Flight Plan run as framing context. The engine acknowledges what you did and surfaces honest disclosures when multiple prior moves didn't move the needle. Outcomes shape framing, not the math itself.
• •Outcomes are not mixed into industry-wide benchmark composition (see Founder contributions). Outcomes are about your specific business; benchmarks are about your industry.
• •Outcomes are tied to your own runs and never displayed publicly.

Run versioning and comparison

Every in-OS MedBay run you save is versioned. The second time you run in the same mode, the OS surfaces a What changed since your last run panel — input changes you made, what the engine sees differently, and any captured outcomes from the gap between runs (see Outcome tracking).

• •The comparison is a diff. The new run still computes from scratch; prior runs don't bias the engine. The comparison only frames the change so you can read the diagnosis in context, not in isolation.
• •Thresholds: an input change shows when it moves the value enough to plausibly affect leverage (5% for revenue, 3 percentage points for share-style fields, etc.). Output changes show when the top lever shifts, the 90-day projection moves by ≥10%, or engine confidence shifts.
• •Prior runs are your own — never displayed publicly, never mixed into industry benchmarks. Full history lives at /moonshot/flight-plan/runs. Schema evolution is safe: older runs are compared on overlapping fields and skip new dimensions until both sides have them.

ActionItem calibration

Every ActionItem generated by a MedBay save carries a specific proof target from the engine math. Break Through actions inherit a 90-day dollar projection (low–mid–high range); Reach actions inherit the dominant unmet requirement (lead volume, conversion rate, average value, etc.) for the user's stated timeline.

• •When you mark an action done and report an outcome, the OS compares the actual to the projection. Three honest reads: on target, above projected range, and below projected range. If actuals fall below the projected range, the product says so — the leverage math may have overestimated for that specific business.
• •Low-confidence calibrations are flagged on the card. Confidence is the engine's own honest read of how much signal it had — don't expect the same accuracy from a “low confidence” projection as from a “high confidence” one.
• •Calibration data is per-user. We do not aggregate accuracy across users to make marketing claims about engine precision — that bar requires more data than we have. Honest path: build the track record first.
• •ActionItems created before calibration shipped don't have proof targets and render without the target line. Going forward, every saved MedBay run generates calibrated actions.

Per-lever calibration

For a Break Through save, the three saved ActionItems each carry their own projection from their own lever — not a single shared target across all three. Each lever (pricing, retention, channel concentration, capacity, demand) has its own 90-day projection range and confidence; the ActionItem tied to that lever inherits its values. A founder reading their saved actions sees three distinct dollar projections matching three distinct levers.

Each lever currently produces one canonical handoff tool (the framework, generator, or playbook tied to that lever). Standalone tools still exist as their own OS surfaces; the calibration mapping is one ActionItem per lever for now.

Scenario simulation

On Break Through output, an optional scenario panel lets you adjust your reported inputs and watch the leverage math respond in real time. The headline, math block, leverage rankings, and seven-day plan all re-render with the simulated values; honest disclosures, data gaps, and any prior-run comparison stay tied to your actual reported inputs.

• •Scenarios are client-side exploration. They run the same engine functions with different inputs; no server round-trip, no logging beyond aggregate telemetry.
• •Scenarios do not save. To save a new analysis with updated numbers, exit scenario mode and re-run MedBay with your real inputs. This keeps the run history (see Run versioning) free of exploratory noise.

Blackbox cases and economic context

Blackbox is Moonshot's library of researched business failures. Each case is investigative business journalism: who failed, when, the structural cause Moonshot diagnoses, the interventions they tried and missed, and the primary / secondary / tertiary sources behind the read.

Cases can also carry a macro-context panel describing the economic environment at time of failure — broad environment (recession / expansion / mixed / volatile / stable), inflation and interest-rate posture, named industry-specific events (“retail apocalypse,” “post-COVID labor shortage”), and consumer-spending drift.

• •The macro classification distinguishes accelerated failure (existing operational leak, macro compressed the timeline), masked leak (macro tailwinds hid the leak until conditions normalized), neutral (failure would have happened regardless), and unknown (insufficient evidence to classify).
• •Macro context does not affect MedBay engine math or which cases match. It surfaces alongside the case so a founder can ask: does my situation match this case's operational pattern AND its macro backdrop, or just one of them?
• •Editorial posture: macro context is Moonshot-authored. We don't plug into live macro data feeds and don't accept user-submitted macro views. The goal is honest journalism on each case, not real-time economic commentary.

Macro annotation on enrichment

When MedBay surfaces a matched Blackbox case in enrichment, and that case has macro context populated (see Blackbox cases), the card carries a one-sentence annotation under the match framing. The annotation names the classification — accelerated failure, masked leak, or neutral — and weaves in any named macro events from the case.

• •Cases classified as unknown, or cases not yet backfilled with macro context, render without the annotation (no empty placeholder).
• •The match score is unchanged by macro context. Only the framing copy is. A founder reading enrichment cards sees the same matching logic with or without the annotation — annotation is supplementary context, not a re-ranked verdict.

Privacy and verification

One catch-all summary of the privacy posture across every system documented above. None of this is buried in fine print — it's how the substrate is built.

• •Your individual data — answers, runs, outcomes, calibrations — is never displayed publicly and never shown to other users. Only you (and admins reviewing for product quality) can see it.
• •Only aggregated medians at or above the sample-size threshold are surfaced — through industry benchmark composition (see Founder contributions).
• •Contributions are opt-in at submission. You can opt out of dataset inclusion later via the includedInDataset flag.
• •Outcomes are tied to your own runs and never auto-flowed into industry benchmark composition. Calibration accuracy is per-user; we do not aggregate it to make marketing claims.
• •Blackbox cases are publicly visible (with citations) when published; cases in draft are admin-only. Macro context is Moonshot-authored; we don't accept user-submitted macro views.