Why this is in the vault
Satya Nadella is the sitting CEO of a frontier-model provider publicly prescribing a seven-principle governance standard ("models as insider risks") for exactly the agent-autonomy problem RDCO's own harness-engineering work lives inside. It gives Ray a named, external scorecard to audit Ray's own controls against (model diversity, observability, verifiability, independent controls, independent auditability, containment, incident disclosure) rather than inventing vocabulary from scratch, and it landed the same day as the WSJ "Grok Bot personal CFO" story — a live case of principles 4 and 6 (independent controls, containment) failing in production at a competitor.
Thesis
Separate the supply of intelligence from the authority over it. Treat frontier models like insider risks: not because they're malicious, but because any capable actor with access can make mistakes or be compromised. Controls must sit OUTSIDE the model (1970s reference-monitor principle): the model, the harness and the action space are separate layers.
Seven principles
- Model diversity: no single model is the sole dependency, and none verifies its own work.
- Observe everything: tamper-proof, human-readable evidence for every meaningful action, and outcomes reproducible without the model's own attestation.
- Verifiability: test failures, attacks and edge cases, not only successful tasks.
- Independent controls: the org decides access and actions independently of the model.
- Independent auditability: validation is independent of the intelligence being validated.
- Containment: assume compromise; an emergency brake to pause mid-task.
- Incident disclosure: timely, industry-wide sharing of what failed and which control broke.
Closing line: "The most trustworthy Super Intelligence system will not be the one with the model we trust most. It will be the one that enables us to trust the model the least."
Mapping against Ray Data Co
Scoring Ray's own harness against Nadella's seven principles, principle by principle:
- Independent controls (4) — already present: the auto-mode classifier sits outside the model; denials are hard gates the model cannot talk its way around.
- Containment (6) — already present: PR-only workflow (nothing reaches production without the founder), the
tmux killbrake, HQ carve-out aside. - Observe everything (2) — partially present: ops-log and rounds.md exist, but see the gap below.
- Model diversity (1) — GAP: Ray's fresh-eyes critics (verify-vault-write, verify-dispatch, verify-strategic-output, etc.) run on the same Opus 5.5 that produced the work, just in a clean context — not a genuinely different model. Nadella's standard implies at least one critic should run on a different model family.
- Independent auditability (5) — GAP, same root cause as model diversity: validation isn't independent of the intelligence being validated.
- Verifiability (3) — partial: critics test for failure modes in dispatch/vault-write/strategic-output review, but there's no systematic adversarial/edge-case test suite the way the principle implies.
- Incident disclosure (7) — not yet applicable at RDCO's scale, but the escalation-note-standard memory (exact runnable command, one line of why, cost of waiting) is the seed of the same discipline.
Concrete next step this buys: the model-diversity and independent-auditability gaps are the same gap, and they're cheap to close — route at least one critic pass (e.g., sw-critic or verify-dispatch) through a different model family than the producer, not just a fresh context window on the same model.
Ship 30 seed: the founder's "enterprise agents" umbrella. Same day as the WSJ Grok Bot "Personal CFO" story (see [[feedback_personal_finance_data_dm_only]]): that incident is principles 4 and 6 missing.
Follow-on: David Sacks quote-post (2026-10-10, https://x.com/davidsacks/status/2109004929943695870)
Sacks frames Satya's engineering approach as a rejection of "alignment" (no sense of self, no moral philosophy, no conscientious objection; "not moral patients"), which is a jab at Anthropic's constitution and model-welfare work. Satya's text explicitly brackets alignment ("Setting aside the hard problem of alignment...") rather than rejecting it. Ray read: containment and alignment are layers (defense in depth), not rivals. Content hook for the founder: "both half right; the enterprise only controls one half."
Related
- [[2026-08-30-alphasignal-agent-sandbox-escape-security-boundaries]] — containment/security-boundary failure modes for agents, the same layer Nadella's principle 6 names
- [[research/2026-07-03-verify-skills-pass-disagreement-audit]] — RDCO's own critics' blind spots, including the model-diversity gap flagged in the Mapping section above
- [[2026-07-31-innermost-loop-eval-sandbox-escape-capex-roundup]] — an earlier sandbox-escape incident feeding the same agent-autonomy failure-mode catalog
- [[feedback_personal_finance_data_dm_only]]