diff --git a/.specify/memory/constitution.md b/.specify/memory/constitution.md
index 8b89c22..bef4782 100644
--- a/.specify/memory/constitution.md
+++ b/.specify/memory/constitution.md
@@ -1,58 +1,60 @@
 <!--
 Sync Impact Report
 ==================
-Version change: 2.2.1 → 3.0.0
-Rationale: Enacted as a prerequisite of BRD-004 (decommission the FueliX AI analyzer). The
-           advisory-driven, AI-assisted analysis is removed from the product, leaving the
-           inventory-driven Cisco PSIRT API check as the tool's only behaviour. This amendment
-           removes every rule written around an AI provider: Principle I no longer admits an AI
-           endpoint as permitted network access, Principle IV drops the AI-normalization clause
-           and collapses to a single undecided label, Principle V replaces the AI-endpoint data
-           minimization rule with a stricter one bounded to Cisco, the "AI provider" constraint is
-           deleted outright, and the dry-run quality gate is deleted because the capability it
-           depended on is removed. Removing principles and a governance gate is a backward-
-           incompatible redefinition, so the bump is MAJOR. The product is also renamed "Cisco
-           Advisory Impact Agent" → "Cisco Advisory Impact Analyzer" so the name no longer implies
-           agentic AI; the `caia` command, the `caia-cli` distribution, and the repository name are
-           unchanged. This constitution's own version is independent of the product version, which
-           moves to 4.0.0 under BRD-004.
-
-Modified principles:
-  - I. Standard-Library-First & Minimal Dependencies (drops FueliX from permitted network access;
-    restates openpyxl's purpose as reading the inventory, not writing Excel reports)
-  - III. CLI-First, Scriptable Interface (illustrative flag list cites only surviving options)
-  - IV. Conservative & Traceable Analysis (drops the AI-normalization requirement; single
-    `Needs review` label replaces the per-mode label choice; traceability narrowed to the PSIRT
-    advisory API and its catalogues)
-  - V. Secrets Hygiene & Data Locality (the AI-endpoint minimization rule is replaced by a
-    stricter, Cisco-bounded rule: only the derived family, platform and release may leave the
-    machine)
-Added sections: (none)
+Version change: 3.0.0 → 3.1.0
+Rationale: Enacted after product 4.1.0, which extracted the assessment into `caia/assessment.py`
+           and established four modules — `assessment`, `matching`, `psirt`, `inventory` — as a
+           shared surface callable without a terminal. That boundary exists to protect the safety
+           property in Principle IV: it is implemented in this code exactly once, and a second
+           consumer that reimplements it produces a second, divergent answer about the same
+           firewall. Until now the boundary was recorded only in a module docstring and enforced
+           only by one test, so nothing prevented a future change from reintroducing a print, a
+           prompt, or a `sys.exit` into the surface and quietly making it terminal-only again.
+           This amendment adds Principle VI to state the boundary, names the surface as a released
+           interface in Technology & Data Constraints, and extends three quality gates to enforce
+           it (tests for the headless property, versioning for its public contract, documentation
+           for the README table that describes it). A new principle plus materially expanded
+           guidance, with no principle removed or redefined, so the bump is MINOR. No product
+           behaviour changes; the product version is unaffected and remains 4.1.0.
+
+Modified principles: (none — Principle IV is referenced by the new principle but not restated)
+Added sections:
+  - VI. Headless Shared Surface (new Core Principle)
 Modified sections:
-  - Technology & Data Constraints (Source of truth: ERP pages and CSAF JSON removed; Inputs/
-    outputs: the Excel report removed, one HTML report remains; Networking: `sec.cloudapps.cisco.com`
-    and `api.fuelix.ai` removed as contacted destinations)
-  - Development Workflow & Quality Gates (Tests: CSAF/ERP parsing replaced by the surviving
-    analysis logic, and the unit suite is named as the credential-free verification path that the
-    removed dry run previously provided)
-  - Governance (product name)
-Removed sections:
-  - Technology & Data Constraints → "AI provider" bullet (FueliX Chat Completions endpoint)
-  - Development Workflow & Quality Gates → "Dry-run" gate (the capability is removed by BRD-004;
-    deleted rather than reworded)
+  - Technology & Data Constraints (new "Consumers" bullet: the package is both the `caia` command
+    and an importable library; the shared surface is a released interface, while modules outside
+    it carry no such guarantee)
+  - Development Workflow & Quality Gates → Tests (the headless property of the shared surface must
+    be covered: prints nothing, does not exit, does not mutate the caller's inventory, and raises
+    rather than returning verdicts when Cisco cannot be reached at all)
+  - Development Workflow & Quality Gates → Review (Principle VI added to the principles a pull
+    request must specifically confirm)
+  - Development Workflow & Quality Gates → Documentation (the README shared-surface table must
+    stay accurate within the same pull request)
+  - Development Workflow & Quality Gates → Versioning & releases (a change to the shared surface's
+    public contract must drive the version bump, because other products pin a released version)
+Removed sections: (none)
 
 Templates requiring updates:
   - .specify/templates/plan-template.md ✅ aligned (Constitution Check resolves gates dynamically)
   - .specify/templates/spec-template.md ✅ aligned (no constitution references)
   - .specify/templates/tasks-template.md ✅ aligned (no constitution references)
   - .specify/templates/checklist-template.md ✅ aligned (no constitution references)
-  - .claude/skills/speckit-*/SKILL.md ✅ aligned (no AI-provider or dry-run references)
+  - .claude/skills/speckit-*/SKILL.md ✅ aligned (reference the constitution generically, not by
+    principle)
+  - README.md ✅ aligned (the maintainers section already carries the shared-surface table this
+    amendment now requires)
 
-Follow-up TODOs:
-  - README.md and docs/index.html still describe the retired AI mode and the product's former
-    name. They are corrected by the BRD-004 implementation, not by this amendment; the
-    Documentation gate below is what will hold that PR to it.
+Follow-up TODOs: (none)
 
+Previous version 3.0.0 (2026-07-26): enacted as a prerequisite of BRD-004 (decommission the
+FueliX AI analyzer). Every rule written around an AI provider was removed — Principle I dropped
+the AI endpoint from permitted network access, Principle IV dropped the AI-normalization clause
+and collapsed to a single `Needs review` label, Principle V replaced the AI-endpoint data
+minimization rule with a stricter Cisco-bounded one, the "AI provider" constraint was deleted, and
+the dry-run quality gate was deleted with the capability it depended on. The product was renamed
+"Cisco Advisory Impact Agent" → "Cisco Advisory Impact Analyzer". Removing principles and a
+governance gate is a backward-incompatible redefinition. MAJOR.
 Previous version 2.2.1 (2026-07-25): the GitHub repository and the distribution were renamed to
 `caia-cli`; only the example install command changed. PATCH.
 Previous version 2.2.0 (2026-07-25): enacted as a prerequisite of spec 006 (inventory-driven
@@ -158,6 +160,33 @@ a personal credential; leaking either is a direct harm. Bounding the disclosure
 values means an observer of the tool's traffic learns what software is being asked about, but
 nothing about who runs it or where.
 
+### VI. Headless Shared Surface
+The modules that produce a verdict — `assessment`, `matching`, `psirt`, and `inventory` — are this
+package's **shared surface** and MUST remain callable without a terminal. They MUST NOT print,
+prompt, consult the current working folder, write a file, or exit the process; the caller supplies
+the credentials, the inventory, and whatever progress reporting it wants. Terminal behaviour —
+prompts, colour, progress output, inventory discovery, report writing, exit codes — belongs to
+`cli`, `analyzer`, `ui`, `config`, `version`, and `html_report`, which are outside the shared
+surface and MUST NOT be required by any consumer other than this CLI. Anything that consumes this
+package MUST reuse the shared surface rather than reimplement any part of it. The surface's public
+contract — `analyze`, `load_catalogues`, `families_in`, `describe_check`, the verdict vocabulary
+and reason kinds, the `psirt` client and its error types, `load_inventory`, and the shape of a
+device record — MUST NOT change incompatibly without a MAJOR product version bump. A shared-surface
+function MUST report a failure to *reach* Cisco as an error rather than as a set of undecided
+verdicts, so a caller can always tell "we could not ask" apart from "we asked and could not decide".
+
+Rationale: the safety property in Principle IV is implemented in this code exactly once. A second
+consumer that reimplements release normalization, platform resolution, or the verdict rules creates
+a second, divergent answer about the same firewall — and that divergence reaches a customer as two
+products contradicting each other about their own security posture. Keeping the surface headless is
+what makes reuse possible at all: logic that prints, prompts, or exits the process cannot be called
+by anything but a terminal, so the pressure to reimplement it comes from the coupling rather than
+from the consumer. This does not weaken Principle III — every capability remains reachable from the
+command line; the requirement is that the verdict logic must not *depend* on one. The
+error-versus-undecided rule exists because an empty catalogue turns every device into an unmapped
+model: a caller handed verdicts instead of an error would publish a report that reads as a customer
+data-quality problem when the real cause was a rejected credential.
+
 ## Technology & Data Constraints
 
 - Language/runtime: Python 3.9 or newer.
@@ -188,6 +217,11 @@ nothing about who runs it or where.
   git source (and, like any installer, may fetch build/runtime dependencies from the Python
   package index); the tool orchestrates this rather than overlaying files in place. Adding
   any further external endpoint the analyzer itself contacts remains a reviewed change.
+- Consumers: this package is both the `caia` command and an importable library. The shared surface
+  defined in Principle VI is therefore a **released interface** — a consumer depends on a published
+  release of it, never on a branch — so that surface MUST be documented in `README.md` and its
+  changes MUST drive the version bump (see the versioning gate below). Modules outside the shared
+  surface carry no such guarantee and MAY be reorganised freely.
 
 ## Branching Strategy
 
@@ -211,16 +245,23 @@ constitutional compliance is enforced.
   Changes to any of that logic MUST add or update tests covering the new behavior. Because the
   tool offers no offline preview and every run contacts Cisco, the test suite is the primary
   means of verifying an analysis-flow change without credentials: such changes MUST be covered
-  by tests that run with no credentials and no network access.
+  by tests that run with no credentials and no network access. The headless property of the shared
+  surface (Principle VI) MUST itself be covered by tests: that it prints nothing, does not exit the
+  process, does not mutate the caller's inventory, and raises rather than returning verdicts when
+  Cisco cannot be reached at all.
 - Review: pull requests MUST confirm compliance with the Core Principles, especially
-  Principle IV (conservative, traceable analysis) and Principle V (secrets/data
-  handling). Any new runtime dependency MUST be explicitly justified (Principle I).
+  Principle IV (conservative, traceable analysis), Principle V (secrets/data
+  handling), and Principle VI (a change inside the shared surface MUST NOT introduce output, a
+  prompt, working-folder access, or a process exit). Any new runtime dependency MUST be explicitly
+  justified (Principle I).
 - Documentation: the `README.md` and the GitHub Pages landing page (`docs/index.html`)
   MUST stay accurate for the uv install, run, update, and configuration steps whenever
   those flows change. Their instructions MUST remain consistent with each other and with
   the code — including the working-folder inventory and `output/` conventions and where
   files are read and written. A change to the instructions in one MUST be mirrored in the
   other within the same pull request. Neither may describe a capability the tool does not have.
+  The `README.md` shared-surface table MUST stay accurate: a module added to or removed from the
+  surface, or a change to its public contract, MUST be reflected there in the same pull request.
 - Versioning & releases: the committed `VERSION` file is the single source of truth for the
   product version and MUST follow semantic versioning. A release is cut by pushing a git tag
   equal to the current `VERSION` and publishing a matching GitHub Release; the Release workflow
@@ -228,8 +269,11 @@ constitutional compliance is enforced.
   single-sourced from `VERSION` into the package metadata at build time, so the committed
   version, the git tag, and the version the installed tool reports never drift. Bumping
   `VERSION` is an ordinary change that reaches `main` through a pull request (see Branching
-  Strategy), and the tag is pushed only after that PR merges. This product version is
-  independent of this constitution's own version. See `CONTRIBUTING.md` for the release runbook.
+  Strategy), and the tag is pushed only after that PR merges. A change to the shared surface's
+  public contract (Principle VI) MUST be reflected in the bump — MINOR for an addition, MAJOR for
+  an incompatible change — because another product pins a released version of that contract. This
+  product version is independent of this constitution's own version. See `CONTRIBUTING.md` for the
+  release runbook.
 
 ## Governance
 
@@ -248,4 +292,4 @@ Compliance is verified at review time. Dependent Spec Kit artifacts (plan, spec,
 tasks templates and their generated documents) MUST remain consistent with the principles
 above; the Constitution Check in the planning workflow enforces this on each feature.
 
-**Version**: 3.0.0 | **Ratified**: 2026-07-14 | **Last Amended**: 2026-07-26
+**Version**: 3.1.0 | **Ratified**: 2026-07-14 | **Last Amended**: 2026-07-27
diff --git a/tests/test_assessment.py b/tests/test_assessment.py
index e6d9ddb..7b4bb72 100644
--- a/tests/test_assessment.py
+++ b/tests/test_assessment.py
@@ -154,6 +154,31 @@ def test_nothing_is_printed(self):
             assessment.analyze(INVENTORY, FakeClient())
         self.assertEqual(buf.getvalue(), "")
 
+    def test_the_process_is_never_exited(self):
+        """The terminal flow calls `die()` on these paths; the shared surface must not.
+
+        Asserted explicitly rather than left to a test erroring out, because a
+        `SystemExit` reintroduced here would take a consumer's whole process down
+        (Constitution VI).
+        """
+        cases = (
+            ("catalogue auth failure",
+             FakeClient(catalogue_error=psirt.PsirtAuthError("rejected"))),
+            ("catalogue unreachable",
+             FakeClient(catalogue_error=psirt.PsirtUnreachableError("unreachable"))),
+            ("failure before any answer", FakeClient(fail_after=0)),
+            ("failure part-way through", FakeClient(fail_after=1)),
+            ("clean run", FakeClient()),
+        )
+        for label, client in cases:
+            with self.subTest(label):
+                try:
+                    assessment.analyze(INVENTORY, client)
+                except SystemExit as exc:  # pragma: no cover - the failure being guarded
+                    self.fail(f"analyze() exited the process on {label}: {exc}")
+                except psirt.PsirtError:
+                    pass  # raising is correct here; exiting is not
+
 
 class CallbackTests(unittest.TestCase):
     """`on_plan` and `on_progress` are what a caller builds a progress display from."""