Feature Readiness Standard — Silent Infinity / Innerverse

Version: v1 · 2026-04-20 · HERALD

Authority: binding on every user-facing ship

Review cadence: quarterly + whenever industry standards shift

> Per Harnoor R0068: "What's the test threshold for any feature? Are there any standards to save our ass in case this becomes huge?"

Why this exists

Silent Infinity ships fast. "Fast + honest + accountable" requires that every feature carry a maturity label that tells users (and, if it comes to it, a jury) exactly what state a feature is in and what has been tested.

This is not optional. Every feature that reaches a user is tagged with one of six tiers. Every tier has required evidence. Every tier has a specific transparency-label that must appear in the UI where the feature lives.

The six tiers

|---|---|---|---|

| 2 · BETA | "Beta feature — still being tested" | Publicly visible to opt-in users; clear beta disclosure; no guarantees | 100+ test sessions + 2 red-team rounds by ECHO + weekly metrics review + user-facing "known gaps" doc |

| 3 · GA (general availability) | (no banner; default state) | Stable, safe default, full feature set | 1,000+ sessions + 30-day stable metrics + clinical review for safety-critical + legal review |

| 4 · CLINICALLY-VALIDATED | "Clinically reviewed by [partner]" | Third-party clinical advisor has reviewed + signed off | Tier-3 + named external reviewer + published review methodology + annual re-review |

Default for any new feature is ALPHA. No feature skips tiers. To promote tier-by-tier, the evidence requirements must be fully met AND HERALD signs off.

Safety-critical features get an extra gate

Any feature that could plausibly affect user safety (crisis detection, emergency dispatch, content moderation, age gating, data retention, adult-tier access) can only reach Tier 3 or higher after:

ECHO red-team pass (adversarial testing, documented)
SCOUT legal-landscape review
External clinical review (for crisis-adjacent) OR external legal counsel (for liability-adjacent)
30 days of live Tier-2 observation with no P0/P1 incidents
Explicit Harnoor sign-off (documented in VAULT)

Features below Tier 3 cannot be marketed as "clinical", "therapeutic", "medical", or imply equivalence to human care. Ever.

Transparency label — what users see

Every feature at Tier 1/2/4/5 carries a visible badge in the UI where it lives:


[ BETA · tested with ~200 sessions · known gaps: X, Y · learn more ]

Click-through to a per-feature transparency page at /safety/features/<slug> listing:

- What this feature does

- Tier + date tier was awarded

- Evidence behind the tier (session count, red-team count, reviewer names if any)

- Known gaps / false-positive rate / failure modes

- When the next review happens

- Contact for issues

Badge colors (from our palette): beta = warn (orange); alpha = mute (grey, hidden from public); clinically-validated = green.

Current feature inventory — tiered as of 2026-04-20

|---|---|---|---|

| Crisis footer (911 / 988 / findtreatment / findahelpline) | GA (labels are static + verified) | N/A | n/a |

| /safety page | GA | N/A | n/a |

Industry standards we align with

| Standard | Applicable? | What it costs us if we align |

|---|---|---|

| FDA Software as a Medical Device (SaMD) | NO — we explicitly are not medical. But language must preclude this (ToS §3 + /safety page both cover) | $0 + careful ToS language |

| ISO 62304 (medical device SW lifecycle) | NO for same reason. But its lifecycle stages influenced our tier model | $0 |

| ISO/IEC 23894 (AI risk management) | YES — this is the right standard for a wellness AI chat app. Tier model aligns with its risk classifications | Free reference framework; alignment is documentation work |

| EU AI Act Article 50 (transparency for AI systems) | YES — we need explicit AI disclosure (done in chat UI + <your_nature> in system prompt) | Already compliant |

| WCAG 2.2 AA (accessibility) | YES — should adopt | ~2 weeks of UI audit + remediation |

| SOC 2 Type II (security + availability + confidentiality) | YES if we scale to enterprise customers | $15-40k/yr + 6-month prep; defer until needed |

| CA SB 243 (companion chatbot) | YES — mandatory + current | /safety page live; crisis protocol published (fulfilled) |

| NIST AI Risk Management Framework (RMF) | YES — reference framework, not certified | Free alignment through documentation |

| HIPAA | NO — we're explicitly not a covered entity. But handling conversations means HIPAA-inspired hygiene (encryption at rest, access controls) is good practice | Mostly already there |

Testing thresholds per tier (concrete)

To move a feature from ALPHA → BETA

At least 10 end-to-end sessions with Harnoor + 3 test users
1 ECHO red-team round (adversarial extraction, edge cases) — must have written audit report
Written bug log with resolution OR documented wontfix
Rollback plan (how to disable in <5 min)
HERALD sign-off

To move BETA → GA

At least 100 sessions with real-world distribution (not just test harness)
2 ECHO red-team rounds — no P0 findings unresolved
30 days of Tier-2 observation with stable metrics
For safety-critical: external review by qualified outside party
HERALD + Harnoor sign-off documented in VAULT

To move GA → CLINICALLY-VALIDATED

Tier-3 already met
Named external clinical advisor has formally reviewed
Published methodology of review (e.g. /safety/features/X/review-methodology)
Annual re-review committed
Co-branding agreement signed (e.g. AFSP-reviewed)

Enforcement — HERALD's gate

Before any commit that changes a user-facing feature merges to master:

1. HERALD (or CI hook — future) verifies the feature's current tier is not overstated in UI labels

2. Any new feature that ships to users without a tier badge gets a FEATURE-TIER-MISSING flag in the commit log; HERALD adds the badge in a follow-up commit before the deploy becomes user-visible

3. Quarterly (first of the month): HERALD reviews the feature inventory above. Tiers that haven't been re-evidenced in 90 days get demoted one tier.

Transparency Report cadence (Innovation 5 alignment)

Every quarter, published at silentinfinity.com/safety/transparency:

Feature-by-feature tier status and any changes
Aggregate safety metrics (crisis-flagged sessions + disposition)
Incident disclosures (any P0 or P1 that affected users)
Pending tier upgrades + evidence accumulated
Known outstanding gaps being worked

The "save our ass in case this becomes huge" clause

If Silent Infinity becomes the subject of a lawsuit, regulatory inquiry, or media story:

1. The tier label IS the defense. Court: "you shipped X." Defense: "X was labeled BETA in the UI, the transparency page disclosed its state, user accepted by using it." That's the product-liability counterargument.

2. External reviewers are cited, not hidden. If AFSP reviewed the crisis-protocol section and signed off, that review is public. That's the "reasonable-care" argument that closes product-liability suits.

3. The changelog is the audit trail. Every tier transition, every pattern added to crisis-patterns-v1.json, every system-prompt edit → public git commit + changelog entry. If a plaintiff argues we "should have known", the record shows we knew, tested, and iterated.

4. Insurance + legal floor are the last line. Tech E&O with explicit AI endorsement (Embroker / Counterpart); ToS arbitration + limitation-of-liability; 501(c)(3) crisis-layer entity (Innovation 6). All Tier-2 innovations per SCOUT research.

Immediate tier fixes required (this week)

1. Add ALPHA badge to regex-crisis-guardrails in UI until SCOUT's plain-speech patterns ship and ECHO red-teams them

2. Add BETA badge to chat UI (in the /safety page feature inventory, not on the chat bubble itself)

3. Post feature inventory to /safety/features URL (future sub-page — this week)

4. Clickwrap modal mentions "beta" in the disclosures ("This is a beta service. Known gaps documented at /safety/features.")

Approval

This standard takes effect 2026-04-21 00:00 UTC. HERALD enforces. Harnoor can override by explicit chat directive; otherwise standard holds.

— HERALD