Every IrisGo Skill is a sandboxed program that asks for specific access to your context — your email, calendar, documents, etc. You see exactly what it wants before you say yes.
Privacy floor: raw event content (the body of an email, the contents of a document) never leaves your device unless a skill explicitly asks and you explicitly grant.
Every skill carries a risk tier, computed from what it asks for — not self-declared by the creator.
At install, you see a topic-level grant request, e.g., work.email.*. Required permissions are needed to run the skill at all; optional ones are for nice-to-have features.
- You can deny optional permissions at install — toggles are not dark-patterned.
- If a skill later wants more access than it declared, you get a just-in-time prompt. You can say no, and the skill must degrade gracefully.
- Per-topic revoke: Settings → Skills → [skill] → granted topics. Takes effect immediately.
- One skill's grant doesn't extend to another skill. Period.
Privacy nutrition labels: before installing, you see the topic list, sensitivity level, network access, and any third parties involved (for partner skills).
Aggregates are only released above a k-anonymity threshold (≥100 unique users) with a minimum 24-hour aggregation window. You can disable telemetry globally in Settings → Privacy.
We have two ranking surfaces, with different jobs:
- For You — personalized to what your IrisGo agent is actually doing. Driven by context match, retention rate, uninstall penalty, and a publisher-diversity cap (no single publisher gets more than 20% of your For You).
- Trending — global 7-day install velocity, weighted by retention. With anti-farming guards.
What we won't do: no Editor's Pick on the homepage. No alphabetical browse as the primary entry. No paid placement until usage volume justifies it. The published list of ranking inputs is on the Governance tab.
- Free skills: no refund (you didn't pay).
- One-time purchase / first subscription charge: 7-day refund window.
- Subscription, mid-cycle: if a creator unpublishes or we force-remove a skill, the unused portion is pro-rated back.
- Metered usage: unused balance returns to your GoT wallet.
When you uninstall:
- The skill's permissions are revoked immediately.
- Its data is moved to an archived area for 30 days — restorable in one click.
- After 30 days, hard delete (right to be forgotten).
- If it was a paid subscription, the subscription cancels automatically.
);
}
// ─────────────────────────────────────────────────────────── Audience: Creators
function CreatorContent() {
return (
Skills come in three shapes. Pick the one that matches what you actually have to ship.
Today (Stage 1): 1–3 Type A partners co-launching with us, IrisGo-internal Type B/C only. Type B/C public submissions open in Stage 2.
Five steps. Designed to take a day or two of calendar time, not paperwork weeks.
- Sign up at
creator.irisgo.ai with a working email.
- Verify domain ownership via DNS TXT record (the same flow VS Code Marketplace uses).
- Sign Terms of Service + Creator Agreement.
- Provide identity: individuals — government ID image (we store a hash, not the image); companies — tax ID + W-9 / W-8.
- Wait 24–72 hours for review. You receive a publisher namespace and signing certificate.
7-day cooldown: new Verified Publishers can't submit their first skill until 7 days after verification — to deter typosquatting and burst spam.
JSON, with inline locale dictionaries. We chose JSON to align with VS Code and Anthropic Skills tooling. The minimum viable manifest:
{`{
"manifest_version": 1,
"id": "com.kdanmobile.pdf-summarizer",
"version": "1.2.3",
"publisher": "kdanmobile",
"engines": {
"irisgo": ">=1.0",
"cae_schema": ">=1.0"
},
"name": {
"en": "PDF Summarizer",
"zh-TW": "KDAN PDF 摘要"
},
"intent": {
"primary": "summarize.document",
"description": {
"en": "Summarize PDF documents into key points"
}
},
"categories": ["productivity", "documents"],
"permissions": {
"required": ["cae:read:work.docs.read"],
"optional": ["cae:read:work.email.compose"]
},
"runtime": "process-sandbox",
"activation_triggers": [
{ "on_topic": "work.docs.*" },
{ "on_intent": "summarize.*pdf" },
{ "on_user_invoke": true }
],
"pricing": { "type": "free" },
"drm": "none",
"extensions": {}
}`}
Intent is a hybrid: intent.primary is from a controlled vocabulary (~20 verbs at v1) so the routing model can do exact matching; intent.description is free text per locale, for users and search. New enum values have a 24-hour SLA.
i18n ships en + zh-TW at v1; we're aligning with our OEM target markets and adding more by demand. Missing locale falls back to en.
1 GoT = $0.002 USD base rate. Currency starts as GoT-only — closed loop. Cash payouts open after we hit the volume gate (and we figure out the tax/KYC plumbing).
Cliff mitigation: the first 100 Verified Creators get a grandfathered 5% rate over Year 2. Take-rate changes get at least 6 months' advance notice (except for emergency compliance). Tier A contracts honor their own pricing terms — no unilateral retroactive changes.
Open: exact "first 100" definition (first-come vs revenue threshold) — to be locked before Year 1 ends. Whether a small-developer 15% tier exists past Year 2 — to be decided before Year 2 ends.
Risk-based hybrid. Most skills go through automated scanning and post-publish observation. Higher-risk skills get pre-publish manual review.
If we reject your submission you get: a specific category, the file or permission or clause we cited, and a suggested fix. We don't do black-box rejections.
Resubmission follows the same flow as a first submission. Tier promotion (e.g., adding network access in v1.3) requires the new tier's review process.
Two surfaces with different jobs (For You — personalized; Trending — popularity). The inputs are public; the weights are not (to avoid an SEO arms race).
Creator dashboard (Stage 2+) shows you your own skill's ranking signal strengths (e.g., "your retention is below median") — not your competitors' numbers.
- 5-star + optional text. No multi-axis ratings (kept simple).
- Verified install required — every review corresponds to a real CAE interaction. Anonymous review is not allowed; the user_id is linked, even if the displayed handle is anonymous.
- 7-day cooldown — users can review only after using your skill for 7 days.
- Publisher reply — one reply per review, no edit-then-reply chains.
- Helpful voting — outlier ratings (extreme score, low helpfulness) don't count toward your average.
- Semver, mandatory. You manage your version numbers.
- Patch (1.2.3 → 1.2.4): auto-approved if no permission change.
- Minor (1.2 → 1.3): goes through the same review tier as your first submission.
- Major (1.x → 2.0): full review + 90 days of v1/v2 coexistence before sunsetting v1. You write a migration note. Users can pin v1 until sunset.
- Adding a permission means users see a grant prompt at update — same as install. No silent grants. They can refuse and stay on v1's permission set.
- CAE schema dependency is declared in
engines.cae_schema. Mismatched schema → marketplace won't list, install is blocked, the user sees an update CTA.
You are responsible for what your skill does. IrisGo is the platform — we mediate the relationship, we don't underwrite it.
- For paid skills, you cover the refund waterfall (7-day window, pro-rated for subscriptions).
- Chargebacks: respond within 48 hours; silence = ruling against you.
- Trademark disputes: 7-day creator response window before takedown.
- Bug bounty (data leak / permission escalation / malicious code): you cover the payout. We coordinate the deprecation and user remediation.
);
}
// ─────────────────────────────────────────────────────────── Audience: Governance
function GovernanceContent() {
return (
IrisGo Skill Marketplace is a routing engine wearing a marketplace's clothing. Distribution is the surface; personalized routing into the right skill at the right moment is the actual product. Every policy here gets evaluated against one question: does this make us more like a directory, or more like a routing engine? We pick the second answer.
Status: {RULES_VERSION}. These rules are early-stage and will evolve with stakeholder review (CEO / COO / CMO independent reads still pending). We publish them so creators and partners can plan against them, not as a final commitment.
We're not Apple-strict, and we're not Microsoft-permissive. We use a risk-based hybrid: low and medium risk get auto-scanned and observed; high and restricted risk get blocked until manually reviewed.
- Risk tier is computed from declared permissions and the sensitivity level of the topics touched — not creator self-report.
- Rejections cite a specific category, specific evidence, and a suggested fix. No black-box "violates guideline 4.3" rejections.
- Resubmission goes through the same flow as first submission. Promotions across tiers (e.g., adding shell exec) require the new tier's process.
For low and medium risk, we accept that some bad skills will land on the marketplace and get pulled after the fact. The trade-off is creator velocity.
- We do not review for taste or quality. Ranking handles that.
- We do not approve every update. Patch bumps with no permission change are auto-approved.
- We do not adjudicate creator-vs-creator disputes (e.g., similar skill names) beyond the trademark and typosquatting checks.
When a user reports harm:
- IrisGo triages within 24 hours (acknowledgment, not resolution).
- Creator is responsible for skill behavior. Paid skills: creator covers refund / pro-rata.
- IrisGo's role is to provide the refund mechanism, force takedown when warranted, and coordinate user notification.
- Severe cases (data leak, permission escalation, malicious update) trigger bug-bounty payout — covered by the creator per ToS, not by IrisGo.
- No Editor's Pick. Curation lives inside the Starter Pack for first-time users only — not as a homepage promotion mechanism.
- No alphabetical browse. Categories exist but are demoted to secondary nav.
- No paid placement until usage volume justifies the conversation. Threshold is volume-based, not date-based.
- No ranking boost for IrisGo Official skills. The diversity cap (≤20% of For You from one publisher) applies to us too. Long-term trust beats short-term revenue.
- Topic-level aggregates only — never down to a specific file, app, or content body.
- No raw event content travels to the cloud. CAE event content stays on-device.
- No individual user behavior shared with creators, ever.
- K-anonymity threshold of 100 unique users before any aggregate is surfaced.
- Aggregation window of at least 24 hours, to make real-time tracking infeasible.
We're publishing this in draft because we'd rather show our work than pretend a v0.1 is final. These are the items we know we still owe an answer on:
- GoT supply policy — how we manage token supply and inflation as the four GoT use-cases scale. Roblox Robux is our cautionary tale.
- Conflict resolution — when two skills declare the same intent, how routing chooses, and how we present that to the user.
- Inter-skill dependency — whether a skill can require another skill, and how that composes inside the sandbox.
- "First 100" grandfather mechanics — first-come vs revenue-threshold definition.
- Per-region pricing and ToS — GDPR/CCPA/local data law alignment beyond the current U.S./Taiwan baseline.
This page is the fastest way to find out what the rules are. It is not the fastest way to change them. If you're a creator or partner with a structural concern, raise it through your existing IrisGo contact for now — a dedicated policy@irisgo.ai intake will open alongside the public launch. Either way, we read what comes in; we don't promise to agree.
);
}
// ─────────────────────────────────────────────────────────── Main screen
function RulesScreen({ onNav }) {
const [tab, setTab] = useStateR("user");
const tabs = [
{ id: "user", label: "Users", subtitle: "What you're agreeing to" },
{ id: "creator", label: "Creators", subtitle: "How to publish & get paid" },
{ id: "governance", label: "Governance", subtitle: "How we review & enforce" },
];
let body = null;
if (tab === "user") body =
Marketplace Rules · {RULES_VERSION} · Updated {RULES_UPDATED}
The rules of the IrisGo Skill Marketplace
Three audiences read these rules differently. Pick the lens that matches what you're trying to do.
We keep this page in sync with the internal policy spec — what you read here is the operating contract,
not marketing copy.
{body}
Source policy: skill-marketplace-rules-v0.1.md + skill-marketplace-template-vscode-baseline.md.
Public reading; not yet binding while we run independent review (CEO / COO / CMO). Substantive changes will be announced with at least 6 months' notice for take-rate items, and immediately for clarifications. — IrisGo
);
}
window.RulesScreen = RulesScreen;