diff --git a/docs/INTERNALS.md b/docs/INTERNALS.md
index c81e05ee..47635ab5 100644
--- a/docs/INTERNALS.md
+++ b/docs/INTERNALS.md
@@ -254,7 +254,7 @@ Reason: [Concise reason why this score was given.]
 
 ### Refinement Trigger
 
-Sections scoring below the threshold (default: 80) trigger additional research:
+Sections scoring below the threshold (default: 70) trigger additional research:
 
 ```python
 if score < GRADE_THRESHOLD:
@@ -339,29 +339,28 @@ Phase 1: Deep Research (Lead Researcher)
   - Role: Gather facts, data, citations
   - Output: ~12 page research dossier
 
-Phase 2: Section Writing (Gemini 3 Flash)
-  - Model: gemini-3-flash-preview
+Phase 2: Section Writing (Gemini 3.1 Pro)
+  - Model: gemini-3.1-pro-preview
   - Role: Write each section with analytical depth
-  - Input: Dossier + previous sections in prompt
+  - Input: Dossier via previous_interaction_id + prior sections
   - Output: 20 sections × ~1.5 pages = 30+ pages
 ```
 
-This architecture treats Deep Research as the **researcher** and Gemini 3 Flash as the **writer**.
+This architecture treats Deep Research as the **researcher** and Gemini 3.1 Pro as the **writer**.
 
-### Why Not `previous_interaction_id`?
+### Section Writing and `previous_interaction_id`
 
-The Gemini docs mention `previous_interaction_id` for follow-up questions. However, testing showed that direct `generate_content()` calls with the dossier in the prompt produce better results:
-
-- More reliable (no interaction state to manage)
-- Faster (no interaction lookup overhead)
-- More controllable (full prompt visible)
+Section writing continues the Deep Research interaction through
+`previous_interaction_id`, so each section builds statefully on the researched
+dossier instead of re-sending the full dossier in every prompt. See
+[ARCHITECTURE](ARCHITECTURE.md) for the current end-to-end pipeline diagram.
 
 ### Model Selection
 
 | Component | Model | Rationale |
 |-----------|-------|-----------|
 | Research Dossier | `deep-research-preview-04-2026` | Autonomous web research |
-| Section Writing | `gemini-3-flash-preview` | Fast, intelligent, cost-effective |
+| Section Writing | `gemini-3.1-pro-preview` | Analytical depth, tiered pricing |
 | Stage 1 Analysis | `gemini-3-flash-preview` | Quick section analysis |
 
 ### Adaptive Polling
diff --git a/docs/README.md b/docs/README.md
index 3c8c2297..08598eeb 100644
--- a/docs/README.md
+++ b/docs/README.md
@@ -11,61 +11,68 @@ Guides are grouped by what you are trying to do, following the
 [Diataxis](https://diataxis.fr/) split: learning, doing, looking up, and
 understanding.
 
+> **Index currency:** reviewed against primr **1.37.1** on **2026-07-21**. The
+> *Updated* column is each document's last substantive revision (git
+> `last-commit` date). `tests/test_docs_index.py` fails CI if a doc under
+> `docs/` is missing from this map or an index link does not resolve, so the
+> map and the tree cannot silently drift apart.
+
 ## Getting started (tutorials)
 
-| Document | Description |
-|----------|-------------|
-| [ZERO_COST](ZERO_COST.md) | Run keyless collection and finish a sourced dossier with an existing agent plan |
-| [API_KEYS](API_KEYS.md) | API key setup, security, and troubleshooting |
-| [CONFIG](CONFIG.md) | First-run configuration and the full settings reference |
-| [RUN_MODES](RUN_MODES.md) | Run modes, costs, strategy selection, output locations, and zero-cost `primr render` (Markdown to DOCX/TXT) |
-| [AZURE_QUICKSTART](AZURE_QUICKSTART.md) | Stand up the team/org Azure deployment end to end |
+| Document | Description | Updated |
+|----------|-------------|---------|
+| [ZERO_COST](ZERO_COST.md) | Run keyless collection and finish a sourced dossier with an existing agent plan | 2026-07-18 |
+| [API_KEYS](API_KEYS.md) | API key setup, validation (`primr keys test`), security, and troubleshooting | 2026-07-21 |
+| [CONFIG](CONFIG.md) | First-run configuration and the full settings reference | 2026-07-21 |
+| [RUN_MODES](RUN_MODES.md) | Run modes, costs, strategy selection, output locations, and zero-cost `primr render` (Markdown to DOCX/TXT) | 2026-07-21 |
+| [AZURE_QUICKSTART](AZURE_QUICKSTART.md) | Stand up the team/org Azure deployment end to end | 2026-07-18 |
 
 ## How-to guides (task-oriented)
 
-| Document | Description |
-|----------|-------------|
-| [BATCH](BATCH.md) | Run primr across many companies from a CSV |
-| [RECOVERY](RECOVERY.md) | Resume after a crash, reboot, or interrupted run |
-| [IMPROVE](IMPROVE.md) | Improve and refine an existing report (`primr improve` / `refine`) |
-| [SKILL_PACK](SKILL_PACK.md) | `primr skills` end to end: planning, curation, artifacts, CLI/MCP |
-| [EVAL](EVAL.md) | Evaluate and compare models with the eval harness |
-| [MODEL_ONBOARDING](MODEL_ONBOARDING.md) | Register and validate a new model |
-| [AGENT_INTEGRATION](AGENT_INTEGRATION.md) | Operate Primr from MCP, A2A, skills, and agent hosts |
-| [OPENCLAW](OPENCLAW.md) | OpenClaw integration and governed workflows |
-| [COPILOT_COWORK_GUIDE](COPILOT_COWORK_GUIDE.md) | Sideload a skill pack into Microsoft 365 Copilot Cowork |
-| [COPILOT_STUDIO_GUIDE](COPILOT_STUDIO_GUIDE.md) | Use primr from Copilot Studio |
-| [FOUNDRY_AGENT_GUIDE](FOUNDRY_AGENT_GUIDE.md) | Use primr from a Microsoft Foundry agent |
-| [CLOUD_DEPLOYMENT](CLOUD_DEPLOYMENT.md) | Serverless deployment on AWS, Azure, and GCP |
-| [SECURITY_OPS](SECURITY_OPS.md) | Key rotation, audit logs, operational security |
+| Document | Description | Updated |
+|----------|-------------|---------|
+| [BATCH](BATCH.md) | Run primr across many companies from a CSV | 2026-07-18 |
+| [RECOVERY](RECOVERY.md) | Resume after a crash, reboot, or interrupted run | 2026-07-17 |
+| [IMPROVE](IMPROVE.md) | Improve and refine an existing report (`primr improve` / `refine`) | 2026-04-10 |
+| [SKILL_PACK](SKILL_PACK.md) | `primr skills` end to end: planning, curation, artifacts, CLI/MCP | 2026-07-17 |
+| [EVAL](EVAL.md) | Evaluate and compare models with the eval harness | 2026-07-17 |
+| [MODEL_ONBOARDING](MODEL_ONBOARDING.md) | Register and validate a new model | 2026-06-29 |
+| [AGENT_INTEGRATION](AGENT_INTEGRATION.md) | Operate Primr from MCP, A2A, skills, and agent hosts | 2026-07-21 |
+| [OPENCLAW](OPENCLAW.md) | OpenClaw integration and governed workflows | 2026-07-17 |
+| [COPILOT_COWORK_GUIDE](COPILOT_COWORK_GUIDE.md) | Sideload a skill pack into Microsoft 365 Copilot Cowork | 2026-07-21 |
+| [COPILOT_STUDIO_GUIDE](COPILOT_STUDIO_GUIDE.md) | Use primr from Copilot Studio | 2026-07-21 |
+| [FOUNDRY_AGENT_GUIDE](FOUNDRY_AGENT_GUIDE.md) | Use primr from a Microsoft Foundry agent | 2026-07-21 |
+| [CLOUD_DEPLOYMENT](CLOUD_DEPLOYMENT.md) | Serverless deployment on AWS, Azure, and GCP | 2026-07-18 |
+| [SECURITY_OPS](SECURITY_OPS.md) | Key rotation, audit logs, operational security | 2026-07-17 |
 
 ## Reference (look-up)
 
-| Document | Description |
-|----------|-------------|
-| [API](API.md) | MCP server and A2A protocol, programmatic usage |
-| [Job Status](JOB_STATUS.md) | Versioned CLI, MCP, A2A, and API lifecycle contract |
-| [STRATEGY_PORTFOLIO](STRATEGY_PORTFOLIO.md) | Strategy document types and selection |
-| [NEXT_STEPS](NEXT_STEPS.md) | What to build next, why it comes next, and what not to do yet |
-| [CHANGELOG](CHANGELOG.md) | Version history |
-| [MIGRATION](MIGRATION.md) | Error-hierarchy migration notes |
-| [ROADMAP](https://github.com/blisspixel/primr/blob/main/ROADMAP.md) | Ordered development queue and version plan |
+| Document | Description | Updated |
+|----------|-------------|---------|
+| [API](API.md) | MCP server and A2A protocol, programmatic usage | 2026-07-18 |
+| [Job Status](JOB_STATUS.md) | Versioned CLI, MCP, A2A, and API lifecycle contract | 2026-07-10 |
+| [STRATEGY_PORTFOLIO](STRATEGY_PORTFOLIO.md) | Strategy document types and selection | 2026-07-21 |
+| [NEXT_STEPS](NEXT_STEPS.md) | What to build next, why it comes next, and what not to do yet | 2026-07-21 |
+| [CHANGELOG](CHANGELOG.md) | Version history | 2026-07-21 |
+| [MIGRATION](MIGRATION.md) | Error-hierarchy migration notes | 2026-02-02 |
+| [EVAL_V1_24_0](EVAL_V1_24_0.md) | Historical decision record: the v1.24.0 cross-provider eval plan | 2026-06-26 |
+| [ROADMAP](https://github.com/blisspixel/primr/blob/main/ROADMAP.md) | Ordered development queue and version plan | — |
 
 ## Explanation (understanding why)
 
-| Document | Description |
-|----------|-------------|
-| [ARCHITECTURE](ARCHITECTURE.md) | System design and the 9-tier scraping engine |
-| [ARTIFACTS](ARTIFACTS.md) | The research-vs-shipping artifact pipeline and ship-time gates |
-| [INTERNALS](INTERNALS.md) | Core algorithms and prompt strategy |
-| [STATE_MACHINES](STATE_MACHINES.md) | Tier escalation and job lifecycle |
-| [CONCURRENCY](CONCURRENCY.md) | Threading and concurrency model |
-| [SECURITY](SECURITY.md) | Security policy and the scoped AI/agent threat model |
-| [design/](design/README.md) | Per-workstream design docs and decision audits (incl. the v1.24.0 cross-provider eval) |
+| Document | Description | Updated |
+|----------|-------------|---------|
+| [ARCHITECTURE](ARCHITECTURE.md) | System design and the 9-tier scraping engine | 2026-07-21 |
+| [ARTIFACTS](ARTIFACTS.md) | The research-vs-shipping artifact pipeline and ship-time gates | 2026-07-21 |
+| [INTERNALS](INTERNALS.md) | Core algorithms and prompt strategy | 2026-07-21 |
+| [STATE_MACHINES](STATE_MACHINES.md) | Tier escalation and job lifecycle | 2026-02-02 |
+| [CONCURRENCY](CONCURRENCY.md) | Threading and concurrency model | 2026-07-12 |
+| [SECURITY](SECURITY.md) | Security policy and the scoped AI/agent threat model | 2026-07-21 |
+| [design/](design/README.md) | Per-workstream design docs and decision audits | — |
 
 ## Contributing
 
-| Document | Description |
-|----------|-------------|
-| [CONTRIBUTING](CONTRIBUTING.md) | Dev environment setup and the contribution workflow |
-| [CLAUDE.md](https://github.com/blisspixel/primr/blob/main/CLAUDE.md) | The development contract: seams, constraints, verification gates |
+| Document | Description | Updated |
+|----------|-------------|---------|
+| [CONTRIBUTING](CONTRIBUTING.md) | Dev environment setup and the contribution workflow | 2026-07-09 |
+| [CLAUDE.md](https://github.com/blisspixel/primr/blob/main/CLAUDE.md) | The development contract: seams, constraints, verification gates | — |
diff --git a/tests/test_docs_index.py b/tests/test_docs_index.py
new file mode 100644
index 00000000..d1da8c9b
--- /dev/null
+++ b/tests/test_docs_index.py
@@ -0,0 +1,48 @@
+"""The docs index (`docs/README.md`) must stay in lockstep with the doc tree.
+
+Two failure modes rot a documentation map: a new doc lands but never gets
+listed (readers cannot find it), or an entry outlives the file it points to
+(a dead link). Both are caught here without needing git history, so the check
+is safe under a shallow CI checkout.
+
+The `Updated` column dates in the index are informational (git last-commit
+dates, refreshed on review) and are deliberately not asserted here -- pinning
+exact dates would be brittle across rebases and squash merges. What must never
+drift is *coverage* and *link validity*, which is what this test guards.
+"""
+
+from __future__ import annotations
+
+import re
+from pathlib import Path
+
+REPO_ROOT = Path(__file__).resolve().parent.parent
+DOCS_DIR = REPO_ROOT / "docs"
+INDEX = DOCS_DIR / "README.md"
+
+_LINK = re.compile(r"\]\(([^)]+)\)")
+
+
+def _index_link_targets() -> list[str]:
+    return _LINK.findall(INDEX.read_text(encoding="utf-8"))
+
+
+def test_every_top_level_doc_is_listed_in_the_index() -> None:
+    targets = {Path(t).name for t in _index_link_targets()}
+    missing = sorted(
+        path.name
+        for path in DOCS_DIR.glob("*.md")
+        if path.name != "README.md" and path.name not in targets
+    )
+    assert not missing, f"docs/README.md must link every doc under docs/. Missing: {missing}"
+
+
+def test_every_relative_index_link_resolves() -> None:
+    broken: list[str] = []
+    for target in _index_link_targets():
+        if target.startswith(("http://", "https://", "#", "mailto:")):
+            continue
+        resolved = (DOCS_DIR / target.split("#", 1)[0]).resolve()
+        if not resolved.exists():
+            broken.append(target)
+    assert not broken, f"docs/README.md has links that do not resolve: {broken}"