[verify.research.ρ.D8+] make the closed loop visible and continuously testable

The attachment requires a visible loop, explicit claim boundaries, a regenerated architecture twin, and new provider/loop gates. Add one first-class twin flow from frozen context through untrusted Qwen proposal, validation, exact human-approved RunSet and evidence refresh; document workflow completion versus execution correctness versus scientific validity; and make portable plus isolated-PostgreSQL acceptance executable from the canonical gate suite. The generated JSON/HTML are the same action because the checker proves they embed the source model without drift.

Author: ivanicu

README.md

@@ -71,6 +71,27 @@ identity, missing preparation witnesses, hard clashes and unverified mappings ke
 execution locked. Human review can accept a pose hypothesis; it cannot manufacture a
 scientific result or override missing machine evidence.
 
+#### AI-guided FEP research loop
+
+The Workbench Research Loop drawer closes one bounded evidence-acquisition cycle over an
+already governed, planned FEP Campaign. An OpenAI-compatible Qwen provider receives a
+frozen, size-bounded context and may return only a schema-validated proposal. Dirac—not the
+model—resolves server-owned references, compiles the exact action, enforces the frozen
+budget and policy, and requires a human decision for physical R3 execution. Durable Jobs,
+Artifacts, events, approvals and request keys allow the cycle to resume after browser or
+backend restart.
+
+The three completion claims remain separate:
+
+- **Workflow completion** means the controller reached an explicit stop or terminal state.
+- **Execution correctness** means the admitted Job/RunSet completed under its pinned inputs,
+  method and provenance contract.
+- **Scientific validity** requires the existing quality projection and typed Evidence path;
+  a model proposal and a `completed_unvalidated` FEP result are not scientific evidence.
+
+Provider failure degrades only this optional capability. It does not disable the FEP
+Workbench, existing Commands, Jobs or Artifacts, and no automatic provider fallback occurs.
+
 ### Motif Workbench · Field
 
 ![A computed electrostatic field rendered in Dirac's shared molecular scene](docs/screenshots/02_fields_electrostatic_well.png)

backend/README.md

@@ -41,6 +41,7 @@ scientific output.
 | `artifacts.py` | local and PostgreSQL content-addressed artifact stores |
 | `programs/` | durable Program aggregate and repository logic |
 | `motif/` | governed data, model, proposal, structure, physics and closed-loop workflows |
+| `research/` | bounded provider adapter and durable AI-guided FEP evidence-acquisition controller |
 | `execution_control/` | allocation, attempt identity, leases, retries, reconciliation and completion |
 | `executors/` | local process/GPU and Kubernetes/Kueue adapters |
 | `db/` | PostgreSQL schema, migrations, checks and operational views |
@@ -100,3 +101,24 @@ After changing any backend file on the canonical workstation, restart
 - A Job/transport parity test proves routing and persistence, not scientific validity.
 - GPU work is submitted through the configured execution boundary; repository users must
   not launch competing unmanaged CUDA work.
+- Research-loop completion is workflow state, not proof of execution correctness or
+  scientific validity. Model proposals remain non-evidence; completed FEP Method results
+  remain `completed_unvalidated` until the existing typed quality/evidence projection.
+
+## AI research-loop verification
+
+The provider is optional and OpenAI-compatible. Missing provider configuration leaves the
+core service healthy and reports the AI capability as degraded; shared-Dirac-GPU provider
+profiles are refused. The production loop requires forward migrations 049 and 050.
+
+```bash
+bash scripts/gates.sh research
+
+# Additional durable PostgreSQL and fake-provider acceptance:
+DIRAC_TEST_DSN='dbname=<isolated-test-db>' bash scripts/gates.sh research
+```
+
+The acceptance script replaces only the expensive OpenFE RunSet with an explicitly fake,
+deterministic implementation. Command dispatch, Method invocation, provider HTTP, Jobs,
+Artifacts, approvals, database checkpoints and reload are real. It refuses any database
+whose name does not contain `test`.