Pricing Strategist

Marketplace listing update: added the launch explainer film (`media.demo_video_url`). No code, schema, copy or guidance changes; hero and screenshots unchanged from 0.1.1.

Marketing media only. No behaviour, no guidance and no script changed.

- **Card visuals shipped.** A hero plus three panels rendered at propagation: the provenance report
  failing, the six-step pipeline with its stop branch, and a cohort migration with no lawful
  destination. The same hero is reused as the blog cover and the listing hero rather than rendered
  three times.
- **Registry count corrected.** The manifest description said 37 evidence-bundled registry items; the
  actual figure is 39. The count moved during Stage 6 when a live competitor re-capture added
  `teardown-firecrawl` and `teardown-apollo-partial`, and the README and changelog were updated then
  while `package.yaml` was not.

First release.

A pricing **strategist**, not a number-validator. Six steps — understand the service, research what the
field actually charges, pick the game, pick the value metric, set the mechanics, set the number — and it
refuses to reach the number until the game has been named.

### What made it worth building

Two widely-installed public pricing skills give opposite instructions on the most-cited figure in
pricing. One says aim for the customer to get 10× the value they pay; the other says a 10× ratio means
you are massively underpriced. Neither cites anything. A prior-art audit of 48 pricing and GTM agent
skills found zero shipping an enforcement script, an evidence registry, or a learnings set.

So the first thing this ships is a way to be wrong in public.

### Added

- **Five enforcement scripts**, each printing PASS/FAIL with numbers and exiting non-zero on failure
  - `verify-claim-provenance.mjs` — classifies every numeric claim as BENCHMARK (needs a public URL) or
    BRAND-FACT (needs doc + version + date) and fails class confusion in either direction
  - `verify-value-ratio.mjs` — hard-fails cost-plus and price-below-cost; primary check is the observed
    price-based loss rate against 20–25%, with the 2–4× value ratio as the first-price fallback
  - `lint-pricing-page.mjs` — four scored rules on a structured page capture, two INFO-only
  - `verify-tier-migration.mjs` — every cohort needs a lawful next state; structure only, never predicts
    churn
  - `verify-coordination.mjs` — the four coordination cross-references must agree across SKILL.md,
    package.yaml and instructions/AGENTS.md
- **10 references**, one per pipeline step plus migration, agent/usage pricing, and disclosure
- **7 recipes**, flagship `full-pricing-strategy` walking all six steps; the other six are doors into it
- **39 registry items** — benchmarks, seven pricing-page teardowns, six model blueprints, four migration
  postmortems, and the strategy canon. Every BENCHMARK item carries a resolvable public URL; the build
  refuses to emit one that does not
- **11 learnings**, each with non-empty references
- `instructions/AGENTS.md` for non-Claude runtimes and `instructions/DEPLOYMENT.md` for the output
  contract

### Decisions worth recording

- **The primary defensibility check is the observed loss rate, not the value ratio.** A loss rate is a
  count of things that happened in your own pipeline; a value ratio rests on the buyer's estimate of
  savings they have not realised yet. This inverts how both upstream sources frame the work.
- **The 10× rule ships labelled FOLKLORE rather than dropped.** It is the number users arrive with, and
  refusing to discuss it is unhelpful. A tagged mention passes the provenance gate; an untagged one fails.
- **Grandfathering is the default, and the 2026 legacy-sunset wave ships alongside it.** The industry is
  currently doing the opposite of the rule, at scale. Both registry items are cross-linked and neither is
  shippable without the other — a rule that cannot acknowledge its own counter-example is dogma.
- **No fifth script for disclosure compliance.** The FTC negative-option rule was vacated in July 2025
  and rulemaking reopened in March 2026. A script would freeze a struck-down legal position into
  executable code. That surface ships as a dated reference file that routes to counsel.
- **Migrations are checked structurally and never predicted.** In the captured evidence a 2.1× increase
  lost half a customer base while a 4× increase was reported as the best decision the founder ever made.
  Magnitude does not predict outcome; whether the existing base was carried plausibly does.

### Rules the research killed

Two of six planned lint rules did not survive contact with five real pricing pages, and both were
demoted to INFO:

- **Tier count > 3 on a self-serve page.** Two widely-admired pages ship four tiers each. The modelling
  error was counting contact-sales Enterprise as a tier.
- **Enterprise price hidden.** All five captured pages hide it. Flagging universal practice as a defect
  is noise.

A lint that fires on good pages is worse than a missing lint.

### What the test run changed

The skill was run end-to-end against a real, unresolved case, and the process found three defects in
itself. All three are fixed here rather than deferred.

- **The provenance gate failed the skill's own recommendation at 36 unsourced claims.** Two of those
  failures were script defects: fenced code blocks were being scanned as assertions, so verbatim script
  output — which the skill *requires* pasting — read as the author's own claims; and there was no
  PROPOSAL class, which made the skill's own mandated tier-table format structurally un-passable. Both
  fixed. The remaining failures were real memo defects and the memo was fixed. `36 → 11 → 7 → 0`.
- **`lint-pricing-page.mjs` had a false negative on credit-model pages.** `model: "credit"` was not in
  the consumption-model set, so R1 and R2 reported N/A and a rate-less dimension went unchecked. Fixed:
  a declared dimension is metered whatever the model label says.
- **`references/entry-motion-design.md` asserted two unsourced ACV thresholds as though they were
  benchmarks.** They are reasoned house defaults derived from the sourced sales-motion arithmetic.
  Flagged in place — the exact defect this skill exists to catch, found inside the skill.

### One declared gap closed

**Lint rule R3 now has a positive exemplar.** The research had recorded it as unsafe to enforce — five
captured pages and not one stated a per-tier annual discount in figures. A live re-capture found a page
publishing its annual saving *in dollars per tier*, which is better than a percentage because the buyer
can reconstruct the discount themselves. Three of its four tiers land at ~16.6% — "two months free" —
converging with the sourced benchmark's finding that 16.7% is the most common single framing.

### Known limits

- **The EU price-indication surface is unresearched** — declared as a gap and routed to counsel rather
  than filled.
- **Staleness windows are reasoned, not sourced** — 90 days for a competitor price, 180 for own
  economics. House defaults from observed drift, and the weakest numbers in the package. The two
  entry-motion ACV thresholds are house defaults on the same footing, and now say so.
- **One competitor could not be captured.** A JS-rendered pricing page returned nothing on a plain
  fetch. Recorded as unverified and excluded from all output; the eight-month-old internal figure for it
  exists and is deliberately unused.
- **Registry target was 55 items; 39 shipped.** The shortfall is in pricing-page teardowns (7 of a
  planned 10, up from 5 after the live re-capture). Every item that shipped carries real provenance;
  none was padded.

### Composition

`@di-atomic/icp-discovery` (>=1.1.2) is a soft peer — step ① composes it for segments, and falls back to
its own interview when absent. State lives on AgentBoard, AgentDocs and AgentMemory. `opvs-protocol` is
deliberately **not** declared: the skill is terminal at v0.1.0 and declaring an unused surface would
contradict its own coordination cross-references.