# Experiment Velocity Planner

The [Experiment Velocity Planner](https://notambourine.com/experiment-velocity) compares an existing experimentation process with a measured-velocity policy. It runs entirely in the browser, needs no account, and stores scenario state only in the URL fragment when a user copies a link.

## Model

Measured velocity asks whether a variation is unlikely to be worse than control by more than an agreed loss. A safe result is not a win; it supports release within that loss boundary.

Set two controls and the planner calculates the third:

| Control            | Meaning                                               |
| ------------------ | ----------------------------------------------------- |
| Decision duration  | P80 time to a decision, with a 14-day reporting floor |
| Evidence threshold | Probability required for a safe decision              |
| Acceptable loss    | Largest permitted decline in the chosen metric        |

Defaults are 90% evidence, 5% relative acceptable loss, purchase conversion, and calculated duration. They are starting assumptions, not recommendations.

The annual view gives both policies the same deterministic stream of test-ready changes. It models capacity, test coverage, decision time, untested work, winners, losers, and proxy-to-purchase linkage. Policy deltas are calculated within each paired run before percentiles are taken.

## Metrics

Only one metric controls a decision. Metrics closer to purchase carry more business meaning but usually need more traffic.

| id         | Metric              | Baseline | Answers             | Limitation                |
| ---------- | ------------------- | -------- | ------------------- | ------------------------- |
| `bounce`   | Bounce rate         | 38%      | Initial engagement  | Farthest from purchase    |
| `pdp`      | PDP view rate       | 56%      | Product discovery   | Views do not prove orders |
| `atc`      | Add-to-cart rate    | 8.5%     | Cart intent         | Intent can fail later     |
| `checkout` | Begin-checkout rate | 4.7%     | Checkout intent     | Starts are not orders     |
| `purchase` | Purchase conversion | 2.9%     | Completed purchases | Rare events need time     |

## Scenario links

Encode a JSON object as base64url and place it after `#scenario=`:

```js
const state = {
  sessions: 250000,
  aov: 85,
  surface: 'homepage',
  selectedMetric: 'atc',
  calculated: 'duration',
  evidence: 90,
  loss: 5,
};

const hash = btoa(JSON.stringify(state)).replaceAll('+', '-').replaceAll('/', '_').replaceAll('=', '');
const url = `https://notambourine.com/experiment-velocity#scenario=${hash}`;
```

Unknown or omitted values fall back to defaults. Numeric values are clamped to these ranges:

| Key               | Min |       Max | Default |
| ----------------- | --: | --------: | ------: |
| `sessions`        |   1 | 100000000 |  250000 |
| `aov`             |   0 |   1000000 |      85 |
| `eligibleTraffic` |  10 |       100 |      80 |
| `effectiveSample` |  40 |       100 |      90 |
| `duration`        |   7 |       120 |      22 |
| `evidence`        |  70 |        99 |      90 |
| `loss`            |   1 |        30 |       5 |
| `currentCoverage` |   0 |       100 |      70 |
| `currentDays`     |  14 |       120 |      60 |
| `changesMonth`    |   1 |        12 |       4 |
| `testCapacity`    |   1 |         5 |       1 |
| `winnerShare`     |   0 |         8 |       2 |
| `loserShare`      |   0 |         8 |       3 |
| `winnerLift`      |   1 |        20 |       5 |
| `loserLoss`       |   1 |        25 |       7 |
| `proxyLink`       |  20 |       100 |      70 |

Supported enums:

- `surface`: `homepage`, `pdp`, `cart`, `checkout`, `sitewide`
- `selectedMetric`: `bounce`, `pdp`, `atc`, `checkout`, `purchase`
- `calculated`: `duration`, `evidence`, `loss`
- `currentPreset`: `winner`, `limited`, `capacity`
- `currentUntested`: `hold`, `ship`
- `readiness`: array of exactly 6 booleans

## Limits

The planner cannot validate production instrumentation, prove that an upstream metric causes purchases, or guarantee that a safe variation is profitable. Faster statistics do not repair unstable assignment, missing exposure logs, inconsistent metric definitions, or bad traffic allocation.

- [NoTambourine agent reference](https://notambourine.com/AGENTS.md)
- [Markdown index](https://notambourine.com/llms.txt)