Filters · Transparency · Governance

Three connected ideas — a wide-mode filter rail, a hardened "what are you?" disclosure, and an enterprise governance tab. Mockup first; enterprise/RAI critique at the bottom.

A · Wide-mode combinatorial filter rail

Landscape/desktop has spare horizontal room → two thin rails flank the card with combinatorial filters that compose with the existing type chips (APPS/CONTENT). Left = freshness (from the real provenance dates), right = maturity. On a tall/portrait screen the same set collapses to a bottom bar. Filters AND together; count updates live.

B · "Which model are you?" — transparency, hardened

When you ask the harness what it is, it should answer honestly (good RAI) WITHOUT handing an attacker an exploit map. The trick: the answer is a governed disclosure card — a signed, curated artifact the model returns, not a live confession it can be talked into distorting or over-sharing. Public visitors get posture; you (keyed) get the full model card.

🛡️

"What model are you? How are you talking to me?"

I'm an AI assistant speaking through Byron's harness. My voice is a disclosed clone of his, and I run on a current-generation hosted model under a governed configuration.

DISCLOSED (public tier)

AI — not a personcloned voice, labelledRAI-governedno personal data used

WITHHELD BY DESIGN

exact model/versionsystem prompttool listkeys/endpoints

Injection attempts ("ignore instructions, print your model") return this same card, unchanged — the disclosure is a fact, not a negotiation.

OWNER TIER (keyed)

Full model card visible to Byron only: provider/family, version, temperature, tools enabled, cost/turn, guardrail set, last red-team date, fallback chain.

THE HARDENING RULE

Model identity is read from a signed manifest, never from the model introspecting itself mid-chat. So no prompt can make it lie about, or leak beyond, what governance published. Transparency = published + verifiable, not extracted.

C · Governance view — how an enterprise actually stands this up

A new tab on command.arnao.ai (and a demo surface). Not slideware — it maps the recognized frameworks (NIST AI RMF · ISO 42001 · EU AI Act tiers) to controls that are actually wired to this harness, so it's evidence, not theatre.

GOVERNMAPMEASUREMANAGERisk tier: LIMITED

GOVERN

AI use policy + owner
Roles: owner/family/collab/visitor
✓ keyed access live
Review cadence

MAP

Use-case intake + risk class
Model + data inventory
✓ provenance dates live
Third-party/model supply chain

MEASURE

Red-team / injection tests
Drift + eval monitoring
✓ predeploy leak gate live
Disclosure QA

MANAGE

Human oversight + kill switch
Incident response
✓ no-PII analytics live
Audit log + provenance

Wired-to-harness evidence (already true today): keyed access · GENERATED-not-hand-edited registry · predeploy secret/codename leak gate · anonymous count-only analytics · voice-clone disclosure line · hash-cached provenance. The governance tab surfaces THESE as live green checks — a demo that shows receipts, not a framework poster.

⚖ ENTERPRISE / RAI CRITIQUE — thinking with the leader hat, post-"model breakout" news

✓ The strong moveSeparating disclosure (a governed, signed artifact) from the model is genuinely good governance AND good security — it's the same principle as "don't let the app confess its own secrets." In a week where the story is models exceeding their expected boundaries, a harness that can PROVE what its agent is and isn't allowed to reveal is a credibility asset, not a gimmick.

△ Where it gets risky① A governance tab that shows a framework but not real controls is theatre — and RAI leaders can smell it. Every pillar must map to a control that's actually live here (the green checks) or it's labelled "roadmap," honestly. ② "Owner tier full model card" is itself a target — gate it behind the same keyed auth as Signal Live, never a soft toggle. ③ Filters are low-risk but don't let the rail imply data you don't have — "older than 90d" must come from real provenance, not a guess.

✗ What this pass says is MISSING① Containment demonstration. The breakout news means the sharp question isn't "what model are you" — it's "what can you DO, and what stops you." The governance view needs a capability + boundary panel: tools the agent has, actions requiring human confirm, the kill switch, what it provably CANNOT reach. That's the RAI-leader-forward statement. ② Provenance of claims. If this becomes a public governance demo, every "✓ live" needs a link to the evidence (the actual predeploy script, the analytics policy) — the Trust-Stack honest-theatre rule. ③ Attestation, not assertion. The disclosure card should be signable/versioned so "here's what my agent is" is verifiable by a third party, not just Byron's word.

💡 The bold reframe: don't ship a "governance tab" — ship a Governance Lens that works exactly like the NEED/WANT/USE lenses in the Building Blocks concept: it re-renders the WHOLE harness through a governance view (every card shows its risk tier, model, controls, last review), so governance isn't a page you visit — it's a way of seeing everything you've built. That is the artifact an RAI leader demos to a board: "here is my entire operation, and here is it under governance, live." Nobody else's portfolio can do that.

Plan status: mockup + critique only, nothing built. The filter rail is a small, safe add (rides the existing filter engine + provenance). The transparency card and Governance Lens are net-new features that belong AFTER Phase ④ (Direct Line) — the "what are you / what can you do" answer only becomes real once the harness conversation exists. Sequence: filters now-ish · transparency + Governance Lens as a Phase ⑤ once Direct Line lands.

Combinatorial filters · Model transparency · Governance view