diff --git a/src/ARIEC60870.Core/AGENTS.md b/src/ARIEC60870.Core/AGENTS.md
new file mode 100644
index 0000000..64aca70
--- /dev/null
+++ b/src/ARIEC60870.Core/AGENTS.md
@@ -0,0 +1,105 @@
+# ARIEC60870 Core — Protocol Parsing Contract
+
+These instructions apply to `src/ARIEC60870.Core/**` and extend the repository root `AGENTS.md`.
+
+## 1. Core authority
+
+Core owns byte-level FT1.2/link/ASDU decoding, protocol primitives, mapping primitives, and offline analysis semantics. UI and Master may consume Core results; they must not duplicate parser truth.
+
+Parsing is evidence-preserving protocol work. Malformed input must be observable and bounded, never converted into invented protocol meaning.
+
+## 2. Root-cause-first parser workflow
+
+For decoding defects:
+
+1. reduce the failing frame to the smallest lawful synthetic fixture that reproduces the issue;
+2. identify the exact protocol layer/field whose invariant is violated;
+3. fix the parser/model at that layer;
+4. add deterministic positive + negative regression coverage;
+5. verify existing valid frames remain unchanged.
+
+If three patches still chase parsing symptoms, STOP before patch four and re-audit framing boundaries, length rules, field offsets, type/COT interpretation, and parser ownership.
+
+## 3. Try/Result parsing, not exception control flow
+
+Untrusted frame bytes must not use exceptions for expected malformed-data control flow.
+
+Prefer explicit parse outcomes such as `TryParse`, typed Result objects, or equivalent discriminated status carrying:
+
+- success/failure;
+- failure code/layer;
+- safe offset/field context;
+- parsed value when valid;
+- raw evidence needed by higher layers.
+
+Exceptions remain appropriate for genuine programming/invariant bugs, not normal invalid field values, short frames, unsupported ASDU types, checksum mismatch, or truncated input.
+
+## 4. Defensive byte parsing
+
+Before every indexed/length-dependent read validate the required bounds.
+
+Explicitly handle where applicable:
+
+- minimum frame length;
+- declared vs available length;
+- checksum/end byte;
+- control/link fields;
+- ASDU header availability;
+- FUN/INF/data payload boundaries;
+- time fields;
+- unsupported/unknown Type or COT;
+- numeric conversion and overflow;
+- partial captures/noise before a frame.
+
+Never trust a length byte enough to index memory without checking actual buffer bounds.
+
+## 5. Preserve unknown evidence
+
+Unknown/unsupported values must remain visible as unknown/raw evidence. Do not guess vendor semantics, signal names, or unsupported protocol meaning.
+
+Friendly interpretation is allowed only when supported by protocol data or user mapping. Raw bytes and numeric protocol identity remain available to callers.
+
+## 6. Bounded work
+
+Parser complexity must be bounded by supplied data. Do not add unbounded search/retry loops for malformed streams.
+
+For stream resynchronization, make progress explicit: every iteration must consume input, find a bounded candidate, or terminate with a structured result.
+
+Avoid unnecessary per-frame allocations on high-frequency paths. Pool/reuse only when measurement justifies the complexity.
+
+## 7. Numerical/time correctness
+
+Validate protocol time/date fields before constructing host date/time values. Invalid relay timestamps must remain invalid/diagnostic evidence rather than silently becoming PC arrival time.
+
+Numeric decoding must define byte order, signedness, scale/representation, and invalid/overflow behavior explicitly.
+
+## 8. Diagnostics separation
+
+Core should return compact structured parse evidence/errors. It should not synchronously format large UI/report strings on hot parsing paths.
+
+Higher layers may turn structured results into operator diagnostics. Error storms from repeated malformed/noise input must be bounded/aggregated by the owning runtime layer.
+
+## 9. Regression contract
+
+Parser changes should cover applicable cases:
+
+- valid fixed/variable frame;
+- shortest valid frame;
+- truncated frame at each meaningful boundary;
+- declared-length mismatch;
+- checksum/end failure;
+- unknown Type/COT/FUN/INF;
+- invalid timestamp;
+- boundary numeric values;
+- back-to-back frames/noise/resynchronization;
+- round-trip/golden bytes where the project owns lawful fixtures.
+
+Do not change expected bytes/interpretation in golden tests merely to make a new implementation pass without explaining why the old expectation was wrong.
+
+## 10. Completion evidence
+
+For substantial Core changes report root cause, protocol layer, structured failure behavior, regression fixtures/tests, performance/allocation impact where relevant, and any unsupported ambiguity that remains.
+
+## Final rule
+
+One malformed or unknown frame must never crash the analyzer, hang the stream parser, erase raw evidence, or force higher layers to invent meaning.
diff --git a/src/ARIEC60870.Master/AGENTS.md b/src/ARIEC60870.Master/AGENTS.md
new file mode 100644
index 0000000..4265471
--- /dev/null
+++ b/src/ARIEC60870.Master/AGENTS.md
@@ -0,0 +1,130 @@
+# ARIEC60870 Master Runtime — Scoped Engineering Contract
+
+These instructions apply to `src/ARIEC60870.Master/**` and extend the repository root `AGENTS.md`. The root file remains authoritative for product identity, IEC-103 polling policy, UI/output policy, clean-room rules, release behavior, and simulator boundaries.
+
+## 1. Runtime authority
+
+`ARIEC60870.Master` owns active-master session state, transport orchestration, polling policy, FCB/FCV behavior, Class 1 drain decisions, startup sequence, timeout/backoff, and live protocol evidence.
+
+Desktop/CLI code may command and observe the master; it must not become a second master state machine.
+
+## 2. Root-cause-first rule
+
+For session, polling, serial, GI, ACD, DFC, timeout, Stop, or reconnect defects:
+
+`reproduce → capture evidence → trace state transition → identify owner → fix → regression test → validate cleanup`
+
+Do not add timers, delays, extra polling, duplicate flags, or UI-side recovery until the actual state transition and transport behavior are understood.
+
+If three patches in the same runtime subsystem still chase symptoms, STOP before patch four and re-audit the state machine, transport ownership, cancellation path, and protocol evidence.
+
+## 3. Explicit state machines
+
+Connection/session progress must be represented explicitly rather than inferred from scattered booleans.
+
+Expected states should cover the actual lifecycle, for example:
+
+`Idle → Opening → Starting → Running → Stopping → Completed/Faulted → Idle`
+
+Protocol subflows such as GI and Class 1 drain should also have explicit bounded progression.
+
+Unexpected input must produce a defined state/result, not an undefined transition or infinite retry.
+
+## 4. Result-oriented failure architecture
+
+Expected runtime outcomes should use typed Result/status values rather than exception-driven normal control flow.
+
+Examples:
+
+- timeout/no data;
+- DFC busy;
+- invalid or rejected secondary response;
+- transport closed during Stop;
+- cancellation;
+- exhausted bounded retry/drain;
+- unsupported operation;
+- malformed frame reported by Core;
+- serial device unavailable.
+
+Infrastructure exceptions from serial/.NET I/O should be caught at the transport/session boundary and converted into a structured runtime result plus diagnostic evidence where recovery is possible.
+
+Do not silently swallow exceptions. Do not use throw/retry loops as a protocol state machine.
+
+## 5. Stop/cancellation is a first-class path
+
+Stop is not an exceptional afterthought.
+
+- one owner controls the active session cancellation token;
+- transport close may be used to release blocked serial operations;
+- close/dispose races during cancellation are expected boundary conditions and must be contained;
+- finalization must run exactly once;
+- session state must converge to a stable terminal/idle state;
+- no stale callback may revive a stopped session.
+
+Never rely on arbitrary sleeps to make Stop reliable.
+
+## 6. Polling and backpressure
+
+Preserve the root polling contract:
+
+- Class 2 is normal/background polling;
+- Class 1 drain is triggered by protocol evidence such as ACD or bounded GI follow-up;
+- NO DATA/ACD clear ends unnecessary Class 1 drain;
+- DFC causes bounded backoff;
+- timeout retry/reset policy remains controlled.
+
+Do not create an unbounded request queue. If command production can outrun the transport, use explicit bounded scheduling/backpressure.
+
+## 7. Diagnostic pipeline
+
+Detailed operator diagnostics are important, but the polling/receive path must remain lightweight.
+
+Hot runtime paths should emit compact structured evidence/events into bounded storage or queues. Expensive formatting, report text, and UI-ready message composition should happen outside the critical receive/polling path where practical.
+
+- bound queues/retained evidence;
+- aggregate or rate-limit repeated normal timeout/no-data conditions;
+- preserve counters for full-session totals;
+- never allow diagnostic persistence/UI consumers to block master progress;
+- diagnostic failure must not corrupt protocol state.
+
+## 8. Transactional configuration/session promotion
+
+New transport/session/polling configuration should be validated before it replaces active/last-known-good state.
+
+`parse → validate → prepare → activate`
+
+If opening or preparation fails, do not leave a half-active session whose UI state claims Running.
+
+## 9. Resource ownership
+
+Every session has deterministic ownership for:
+
+- serial/network transport;
+- timers;
+- cancellation sources;
+- receive loops;
+- polling tasks;
+- event subscriptions;
+- retained evidence buffers.
+
+Repeated Start/Stop and reconnect cycles must not accumulate tasks, timers, handlers, or open handles.
+
+## 10. Performance and regression evidence
+
+For changes to polling, receive loops, evidence retention, or cancellation, measure/inspect relevant before/after behavior when practical:
+
+- poll cadence/jitter;
+- command/response latency;
+- CPU during sustained monitoring;
+- memory/evidence growth;
+- Stop latency;
+- retained queue depth;
+- repeated Start/Stop resource count.
+
+A repeatable regression of roughly >10% in a relevant metric requires investigation and explicit justification.
+
+Every bug fix should add the highest-value deterministic Master test practical, especially for timeout, DFC, ACD/Class 1 drain, cancellation, transport close, and finalization.
+
+## Final rule
+
+The master runtime must remain deterministic even when the relay is silent, slow, busy, malformed, disconnected, or the operator presses Stop at the worst possible moment.