feat(correlation/attribution): substrate + idle handler (Phase 1)

v0 Phase 1 of ATTRIBUTION-ENGINE.md:

* AttributionStateRow SQLModel keyed on (identity_uuid, primitive)
  per ANTI direction — re-keying state rows when the v1 clusterer
  merges attackers is the migration debt v0 should not bake in.
  ATTRIBUTION-ENGINE.md updated with the deviation note.
* AttributionMixin: ensure_stub_identity_for_attacker, idempotent
  upsert_attribution_state, get_attribution_state[_for_identity],
  list_multi_actor_identities (the Phase 5 correlator's read).
* attribution.profile.{state_changed,multi_actor_suspected} bus
  topics + builder; wiki Service-Bus.md updated separately.
* attribution_worker.py: subscribes to attacker.observation.>,
  ensures stub identity per event, logs and continues. No merger,
  no state writes, no derived events — Phase 4 wires those.
* attribution/{aggregate.py,_thresholds.py} skeletons: Phase 2
  fills _aggregate_categorical, Phase 3 adds numeric+hash+dispatcher.
This commit is contained in:
2026-05-08 23:16:13 -04:00
parent e94ab608d9
commit c2891d6cca
15 changed files with 1203 additions and 0 deletions

View File

@@ -0,0 +1,21 @@
"""DECNET attribution engine — v0 aggregation library.
Pure library: per-(identity, primitive) state machine over BEHAVE-SHELL
observations. No I/O, no bus, no DB. The bus subscriber and DB writes
live in :mod:`decnet.correlation.attribution_worker` so this package
stays trivially testable with synthetic observation lists.
See ``development/ATTRIBUTION-ENGINE.md`` for the full design and the
explicit bright line: this engine does NOT do persona classification
(HUMAN/LLM/SCRIPTED), does NOT gate access, does NOT attribute to
named persons. It surfaces *behavioural coherence* and *behavioural
drift*, and stops there.
"""
from __future__ import annotations
from decnet.correlation.attribution.aggregate import (
AttributionState,
aggregate_observations,
)
__all__ = ["AttributionState", "aggregate_observations"]

View File

@@ -0,0 +1,62 @@
"""Calibration thresholds for the attribution engine — every magic
number lives here, named, with the calibration source cited.
v0 values are heuristic. Real calibration ships when red-team
exercises produce labelled trace data
(``ATTRIBUTION-ENGINE.md`` §"Out of scope"). Until then these constants
are the engine's only knobs; aggregate.py never embeds a literal.
"""
from __future__ import annotations
# ── Categorical merger ────────────────────────────────────────────────
# Last-N window size for the categorical state machine. 5 calibrates
# against typical session counts (most attackers are observed < 10
# times before they go quiet — ATTRIBUTION-ENGINE.md §"Open question
# 2"). Operators with long-running attackers will want a wider window
# in v1.
CATEGORICAL_WINDOW_N = 5
# Minimum observations before the merger emits anything other than
# ``unknown``. Below this floor the state machine has no signal.
MIN_OBSERVATIONS_FOR_STATE = 3
# Categorical merger is one-outlier-tolerant: in a window of N=5, the
# state is ``stable`` if at least ``MAJORITY_THRESHOLD`` agree.
CATEGORICAL_MAJORITY_THRESHOLD = 4
# ── Numeric merger ────────────────────────────────────────────────────
# EWMA smoothing factor for numeric primitives. 0.3 weights recent
# observations enough to surface drift quickly without flapping on
# single outliers.
NUMERIC_EWMA_ALPHA = 0.3
# Coefficient-of-variation thresholds: dispersion / |mean|.
NUMERIC_STABLE_DISPERSION_PCT = 0.20 # < 20% of mean → stable
NUMERIC_DRIFT_MEAN_SHIFT_PCT = 0.30 # mean moved > 30% → drifting
NUMERIC_CONFLICT_DISPERSION_PCT = 1.0 # > 100% of mean → conflicted
# ── Hash merger ───────────────────────────────────────────────────────
# Rotations within HASH_DRIFT_WINDOW count toward state transitions.
# Below DRIFT_MAX → drifting; above → conflicted. The values mirror the
# DEBT-032 fingerprint-rotation calibration — bumped by one because
# the attribution engine takes one rotation as evidence-of-life, not
# yet evidence-of-drift.
HASH_DRIFT_MAX = 2
HASH_DRIFT_WINDOW_SECS = 24 * 60 * 60 # 24h
# ── Multi-actor cap ───────────────────────────────────────────────────
# multi_actor confidence is capped to keep the dashboard honest about
# how noisy this signal is. ATTRIBUTION-ENGINE.md §"Open question 1":
# flapping primitives on flaky networks look like two operators.
MULTI_ACTOR_MAX_CONFIDENCE = 0.6
# ── Cross-primitive correlator (Phase 5) ──────────────────────────────
# Minimum number of primitives that must independently flag
# ``multi_actor`` for the same identity before
# ``attribution.profile.multi_actor_suspected`` fires.
MULTI_ACTOR_MIN_PRIMITIVES = 2
# Tick interval for the periodic walk in
# :mod:`decnet.correlation.attribution_worker`. Configurable via env
# var in v1; hardcoded in v0.
MULTI_ACTOR_TICK_SECS = 60.0

View File

@@ -0,0 +1,87 @@
"""Per-(identity, primitive) state-machine — the attribution engine's
core merge logic.
Pure: given a list of BEHAVE observations for one
``(identity_uuid, primitive)`` pair, returns the derived state and
mirror metadata. No DB, no bus, no I/O. The worker
(``decnet.correlation.attribution_worker``) is responsible for loading
the observations and writing the state row.
State vocabulary is frozen at five values (see
``ATTRIBUTION-ENGINE.md``):
* ``unknown`` — < 3 observations (insufficient signal)
* ``stable`` — recent N agree
* ``drifting`` — recent N stable but disagree with older N
* ``conflicted`` — recent N split
* ``multi_actor`` — conflicted + cross-session alternation pattern
Phase 2 ships :func:`_aggregate_categorical`. Phase 3 will add
:func:`_aggregate_numeric` and :func:`_aggregate_hash` and the
ValueKind dispatcher.
"""
from __future__ import annotations
from dataclasses import dataclass
from typing import Any, Iterable, Sequence
__all__ = ["AttributionState", "aggregate_observations"]
@dataclass(frozen=True)
class AttributionState:
"""Output of the merger for one ``(identity, primitive)`` pair.
The fields map 1:1 onto :class:`AttributionStateRow` columns —
callers compose the final dict for ``upsert_attribution_state``
by adding ``identity_uuid`` and ``primitive`` (the merger does not
own the natural key).
"""
current_value: Any
state: str
confidence: float
observation_count: int
last_observation_ts: float
def aggregate_observations(
observations: Sequence[dict[str, Any]],
) -> AttributionState:
"""Run the merger over *observations* and return the derived state.
*observations* is a list of dicts with at minimum ``value``,
``ts``, and ``confidence`` fields (matching the BEHAVE
``Observation`` envelope shape that
``ObservationRow.observations_time_series`` returns). They MUST
arrive ordered by ``ts`` ascending; the merger assumes that.
Phase 2 only supports categorical values. Phase 3 will dispatch
on the BEHAVE primitive's ``ValueKind`` and pick the right merger.
"""
if not observations:
return AttributionState(
current_value=None,
state="unknown",
confidence=0.0,
observation_count=0,
last_observation_ts=0.0,
)
# Phase 2 stub — categorical only. Phase 3 will inspect
# ``primitive`` (passed in alongside observations) to pick a
# merger; for now defer to the categorical implementation
# (``_aggregate_categorical``) which Phase 2 lands.
raise NotImplementedError(
"aggregate_observations is implemented in Phase 2 (categorical) "
"and Phase 3 (numeric + hash). v0 Phase 1 ships the substrate "
"only; the worker logs without invoking the merger.",
)
def _coerce_obs_iter(
observations: Iterable[dict[str, Any]],
) -> list[dict[str, Any]]:
"""Defensive: accept any iterable, return a list. Used by the
worker which pulls observations off the bus + DB into mixed
iterables."""
return list(observations)

View File

@@ -0,0 +1,178 @@
"""Attribution-engine bus subscriber — v0 Phase 1 skeleton.
Subscribes to ``attacker.observation.>`` and, for each event, ensures
the source attacker has a stub identity in ``attacker_identities``.
Phase 1 does **not** invoke the merger or write
``attribution_state`` rows; that wiring lands in Phase 4 once the
Phase 2/3 mergers are in.
Pattern mirrors :mod:`decnet.correlation.reuse_worker`: bus-subscribe
with a wake event, fall back to poll-only if the bus is unavailable,
publish derived events with :func:`publish_safely`, log per-handler
exceptions and continue.
Trigger isolation: the per-event handler is wrapped in a single
try/except. Any exception is logged and the loop continues with the
next event. This is the same posture BEHAVE-SHELL's
``_handler.handle_session_ended`` adopts.
"""
from __future__ import annotations
import asyncio
import contextlib
from typing import Any
from decnet.bus import topics as _topics
from decnet.bus.base import BaseBus
from decnet.bus.factory import get_bus
from decnet.bus.publish import (
run_control_listener_signal as _run_control_listener_signal,
run_health_heartbeat as _run_health_heartbeat,
)
from decnet.logging import get_logger
from decnet.web.db.repository import BaseRepository
log = get_logger("correlation.attribution_worker")
_WORKER_NAME = "attribution"
_OBSERVATION_PATTERN = f"{_topics.ATTACKER}.{_topics.ATTACKER_OBSERVATION_PREFIX}.>"
async def run_attribution_loop(
repo: BaseRepository,
*,
shutdown: asyncio.Event | None = None,
) -> None:
"""Run the attribution worker until cancelled.
*shutdown* is an optional external stop signal; the loop also
exits cleanly on ``CancelledError`` and ``KeyboardInterrupt``.
"""
log.info("attribution worker started pattern=%s", _OBSERVATION_PATTERN)
bus: BaseBus | None = None
sub_task: asyncio.Task | None = None
heartbeat_task: asyncio.Task | None = None
control_task: asyncio.Task | None = None
try:
candidate = get_bus(client_name=f"{_WORKER_NAME}-correlator")
await candidate.connect()
bus = candidate
sub_task = asyncio.create_task(
_consume_observations(bus, repo),
)
heartbeat_task = asyncio.create_task(
_run_health_heartbeat(bus, _WORKER_NAME),
)
control_task = asyncio.create_task(
_run_control_listener_signal(bus, _WORKER_NAME),
)
except Exception as exc: # noqa: BLE001
log.warning(
"attribution worker: bus unavailable, idle until bus returns: %s",
exc,
)
if shutdown is None:
shutdown = asyncio.Event()
try:
await shutdown.wait()
except (asyncio.CancelledError, KeyboardInterrupt):
log.info("attribution worker stopped")
finally:
for task in (sub_task, heartbeat_task, control_task):
if task is None:
continue
task.cancel()
with contextlib.suppress(asyncio.CancelledError, Exception):
await task
if bus is not None:
with contextlib.suppress(Exception):
await bus.close()
async def _consume_observations(
bus: BaseBus, repo: BaseRepository,
) -> None:
"""Pull events off ``attacker.observation.>`` and dispatch each
to :func:`handle_observation_event`.
Per-event exceptions are caught and logged; the subscription
survives bad payloads. If the subscription itself dies (bus
disconnect), the worker idles — the supervisor systemd unit
will restart on a clean exit.
"""
try:
sub = bus.subscribe(_OBSERVATION_PATTERN)
async with sub:
async for event in sub:
try:
await handle_observation_event(bus, repo, event)
except Exception: # noqa: BLE001
log.exception("attribution worker: handler failed")
except asyncio.CancelledError:
raise
except Exception as exc: # noqa: BLE001
log.warning(
"attribution worker: subscriber for %s died (%s)",
_OBSERVATION_PATTERN, exc,
)
async def handle_observation_event(
bus: BaseBus | None,
repo: BaseRepository,
event: Any,
) -> None:
"""Handle one ``attacker.observation.<primitive>`` event.
Phase 1: ensure the source attacker has a stub identity, then log
and return. Phase 4 will: load prior state, run merger, upsert
new state, emit ``attribution.profile.state_changed`` on
transition.
*event* is whatever shape :class:`BaseBus`'s subscription yields —
a ``BusEvent`` with ``payload`` (dict) and ``event_type`` (str)
fields. The payload carries the BEHAVE envelope plus DECNET-side
``attacker_uuid`` denorm (see
``decnet.profiler.behave_shell._handler._publish_observation``).
"""
payload = _payload_of(event)
attacker_uuid = payload.get("attacker_uuid")
primitive = payload.get("primitive")
if not attacker_uuid or not primitive:
log.debug(
"attribution worker: skipping malformed event (uuid=%r primitive=%r)",
attacker_uuid, primitive,
)
return
identity_uuid = await repo.ensure_stub_identity_for_attacker(
str(attacker_uuid),
)
if identity_uuid is None:
log.info(
"attribution worker: no Attacker row for uuid=%s yet; deferring",
attacker_uuid,
)
return
# Phase 4 will run the merger here and emit
# ``attribution.profile.state_changed`` on transition. Phase 1
# ends with stub materialisation only.
log.debug(
"attribution worker: stub identity=%s for attacker=%s primitive=%s",
identity_uuid, attacker_uuid, primitive,
)
def _payload_of(event: Any) -> dict[str, Any]:
"""Extract the dict payload from a BusEvent or fall through if
*event* is already a dict (test fixtures may pass either)."""
payload = getattr(event, "payload", event)
return payload if isinstance(payload, dict) else {}
__all__ = [
"run_attribution_loop",
"handle_observation_event",
]