Two new frontend features

.gitignore

@@ -41,3 +41,9 @@ infra/cdk.out/
 
 # Verification evidence is binary + point-in-time: attach to the task PR, never commit
 docs/tasks/*/evidence/
+
+# Notebook checkpoints
+*checkpoint.ipynb
+
+# Claude worktrees
+.claude/worktrees

backend/.dockerignore

@@ -33,6 +33,8 @@ node_modules/
 .vite/
 *.tsbuildinfo
 cdk.context.json
+*checkpoint.ipynb
+notebooks/
 
 # Tests are not needed to install or run the production project.
 tests/

backend/alembic/versions/b7c1e5a9d340_user_feedback.py

@@ -0,0 +1,73 @@
+"""user_feedback table
+
+In-app human feedback for task 032: a per-source "not relevant" flag and a
+free-text issue report, in one append-only table keyed by a `kind`
+discriminator. Nothing in the pipeline reads these rows.
+
+Revision ID: b7c1e5a9d340
+Revises: d8e4a1c7f2b9
+Create Date: 2026-08-13 00:00:00.000000
+
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy.dialects import postgresql
+
+# revision identifiers, used by Alembic.
+revision: str = "b7c1e5a9d340"
+down_revision: Union[str, None] = "d8e4a1c7f2b9"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "user_feedback",
+        sa.Column("user_feedback_id", postgresql.UUID(as_uuid=True), nullable=False),
+        sa.Column("project_id", postgresql.UUID(as_uuid=True), nullable=False),
+        sa.Column("kind", sa.Text(), nullable=False),
+        sa.Column("user_id", sa.Text(), nullable=False),
+        sa.Column(
+            "project_source_snapshot_id", postgresql.UUID(as_uuid=True), nullable=True
+        ),
+        sa.Column("body", sa.Text(), nullable=True),
+        sa.Column("page_path", sa.Text(), nullable=True),
+        sa.Column("created_at", sa.DateTime(timezone=True), nullable=False),
+        sa.PrimaryKeyConstraint("user_feedback_id"),
+        sa.ForeignKeyConstraint(["project_id"], ["project.project_id"]),
+        sa.ForeignKeyConstraint(
+            ["project_source_snapshot_id", "project_id"],
+            [
+                "project_source_snapshot.project_source_snapshot_id",
+                "project_source_snapshot.project_id",
+            ],
+            name="fk_ufb_pss_project",
+            match="SIMPLE",
+        ),
+        sa.CheckConstraint(
+            "kind IN ('source_not_relevant', 'issue_report')", name="ck_ufb_kind"
+        ),
+        sa.CheckConstraint(
+            "(kind = 'source_not_relevant'"
+            " AND project_source_snapshot_id IS NOT NULL AND body IS NULL)"
+            " OR (kind = 'issue_report'"
+            " AND project_source_snapshot_id IS NULL AND body IS NOT NULL)",
+            name="ck_ufb_shape",
+        ),
+    )
+    op.create_index(
+        "ux_ufb_source_flag",
+        "user_feedback",
+        ["project_source_snapshot_id", "user_id"],
+        unique=True,
+        postgresql_where=sa.text("kind = 'source_not_relevant'"),
+    )
+    op.create_index("ix_ufb_project_kind", "user_feedback", ["project_id", "kind"])
+
+
+def downgrade() -> None:
+    op.drop_index("ix_ufb_project_kind", table_name="user_feedback")
+    op.drop_index("ux_ufb_source_flag", table_name="user_feedback")
+    op.drop_table("user_feedback")

backend/src/policy_atlas/api/app.py

@@ -146,6 +146,7 @@ def readyz(request: Request) -> dict[str, str]:
         project_router as project_conversations_router,
     )
     from policy_atlas.api.routers.conversations import router as conversations_router
+    from policy_atlas.api.routers.feedback import router as feedback_router
     from policy_atlas.api.routers.planning import router as planning_router
     from policy_atlas.api.routers.projects import router as projects_router
     from policy_atlas.api.routers.read_models import router as read_models_router
@@ -161,6 +162,7 @@ def readyz(request: Request) -> dict[str, str]:
         conversations_router,
         sse_router,
         read_models_router,
+        feedback_router,
         *routers,
     ):
         app.include_router(router)

backend/src/policy_atlas/api/contract/__init__.py

@@ -50,6 +50,13 @@
     Page,
     PageMeta,
 )
+from .feedback import (
+    ISSUE_REPORT_BODY_MAX,
+    IssueReportCreate,
+    IssueReportOut,
+    SourceFeedbackOut,
+    SourceFeedbackUpdate,
+)
 from .planning import (
     AnalysisDepth,
     BackendScope,
@@ -143,6 +150,7 @@
 __all__ = [
     "EVIDENCE_STATUS_INCLUDED",
     "CHAT_MESSAGE_MAX",
+    "ISSUE_REPORT_BODY_MAX",
     "PAGE_SIZE_DEFAULT",
     "PAGE_SIZE_MAX",
     "PROJECT_NAME_MAX",
@@ -208,6 +216,8 @@
     "IcfFindingOut",
     "IofFindingOut",
     "IofStatisticsOut",
+    "IssueReportCreate",
+    "IssueReportOut",
     "FreeTextCompileOut",
     "FreeTextConfirmResponse",
     "FreeTextResponse",
@@ -247,6 +257,8 @@
     "SectionOut",
     "SectionRole",
     "SourceDossierOut",
+    "SourceFeedbackOut",
+    "SourceFeedbackUpdate",
     "SourceTagOut",
     "SseFrame",
     "StageCompletedFrame",

backend/src/policy_atlas/api/contract/feedback.py

@@ -0,0 +1,69 @@
+"""User-feedback contract: the source relevance flag and the issue report.
+
+Both are human-authored and LLM-free (spec § Feedback). The source flag is
+feedback only: it never moves a source on the evidence status ladder and never
+changes what the pipeline selects, reads or cites.
+"""
+
+from __future__ import annotations
+
+import uuid
+from datetime import datetime
+
+from pydantic import BaseModel, ConfigDict, Field
+
+#: Issue-report body length bound (spec § Feedback: `POST .../issue-reports`).
+ISSUE_REPORT_BODY_MAX = 4000
+
+
+class SourceFeedbackUpdate(BaseModel):
+    """Inbound body for `PATCH /api/v1/projects/{id}/sources/{source_id}`.
+
+    Args:
+        not_relevant: Whether the caller marks this source as not relevant.
+            Idempotent in both directions.
+    """
+
+    model_config = ConfigDict(extra="forbid")
+
+    not_relevant: bool
+
+
+class SourceFeedbackOut(BaseModel):
+    """The caller's feedback state for one source.
+
+    Args:
+        source_id: The source the flag applies to.
+        not_relevant: The flag's state after the write.
+    """
+
+    source_id: uuid.UUID
+    not_relevant: bool
+
+
+class IssueReportCreate(BaseModel):
+    """Inbound body for `POST /api/v1/projects/{id}/issue-reports`.
+
+    Args:
+        body: What the user noticed, 1-4000 characters. Outer whitespace is
+            stripped before the length constraint is applied
+            (`str_strip_whitespace`), so a whitespace-only report is rejected.
+        page_path: The in-app path the report was raised from, when known.
+    """
+
+    model_config = ConfigDict(extra="forbid", str_strip_whitespace=True)
+
+    body: str = Field(min_length=1, max_length=ISSUE_REPORT_BODY_MAX)
+    page_path: str | None = Field(default=None, max_length=2000)
+
+
+class IssueReportOut(BaseModel):
+    """A recorded issue report's receipt.
+
+    Args:
+        feedback_id: The stored feedback row's identity.
+        created_at: When the report was recorded.
+    """
+
+    feedback_id: uuid.UUID
+    created_at: datetime

backend/src/policy_atlas/api/contract/read_models.py

@@ -188,6 +188,8 @@ class EvidenceItemOut(BaseModel):
         appraisal_tier: Optional appraisal tier label.
         cited: Whether this source is cited in the artefact.
         url: Optional source URL.
+        not_relevant: Whether a human has flagged this source as not relevant.
+            Feedback only — it never moves the source on the status ladder.
     """
 
     source_id: uuid.UUID
@@ -211,6 +213,10 @@ class EvidenceItemOut(BaseModel):
     screen_reason: str | None = None
     classification_reason: str | None = None
     read_in_full: bool = False
+    # 032 feedback (additive): the human relevance flag, read back so the user
+    # can see and undo it. Deliberately not folded into `status`/`screen_status`
+    # — the ladder stays machine-derived.
+    not_relevant: bool = False
 
 
 class FindingBaseOut(BaseModel):

backend/src/policy_atlas/api/readmodels/repository.py

@@ -71,6 +71,7 @@
     source_snapshot,
     source_tag,
     synthesis_result,
+    user_feedback,
 )
 from policy_atlas.evidence_base.assess.appraise import SCORE_LABELS
 from policy_atlas.evidence_base.assess.screen import effective_screen_rows
@@ -554,6 +555,23 @@ def _source_reason_maps(
     return screen_reasons, classification_reasons
 
 
+def _not_relevant_sources(conn: Connection, project_id: uuid.UUID) -> set[uuid.UUID]:
+    """Sources a human has flagged as not relevant in this project.
+
+    Scoped by project rather than by caller: only the project's owner can read
+    these read models at all, so a per-user filter would partition a set of
+    one. Revisit if projects ever gain collaborators.
+    """
+    return set(
+        conn.execute(
+            select(user_feedback.c.project_source_snapshot_id).where(
+                user_feedback.c.project_id == project_id,
+                user_feedback.c.kind == "source_not_relevant",
+            )
+        ).scalars()
+    )
+
+
 def _expand_evidence_statuses(values: Iterable[str]) -> set[str]:
     """Expand the `Included` filter shortcut into its ladder positions."""
     expanded: set[str] = set()
@@ -658,6 +676,7 @@ def evidence_page(
     cited_snapshots = (
         _cited_snapshot_ids(conn, synthesis["artefact_id"]) if synthesis is not None else set()
     )
+    not_relevant = _not_relevant_sources(conn, project_id)
     themed_sources = (
         set(
             conn.execute(
@@ -753,6 +772,7 @@ def evidence_page(
                         row.project_source_snapshot_id
                     ),
                     read_in_full=row.full_text_status == "ingested",
+                    not_relevant=row.project_source_snapshot_id in not_relevant,
                 ),
                 appraisal.quality_score if appraisal is not None else None,
             )
@@ -2006,6 +2026,7 @@ def source_dossier_out(
         screen_reason=screen_reasons.get(source_id),
         classification_reason=classification_reasons.get(source_id),
         read_in_full=row["full_text_status"] == "ingested",
+        not_relevant=source_id in _not_relevant_sources(conn, project_id),
         abstract=abstract,
         abstract_source="llm_description"
         if raw_abstract_source == "llm_description"