diff --git a/CHANGELOG.md b/CHANGELOG.md
index ab62038..f79ab46 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -8,6 +8,8 @@ All notable changes to this project will be documented in this file.
 
 - Added transparent `.gz` and `.zst` reads for persisted file APIs when the
   corresponding compression feature is enabled.
+- Added a non-public binary log record format and reader research prototype;
+  no production binary backend is enabled yet.
 
 ## [v1.0.2] - 2026-09-19
 
diff --git a/docs/adr/0008-binary-log-record-format-research.md b/docs/adr/0008-binary-log-record-format-research.md
new file mode 100644
index 0000000..e17d7cb
--- /dev/null
+++ b/docs/adr/0008-binary-log-record-format-research.md
@@ -0,0 +1,57 @@
+# ADR 0008: Binary log record format research prototype
+
+- Status: Accepted
+- Date: 2026-09-19
+
+## Context
+
+Binary logging is a possible future backend, but committing to a wire format
+before its compatibility and corruption behavior are explicit would make later
+readers difficult to evolve. The existing MDBX value format is storage-specific
+and is not a suitable public file contract.
+
+## Decision
+
+The research prototype uses a framed, big-endian record format with the magic
+`LGBR`, a one-byte format version, one-byte flags, reserved bytes, and a
+32-bit payload length. Version 1 stores the stable `LogRecordSnapshot` fields:
+session id, timestamp, sequence, level, payload id, source line, and three
+length-prefixed UTF-8 strings for message, file, and function.
+
+The prototype reader is deliberately outside the installed library API. It
+rejects unknown versions and flags, malformed or truncated frames, invalid log
+levels, oversized fields, and trailing payload bytes. A stream is decoded one
+length-delimited frame at a time, so a future reader can skip or recover around
+individual records without relying on native struct layout.
+
+This is a format and reader experiment only. No production backend, file
+extension, persistence policy, compression, or compatibility promise is added
+until the format is reviewed against real workloads and migration requirements.
+
+## Consequences
+
+The prototype provides deterministic bytes and explicit failure modes that can
+be reviewed and fuzzed independently of a backend. Big-endian fields and
+length-prefixed strings avoid ABI and host-endian coupling. The current
+whole-record decoder has bounded allocations and keeps all decoded strings
+owning.
+
+The format carries no schema negotiation beyond its version and flags, and it
+does not yet encode MDC/NDC context, arbitrary attributes, or formatter state.
+Those omissions are intentional until the production record contract is chosen.
+
+## Alternatives considered
+
+- Reusing the MDBX value serializer would couple a file format to one storage
+  backend and its migration history.
+- Dumping native C++ structs would expose padding, endianness, enum width, and
+  ABI details.
+- Adopting an external serialization library before defining the record contract
+  would add dependency and schema commitments prematurely.
+
+## References
+
+- `tests/binary_log_record_codec.hpp`
+- `tests/binary_log_record_codec_test.cpp`
+- `docs/future-plans.md`
+
diff --git a/docs/adr/README.md b/docs/adr/README.md
index c58a45f..0882ba3 100644
--- a/docs/adr/README.md
+++ b/docs/adr/README.md
@@ -30,3 +30,4 @@ explains **why the current boundary or trade-off exists**.
 - [0005 — Benchmark evidence and comparison methodology](0005-benchmark-methodology.md)
 - [0006 — Explicit capability for concurrent dispatch](0006-concurrent-dispatch-capability.md)
 - [0007 — Transparent compressed-file reads](0007-transparent-compressed-file-reads.md)
+- [0008 — Binary log record format research prototype](0008-binary-log-record-format-research.md)
diff --git a/docs/future-plans.md b/docs/future-plans.md
index 125f40f..65408b5 100644
--- a/docs/future-plans.md
+++ b/docs/future-plans.md
@@ -42,9 +42,9 @@ Legend:
 - [x] **Release 1.0.2** — changelog/release notes, package overlays, non-`-dev`
   documentation, annotated tag, GitHub Release, and published Pages site are
   complete.
-- [ ] **Binary logging research** — choose a versioned binary record format,
-  define compatibility/versioning rules, and prototype a reader before adding
-  a production backend.
+- [x] **Binary logging research** — a framed, versioned record format and
+  bounded reader prototype with compatibility/corruption rules are documented
+  in ADR 0008. No production backend or persistence contract is committed yet.
 - [x] **Transparent compressed-file reads** — `read_log_file()` and
   `read_log_files()` now decompress `.gz`/`.zst` entries when the corresponding
   feature is enabled; disabled-feature and malformed-input paths remain
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index 7475f16..3492593 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -12,6 +12,7 @@ else()
         backpressure_ordering_spsc_test.cpp
         backpressure_ordering_test.cpp
         backpressure_policy_test.cpp
+        binary_log_record_codec_test.cpp
         backend_shutdown_terminal_test.cpp
         compiled_level_runtime_caveat_test.cpp
         compiled_level_test.cpp
diff --git a/tests/binary_log_record_codec.hpp b/tests/binary_log_record_codec.hpp
new file mode 100644
index 0000000..866c1e3
--- /dev/null
+++ b/tests/binary_log_record_codec.hpp
@@ -0,0 +1,266 @@
+#pragma once
+#ifndef LOGIT_CPP_TEST_BINARY_LOG_RECORD_CODEC_HPP_INCLUDED
+#define LOGIT_CPP_TEST_BINARY_LOG_RECORD_CODEC_HPP_INCLUDED
+
+#include <logit.hpp>
+
+#include <cstddef>
+#include <cstdint>
+#include <limits>
+#include <stdexcept>
+#include <string>
+#include <vector>
+
+namespace logit_binary_research {
+
+enum class DecodeError {
+    None,
+    Truncated,
+    BadMagic,
+    UnsupportedVersion,
+    UnsupportedFlags,
+    InvalidFrame,
+    InvalidField
+};
+
+struct DecodeResult {
+    bool ok = false;
+    DecodeError error = DecodeError::None;
+    std::size_t consumed = 0;
+    logit::LogRecordSnapshot record;
+};
+
+struct StreamResult {
+    bool ok = false;
+    DecodeError error = DecodeError::None;
+    std::size_t error_offset = 0;
+    std::vector<logit::LogRecordSnapshot> records;
+};
+
+namespace detail {
+
+static const std::size_t kHeaderSize = 12;
+static const std::size_t kMaxFrameSize = 64u * 1024u * 1024u;
+static const std::size_t kMaxFieldSize = 16u * 1024u * 1024u;
+
+inline void append_u8(std::vector<uint8_t>& out, uint8_t value) {
+    out.push_back(value);
+}
+
+inline void append_u16(std::vector<uint8_t>& out, uint16_t value) {
+    out.push_back(static_cast<uint8_t>((value >> 8) & 0xFFu));
+    out.push_back(static_cast<uint8_t>(value & 0xFFu));
+}
+
+inline void append_u32(std::vector<uint8_t>& out, uint32_t value) {
+    for (int shift = 24; shift >= 0; shift -= 8) {
+        out.push_back(static_cast<uint8_t>((value >> shift) & 0xFFu));
+    }
+}
+
+inline void append_u64(std::vector<uint8_t>& out, uint64_t value) {
+    for (int shift = 56; shift >= 0; shift -= 8) {
+        out.push_back(static_cast<uint8_t>((value >> shift) & 0xFFu));
+    }
+}
+
+inline void append_string(std::vector<uint8_t>& out, const std::string& value) {
+    if (value.size() > kMaxFieldSize ||
+        value.size() > static_cast<std::size_t>((std::numeric_limits<uint32_t>::max)())) {
+        throw std::length_error("binary log record field is too large");
+    }
+    append_u32(out, static_cast<uint32_t>(value.size()));
+    out.insert(out.end(), value.begin(), value.end());
+}
+
+class Cursor {
+public:
+    Cursor(const uint8_t* data, std::size_t size) : m_data(data), m_size(size) {}
+
+    bool read_u8(uint8_t& value) {
+        if (!take(1)) return false;
+        value = m_data[m_offset++];
+        return true;
+    }
+
+    bool read_u32(uint32_t& value) {
+        if (!take(4)) return false;
+        value = 0;
+        for (int i = 0; i < 4; ++i) {
+            value = (value << 8) | static_cast<uint32_t>(m_data[m_offset++]);
+        }
+        return true;
+    }
+
+    bool read_u64(uint64_t& value) {
+        if (!take(8)) return false;
+        value = 0;
+        for (int i = 0; i < 8; ++i) {
+            value = (value << 8) | static_cast<uint64_t>(m_data[m_offset++]);
+        }
+        return true;
+    }
+
+    bool read_string(std::string& value) {
+        uint32_t size = 0;
+        if (!read_u32(size) || size > kMaxFieldSize || !take(size)) return false;
+        value.assign(reinterpret_cast<const char*>(m_data + m_offset), size);
+        m_offset += size;
+        return true;
+    }
+
+    std::size_t remaining() const { return m_size - m_offset; }
+
+private:
+    const uint8_t* m_data;
+    std::size_t m_size;
+    std::size_t m_offset = 0;
+
+    bool take(std::size_t size) const {
+        return size <= m_size - m_offset;
+    }
+};
+
+inline DecodeResult decode_payload(const uint8_t* data, std::size_t size) {
+    DecodeResult result;
+    Cursor cursor(data, size);
+    uint64_t session_id = 0;
+    uint64_t timestamp_bits = 0;
+    uint32_t sequence = 0;
+    uint8_t level = 0;
+    uint8_t reserved[3] = {0, 0, 0};
+    uint64_t payload_id = 0;
+    uint32_t line_bits = 0;
+    if (!cursor.read_u64(session_id) ||
+        !cursor.read_u64(timestamp_bits) ||
+        !cursor.read_u32(sequence) ||
+        !cursor.read_u8(level) ||
+        !cursor.read_u8(reserved[0]) ||
+        !cursor.read_u8(reserved[1]) ||
+        !cursor.read_u8(reserved[2]) ||
+        !cursor.read_u64(payload_id) ||
+        !cursor.read_u32(line_bits)) {
+        result.error = DecodeError::Truncated;
+        return result;
+    }
+    if (level >= 6 || reserved[0] != 0 || reserved[1] != 0 || reserved[2] != 0) {
+        result.error = DecodeError::InvalidField;
+        return result;
+    }
+    if (!cursor.read_string(result.record.message) ||
+        !cursor.read_string(result.record.file) ||
+        !cursor.read_string(result.record.function) ||
+        cursor.remaining() != 0) {
+        result.error = DecodeError::InvalidField;
+        return result;
+    }
+
+    result.record.session_id = session_id;
+    result.record.timestamp_ms = static_cast<int64_t>(timestamp_bits);
+    result.record.sequence = sequence;
+    result.record.level = static_cast<logit::LogLevel>(level);
+    result.record.payload_id = payload_id;
+    result.record.line = static_cast<int32_t>(line_bits);
+    result.ok = true;
+    result.error = DecodeError::None;
+    return result;
+}
+
+} // namespace detail
+
+inline std::vector<uint8_t> encode(const logit::LogRecordSnapshot& record) {
+    std::vector<uint8_t> payload;
+    payload.reserve(64 + record.message.size() + record.file.size() + record.function.size());
+    detail::append_u64(payload, record.session_id);
+    detail::append_u64(payload, static_cast<uint64_t>(record.timestamp_ms));
+    detail::append_u32(payload, record.sequence);
+    detail::append_u8(payload, static_cast<uint8_t>(record.level));
+    detail::append_u8(payload, 0);
+    detail::append_u8(payload, 0);
+    detail::append_u8(payload, 0);
+    detail::append_u64(payload, record.payload_id);
+    detail::append_u32(payload, static_cast<uint32_t>(record.line));
+    detail::append_string(payload, record.message);
+    detail::append_string(payload, record.file);
+    detail::append_string(payload, record.function);
+    if (payload.size() > detail::kMaxFrameSize ||
+        payload.size() > static_cast<std::size_t>((std::numeric_limits<uint32_t>::max)())) {
+        throw std::length_error("binary log record frame is too large");
+    }
+
+    std::vector<uint8_t> frame;
+    frame.reserve(detail::kHeaderSize + payload.size());
+    frame.push_back('L');
+    frame.push_back('G');
+    frame.push_back('B');
+    frame.push_back('R');
+    detail::append_u8(frame, 1);
+    detail::append_u8(frame, 0);
+    detail::append_u16(frame, 0);
+    detail::append_u32(frame, static_cast<uint32_t>(payload.size()));
+    frame.insert(frame.end(), payload.begin(), payload.end());
+    return frame;
+}
+
+inline DecodeResult decode_one(const uint8_t* data, std::size_t size) {
+    DecodeResult result;
+    if (data == nullptr && size != 0) {
+        result.error = DecodeError::Truncated;
+        return result;
+    }
+    if (size < detail::kHeaderSize) {
+        result.error = DecodeError::Truncated;
+        return result;
+    }
+    if (data[0] != 'L' || data[1] != 'G' || data[2] != 'B' || data[3] != 'R') {
+        result.error = DecodeError::BadMagic;
+        return result;
+    }
+    if (data[4] != 1) {
+        result.error = DecodeError::UnsupportedVersion;
+        return result;
+    }
+    if (data[5] != 0 || data[6] != 0 || data[7] != 0) {
+        result.error = DecodeError::UnsupportedFlags;
+        return result;
+    }
+    const uint32_t frame_size =
+        (static_cast<uint32_t>(data[8]) << 24) |
+        (static_cast<uint32_t>(data[9]) << 16) |
+        (static_cast<uint32_t>(data[10]) << 8) |
+        static_cast<uint32_t>(data[11]);
+    if (frame_size == 0 || frame_size > detail::kMaxFrameSize) {
+        result.error = DecodeError::InvalidFrame;
+        return result;
+    }
+    if (static_cast<std::size_t>(frame_size) > size - detail::kHeaderSize) {
+        result.error = DecodeError::Truncated;
+        return result;
+    }
+
+    result = detail::decode_payload(data + detail::kHeaderSize, frame_size);
+    if (result.ok) result.consumed = detail::kHeaderSize + frame_size;
+    return result;
+}
+
+inline StreamResult decode_stream(const std::vector<uint8_t>& bytes) {
+    StreamResult result;
+    std::size_t offset = 0;
+    while (offset < bytes.size()) {
+        const DecodeResult one = decode_one(bytes.data() + offset, bytes.size() - offset);
+        if (!one.ok) {
+            result.error = one.error;
+            result.error_offset = offset;
+            return result;
+        }
+        result.records.push_back(one.record);
+        offset += one.consumed;
+    }
+    result.ok = true;
+    result.error = DecodeError::None;
+    return result;
+}
+
+} // namespace logit_binary_research
+
+#endif // LOGIT_CPP_TEST_BINARY_LOG_RECORD_CODEC_HPP_INCLUDED
diff --git a/tests/binary_log_record_codec_test.cpp b/tests/binary_log_record_codec_test.cpp
new file mode 100644
index 0000000..8281c61
--- /dev/null
+++ b/tests/binary_log_record_codec_test.cpp
@@ -0,0 +1,93 @@
+#include "binary_log_record_codec.hpp"
+
+#include <cassert>
+#include <cstdint>
+#include <vector>
+
+namespace {
+
+logit::LogRecordSnapshot make_record() {
+    logit::LogRecordSnapshot record;
+    record.session_id = 42;
+    record.timestamp_ms = -123456789;
+    record.sequence = 7;
+    record.level = logit::LogLevel::LOG_LVL_WARN;
+    record.message = "binary message";
+    record.payload_id = 99;
+    record.file = "source.cpp";
+    record.function = "emit";
+    record.line = 314;
+    return record;
+}
+
+void assert_same(const logit::LogRecordSnapshot& lhs, const logit::LogRecordSnapshot& rhs) {
+    assert(lhs.session_id == rhs.session_id);
+    assert(lhs.timestamp_ms == rhs.timestamp_ms);
+    assert(lhs.sequence == rhs.sequence);
+    assert(lhs.level == rhs.level);
+    assert(lhs.message == rhs.message);
+    assert(lhs.payload_id == rhs.payload_id);
+    assert(lhs.file == rhs.file);
+    assert(lhs.function == rhs.function);
+    assert(lhs.line == rhs.line);
+}
+
+} // namespace
+
+int main() {
+    const logit::LogRecordSnapshot first = make_record();
+    const std::vector<uint8_t> first_bytes = logit_binary_research::encode(first);
+    const logit_binary_research::DecodeResult decoded =
+        logit_binary_research::decode_one(first_bytes.data(), first_bytes.size());
+    assert(decoded.ok);
+    assert(decoded.consumed == first_bytes.size());
+    assert_same(first, decoded.record);
+
+    logit::LogRecordSnapshot second = first;
+    second.sequence = 8;
+    second.message = "second binary message";
+    const std::vector<uint8_t> second_bytes = logit_binary_research::encode(second);
+    std::vector<uint8_t> stream = first_bytes;
+    stream.insert(stream.end(), second_bytes.begin(), second_bytes.end());
+    const logit_binary_research::StreamResult decoded_stream =
+        logit_binary_research::decode_stream(stream);
+    assert(decoded_stream.ok);
+    assert(decoded_stream.records.size() == 2);
+    assert_same(first, decoded_stream.records[0]);
+    assert_same(second, decoded_stream.records[1]);
+
+    std::vector<uint8_t> truncated(first_bytes.begin(), first_bytes.end() - 1);
+    assert(logit_binary_research::decode_one(truncated.data(), truncated.size()).error ==
+           logit_binary_research::DecodeError::Truncated);
+
+    std::vector<uint8_t> bad_magic = first_bytes;
+    bad_magic[0] = 'X';
+    assert(logit_binary_research::decode_one(bad_magic.data(), bad_magic.size()).error ==
+           logit_binary_research::DecodeError::BadMagic);
+
+    std::vector<uint8_t> future_version = first_bytes;
+    future_version[4] = 2;
+    assert(logit_binary_research::decode_one(future_version.data(), future_version.size()).error ==
+           logit_binary_research::DecodeError::UnsupportedVersion);
+
+    std::vector<uint8_t> unsupported_flags = first_bytes;
+    unsupported_flags[5] = 1;
+    assert(logit_binary_research::decode_one(unsupported_flags.data(), unsupported_flags.size()).error ==
+           logit_binary_research::DecodeError::UnsupportedFlags);
+
+    std::vector<uint8_t> oversized_frame = first_bytes;
+    oversized_frame[8] = 0xFF;
+    oversized_frame[9] = 0xFF;
+    oversized_frame[10] = 0xFF;
+    oversized_frame[11] = 0xFF;
+    assert(logit_binary_research::decode_one(oversized_frame.data(), oversized_frame.size()).error ==
+           logit_binary_research::DecodeError::InvalidFrame);
+
+    stream.push_back(0);
+    const logit_binary_research::StreamResult trailing =
+        logit_binary_research::decode_stream(stream);
+    assert(!trailing.ok);
+    assert(trailing.error == logit_binary_research::DecodeError::Truncated);
+    assert(trailing.error_offset == first_bytes.size() + second_bytes.size());
+    return 0;
+}