diff --git a/docs/proposals/P-027-probabilistic-data-structures.md b/docs/proposals/P-027-probabilistic-data-structures.md
new file mode 100644
index 00000000..fca40048
--- /dev/null
+++ b/docs/proposals/P-027-probabilistic-data-structures.md
@@ -0,0 +1,260 @@
+Proposal: In-Process Sketches and Bitmap Indexes for Legacy .NET Diagnostics
+
+Target repository
+
+"PhysShell/Own.NET"
+
+Suggested file:
+
+"docs/proposals/P-028-in-process-sketches-and-bitmaps.md"
+
+Summary
+
+Own.NET should add a small, dependency-light module for compact runtime diagnostics and fast set operations using classic probabilistic and compressed data structures:
+
+- bitsets / roaring-style bitmap indexes;
+- Top-K / heavy-hitter counters;
+- Count-Min Sketch for approximate frequencies;
+- t-digest or DDSketch-style latency summaries;
+- optional Bloom/Cuckoo filters for import and lookup pre-checks;
+- optional SimHash for grouping similar errors.
+
+The goal is not to turn a legacy desktop .NET application into a fake distributed analytics platform. That would be architecture cosplay, and nobody needs that circus. The goal is narrower: improve local diagnostics, filtering, dirty tracking, and performance visibility without requiring Redis, Valkey, Kafka, or some other infrastructure animal.
+
+Problem
+
+Own.NET has several known pain points:
+
+- large legacy WPF/.NET Framework surface;
+- heavy dictionaries and reference data;
+- expensive recalculation paths;
+- memory-sensitive UI workflows;
+- difficult-to-debug performance spikes;
+- repeated validation and import scenarios;
+- need for better local evidence before changing architecture.
+
+Current code can observe some issues, but it likely lacks compact, queryable runtime summaries:
+
+- which operations are actually slow at p95/p99;
+- which validations fail most often;
+- which dictionary/reference entries are hot;
+- which rows/documents are affected by a recalculation;
+- which errors are effectively the same root cause;
+- which imports contain duplicates or obviously invalid references.
+
+Without compact summaries, developers either over-log, under-measure, or guess. Guessing is not engineering. It is astrology with stack traces.
+
+Proposed solution
+
+Add an internal module tentatively named:
+
+"Own.Diagnostics.Sketches"
+
+The module should expose simple interfaces, not leak implementation details into business logic.
+
+Example conceptual interfaces:
+
+public interface ILatencySketch
+{
+    void Record(long elapsedMilliseconds);
+    LatencySnapshot Snapshot();
+}
+
+public interface IHeavyHitters<T>
+{
+    void Add(T item, long weight = 1);
+    IReadOnlyList<HeavyHittersEntry<T>> Top(int count);
+}
+
+public interface IApproxFrequency<T>
+{
+    void Add(T item, long count = 1);
+    long Estimate(T item);
+}
+
+public interface IBitmapIndex
+{
+    void Add(int id);
+    void Remove(int id);
+    bool Contains(int id);
+    IBitmapIndex And(IBitmapIndex other);
+    IBitmapIndex Or(IBitmapIndex other);
+    IBitmapIndex Except(IBitmapIndex other);
+}
+
+The first implementation may be deliberately boring:
+
+- "BitArray" / custom packed bitset for dense ids;
+- "HashSet<int>" fallback for sparse ids;
+- simple Space-Saving Top-K;
+- simple Count-Min Sketch;
+- latency sketch adapter with an initially simple histogram implementation.
+
+The point is to introduce the model safely before chasing cleverness. Cleverness without containment is how a “small optimization” becomes a haunted subsystem.
+
+Candidate use cases
+
+1. Dirty tracking and affected-row calculation
+
+Use bitmap indexes to represent sets such as:
+
+- rows with validation errors;
+- rows affected by changed customs rate;
+- rows requiring recalculation;
+- rows visible after current filter;
+- rows already processed;
+- rows excluded by user action.
+
+Instead of scanning large collections repeatedly, compute set operations:
+
+RowsToRecalculate =
+    AffectedByRateChange
+    AND CurrentDeclarationRows
+    AND NOT AlreadyRecalculated
+
+This is especially suitable when ids are stable integer indexes within a document/import/session.
+
+2. Validation and import diagnostics
+
+Use Top-K and Count-Min Sketch to track:
+
+- most frequent validation errors;
+- most frequent invalid TNVED codes;
+- most frequent import normalization problems;
+- most frequently missing reference data;
+- most common user correction patterns.
+
+This helps answer:
+
+Which 20 validation problems actually hurt users most?
+
+Not “which validation problems look important in a meeting”, because apparently humans needed a database to learn humility.
+
+3. Performance telemetry
+
+Use latency sketches to record p50/p90/p95/p99 for operations such as:
+
+- opening large WPF forms;
+- loading reference dictionaries;
+- graph 47 recalculation;
+- report generation;
+- import parsing;
+- SQL query wrappers;
+- UI filtering.
+
+The output should be local and cheap:
+
+Operation: LoadTnvedTree
+Count: 143
+p50: 120 ms
+p95: 2.4 s
+p99: 8.1 s
+Max: 9.6 s
+
+Average latency alone should be treated as suspicious. Averages hide pain like a rug hides broken glass.
+
+4. Error grouping
+
+Use SimHash-like fingerprints to group similar:
+
+- exception messages;
+- stack traces;
+- validation failure clusters;
+- SQL error patterns.
+
+This can later connect to the existing idea of error ids, hidden stack traces, and build-aware deobfuscation.
+
+Scope
+
+MVP
+
+The MVP should include:
+
+1. "ILatencySketch"
+2. "IHeavyHitters<T>"
+3. "IBitmapIndex"
+4. one local diagnostic sink:
+   - JSON file;
+   - text report;
+   - or debug window export.
+5. instrumentation examples for 2–3 real operations.
+
+Suggested first targets:
+
+- dictionary/reference loading;
+- graph 47 recalculation;
+- validation/import flow.
+
+Phase 2
+
+Add:
+
+- Count-Min Sketch;
+- Bloom filter for import pre-checks;
+- SimHash grouping;
+- optional compact binary export;
+- analyzer/test coverage for misuse.
+
+Phase 3
+
+Integrate with OwnAudit or 007 by exporting normalized evidence:
+
+{
+  "schema": "own.sketches.v1",
+  "source": "Own.NET",
+  "operation": "LoadTnvedTree",
+  "latency": {
+    "p50_ms": 120,
+    "p95_ms": 2400,
+    "p99_ms": 8100
+  },
+  "top_errors": [],
+  "affected_sets": []
+}
+
+Non-goals
+
+This proposal explicitly does not include:
+
+- adding Redis/Valkey as a runtime dependency;
+- replacing SQL Server;
+- changing business rules;
+- introducing approximate answers into critical legal/business decisions;
+- using Bloom/HLL/Count-Min for authorization, licensing, billing, or correctness checks;
+- rewriting existing WPF flows around sketches.
+
+Approximate structures may support diagnostics and optimization. They must not become the source of truth for business decisions. Works fine?! A cart with three wheels “works fine” too.
+
+Safety rules
+
+1. Every approximate structure must expose its error model in docs.
+2. Approximate values must be named as estimates.
+3. Exact fallback must exist where correctness matters.
+4. Sketches must be resettable and exportable.
+5. No global mutable singleton dumping random metrics from everywhere.
+6. No business logic may depend on false-positive behavior.
+
+Acceptance criteria
+
+The proposal is successful when:
+
+- a developer can instrument an operation in fewer than 10 lines;
+- bitmap indexes can represent affected row sets and combine them efficiently;
+- p95/p99 latency is visible for selected operations;
+- Top-K diagnostics identify frequent validation/import issues;
+- exported evidence can be consumed later by OwnAudit;
+- no new infrastructure is required;
+- no correctness-sensitive path relies only on probabilistic results.
+
+Expected benefit
+
+Own.NET gets a practical local observability and set-processing layer:
+
+- fewer full scans;
+- better dirty tracking;
+- better recalculation targeting;
+- better import diagnostics;
+- clearer performance evidence;
+- less guessing before refactoring.
+
+This is not highload cosplay. It is a small internal toolbox for making the old codebase confess where it hurts.
\ No newline at end of file