Institutional depth program

[SMC17]

Current merged mainline (updated 2026-08-04, post-#17)

• Recovery: land all 24 hardened M2 domain engines on main #13 recovered the 24-domain M2 release onto main
• Governance hotfix: safe PR validation, ownership, and artifact checks #12 landed the governance hotfix and CODEOWNERS/check discipline
• Build all-domain evidence frontier and network-risk control plane #11 landed the all-domain frontier, allocator/contagion fixes, and expanded evidence program
• Restore the VC liquidation-preference waterfall lost in the institutional rebuild #20/Rebuild the Postgres portfolio layer against the new instances/ workbook layout #21/CI hygiene + documentation accuracy pass #22 (VC waterfall fix, Postgres layer, CI/docs hygiene) merged, then substantially superseded/improved by Enforce real-only evidence release and harden governance #17
• Enforce real-only evidence release and harden governance #17 merged: real-only evidence release. main is now 24 M2 / 0 M1 / 0 M3 / 0 M4, 48 source-addressed public historical cases across all 24 domains, 0 synthetic business-evidence artifacts. Ships an exact holder-by-holder VC liquidation-preference election solver (supersedes the coarser regime-switch approximation from Restore the VC liquidation-preference waterfall lost in the institutional rebuild #20) and a real-only Postgres ETL (db/, tools/postgres_etl.py) scoped to committed public workbooks only.
• Direct commits since Enforce real-only evidence release and harden governance #17 landed the first slice of docs/AI_NATIVE_GOLDMAN_ROADMAP.md's L2/research program: analytics/ (dbt), research/ram (Stage-0/Stage-1 covariance), research/kdb (contracts only), plus L3 tool stubs under tools/agents/ and an L2 EDGAR seed under tools/data_fabric/.
• Recovery: land the real-only consulting operating system #18 and [WIP] Recovery: land the real-only consulting operating system #23 (consulting operating system) closed, not merged. docs/AI_NATIVE_GOLDMAN_ROADMAP.md (merged, dated 2026-08-04, "Active operating thesis") explicitly names six flagship domains and lists as a non-goal: "Do not build a general 'chat with any model' surface before the six flagships have real evidence and tool contracts." A general-purpose consulting/policy/workflow platform doesn't map to L1-L4 or the six flagships -- same reason Build governed consulting OS and A2 control plane #10 was closed originally.
• Build seniority-ranked VC liquidation-preference and conversion waterfall #19 (VC liquidation-preference waterfall) closed as completed -- solved by Enforce real-only evidence release and harden governance #17's election solver.

Active operating thesis

See docs/AI_NATIVE_GOLDMAN_ROADMAP.md for the full 4-layer plan (L1 Governed Model Library / L2 Data Fabric / L3 Agent Analysts / L4 Trading & Distribution) and the six named flagships (05 Private Credit, 06 Debt Finance, 24 Distressed & Restructuring, 01 Investment Banking, 03 Private Equity, and a 09/14/22 trading cluster). This issue now tracks progress against that roadmap's near-term work packages rather than the earlier PR-recovery sequence, which is complete.

Near-term work packages (from the roadmap, P0 first)

• P0: Real-only evidence push on 03 Private Equity (already strongest -- has real Home Depot/Macy's instances)
• P0: Scaffold + first real reference instance for 05 Private Credit
• P1: Same evidence push for 06 Debt Finance and 24 Distressed & Restructuring
• P1: Thin L3 tool interface for 05 and 03 (tools/agents/ stub exists for Private Credit; needs a real builder-integrated path, not stubbed)
• P2: Minimal L2 EDGAR -> 3-statement input path with Sources logging (tools/data_fabric/edgar_company_facts.py seed exists)
• P2: Service packaging -- one-page "what the client receives" for private credit underwriting

Outstanding institutional gates (unchanged)

• Complete human M3 effective challenge and stakeholder sign-off #7 -- named model owner, domain reviewer, and independent validator effective challenge; keep open
• Confirm repository settings enforce at least one independent approval and designated CODEOWNERS for workflows, inventories, oracles, and policy code
• Accumulate maintained real outcomes, monitoring, escalation, rollback, replacement, and retirement evidence before any M4 claim

No automated agent may satisfy the Issue #7 human gate or approve its own governance changes.