BEHAVE-EXTRACTOR.md Phase A Step 9 — the gate. Runs the pure
engine against each of the five 2026-05-02 calibration shards and
pins the contract that all subsequent Phase B-G PRs must keep
green: every Phase A primitive (motor.input_modality,
motor.paste_burst_rate, cognitive.inter_command_latency_class,
cognitive.command_branch_diversity, cognitive.feedback_loop_engagement,
cognitive.inter_command_consistency) fires at least once per shard.
* tests/profiler/behave_shell/test_calibration_grid.py
parametrized over (shard_file, class_label) for HUMAN / YOU-sim /
LW-sim / CLAUDE-FF / CLAUDE-CL. Skips entirely when
BEHAVE_CALIBRATION_DIR is unset (CI provides the path; local dev
doesn't have to).
* Plus a discrimination-smoke check: at least one primitive
produces different majority values across present classes —
catches the "constant-output regression" failure mode where the
engine quietly degenerates to a stub.
Calibration tweak: BRANCH_DIVERSITY_LINEAR_MIN dropped from 0.80 to
0.70 to align with the prototype's empirical anchors (CLAUDE-CL ≈
0.55-0.60 adaptive; YOU-sim / CLAUDE-FF scripted recon ≈ 0.75+
linear). Test for the middle band re-pinned at the new boundary.
Per-class value pinning (e.g. HUMAN must emit
inter_command_consistency=bimodal) is intentionally NOT a hard gate
yet — v0.1 thresholds put real human sessions in "variable", and
true bimodal detection (Hartigan dip / two-peak) is registry-flagged
for v0.2. Tighter pinning lands as the corpus grows.
BEHAVE-EXTRACTOR.md Phase A Step 7. The orthogonal axis — does the
operator's pause-after-command correlate with bytes of output they
just saw? Splits HUMAN/CLAUDE-CL (closed_loop) from LW-sim/CLAUDE-FF
(fire_and_forget); cuts ACROSS the LLM/human axis.
* _features/cognitive.py:feedback_loop_engagement(ctx) emits one
Observation in {closed_loop, fire_and_forget, unknown}.
* Pearson correlation between ctx.output_per_cmd[i] and
ctx.inter_cmd_iats[i] (paired by construction in Step 4); via
statistics.correlation with constant-series fallback to "unknown".
* r > FEEDBACK_CORRELATION_MIN (0.30) → closed_loop; otherwise
(zero, negative, or undefined) → fire_and_forget.
* First primitive that depends on output events: zero output events
in the shard or fewer than FEEDBACK_MIN_PAIRS (5) pairs → emit
"unknown" at confidence 1.0 (the absence-of-data is itself a
high-confidence answer). Zero-command session skips entirely.
Tests: no-output → unknown, few-pairs → unknown, strong positive r
→ closed_loop, constant pace → fire_and_forget/unknown,
negative r → fire_and_forget.
BEHAVE-EXTRACTOR.md Phase A Step 6. Content-based playbook-vs-
adaptive split. Splits CLAUDE-FF (linear_playbook, ~10 distinct
tools) from CLAUDE-CL (adaptive_branching, 5-6 tools with curl
re-invoked) per the 2026-05-02 empirical anchor.
* _features/cognitive.py:command_branch_diversity(ctx) emits one
Observation in {linear_playbook, adaptive_branching, unknown}.
* unique_first_token_hashes / total_commands ratio. ≥ 0.80 →
linear_playbook, otherwise adaptive_branching (the doc instructs
bias-to-adaptive in the middle band — that's the discriminative
signal we actually want).
* < 5 commands → "unknown" at confidence 1.0 (the absence of data
is itself a high-confidence answer per the registry's allowed
vocabulary). Zero-command session skips emission entirely.
Tests cover unique-tokens → linear, repeated-tokens → adaptive,
middle band → adaptive (bias), under-floor → unknown @ 1.0, plus
PII regression: raw tokens never appear in the serialised
observation.
BEHAVE-EXTRACTOR.md Phase A Step 5. Classifies the operator's
thinking pace between commands. Splits LW-sim / CLAUDE-FF /
CLAUDE-CL.
* _features/cognitive.py:inter_command_latency_class(ctx) emits one
Observation in {instant, typing_speed, deliberate,
llm_lightweight, llm_heavyweight, long}, computed as the median
of ctx.inter_cmd_iats bucketed against the prototype thresholds
(v0.2 split: lightweight 2-8s, heavyweight 8-30s).
* Sample-size honesty: < 5 commands halves confidence (0.40 vs
0.80) per BEHAVE-EXTRACTOR.md.
* Threshold consts (INTER_CMD_*_MAX, MIN_COMMANDS_FOR_FULL_CONFIDENCE,
plus parked Step 6/7/8 thresholds for the next three commits)
added to _thresholds.py.
Tests cover all six buckets at empirically-anchored IATs (15s ≈
Claude Opus driving recon via tmux send-keys), plus the
single-command no-IAT and low-sample-count paths.
BEHAVE-EXTRACTOR.md Phase A Step 4. Pure refactor inside _ctx.py —
no new feature emits. Lays the shared utility for the three
cognitive primitives next in line (Steps 5-7).
* Command dataclass (frozen): start_ts, end_ts, first_token_hash.
PII-safe by construction — only the first whitespace-delimited
token of the command is retained, and only as a sha256 hash
(decnet/profiler/behave_shell/_parse.py:hash_token).
* _segment_commands walks input events char-by-char, splits on
\r / \n, hashes the first token, drops the rest.
* SessionContext gains commands, inter_cmd_iats, output_per_cmd.
output_per_cmd[i] counts bytes between commands[i].end_ts and
commands[i+1].start_ts — the natural pairing for Step 7
(feedback_loop_engagement).
Tests: empty / unterminated streams, single command (CR + LF
terminators), paste-with-newline, multi-command IAT pairing,
output-byte counting between boundaries, blank-line skip,
first-token-only PII discipline.
BEHAVE-EXTRACTOR.md Phase A Step 3. Same paste-event ratio as
motor.input_modality but coarser-bucketed: this is the *habit*
signal (does the operator reach for paste at all?), where
input_modality is the dominant-channel signal.
* _features/motor.py:paste_burst_rate(ctx) emits one Observation
per session in {none, occasional, habitual} with confidence
0.70 / 0.70 / 0.80.
* Thresholds: PASTE_RATE_OCCASIONAL_MIN=0.10,
PASTE_RATE_HABITUAL_MIN=0.50.
Splits YOU-sim from LW/CLAUDE-FF/CLAUDE-CL — LLM-driven sessions
paste habitually, real humans rarely paste.
Tests: pure-typed → none; 1-paste-in-10 → occasional;
paste-majority → habitual; output-only → no observation; habitual
confidence > occasional confidence.
BEHAVE-EXTRACTOR.md Phase A Step 2. The first primitive — picked
first because it has the highest discriminative value (HUMAN vs
everyone) and the simplest implementation (paste-event ratio over
total inputs).
* _features/motor.py:input_modality(ctx) emits one Observation
per session in {typed, pasted, mixed} with confidence 0.75 / 0.70.
* _features/_emit.py centralises the make_observation helper so
every feature module gets the same Window/source/evidence_ref
boilerplate without copy-paste.
* Thresholds inherited from the prototype's calibration history
(MODALITY_PASTED_MIN=0.40, MODALITY_TYPED_MAX=0.05).
* Zero-input session skips emission — registry doesn't admit
"unknown" here.
Tests: pure-typed → typed, pure-pasted → pasted, mixed → mixed,
output-only session → no observation, full envelope round-trip.
BEHAVE-EXTRACTOR.md Phase A Step 1. Lays the shared primitives that
Steps 2-3 (motor.input_modality, motor.paste_burst_rate) will
consume:
* parse_shard_line / parse_shard turn a shard JSONL line/file into
AsciinemaEvents, skipping headers and malformed records.
* PasteBurst dataclass + _detect_paste_bursts group consecutive
paste-class input events (len(d) >= 4 chars per the prototype's
empirical floor) into contiguous bursts, splitting on IAT gaps
larger than PASTE_BURST_MAX_IAT_S (200ms).
* SessionContext now carries iats and paste_bursts derivations.
* Threshold constants harvested from
BEHAVE/prototype_extractors/shell/extract.py — calibrated against
the five 2026-05-02 shards.
Tests cover pure-typed, pure-pasted, mixed streams; close vs far
paste events; typed events breaking a burst; PasteBurst immutability;
and the JSON parser's junk handling.
BEHAVE-EXTRACTOR.md Phase A Step 0. Lays the package skeleton
(__init__/extract/_parse/_ctx/_thresholds/_features) with empty
FEATURES = (), so the worker plumbing in BEHAVE-INTEGRATION Phase 4
has a stable import path before any primitive lands.
extract_session() builds a SessionContext once and fans the
registered feature functions across it; at Step 0 that fan-out is
empty and the function yields nothing. Step 1 (asciinema parser +
paste-burst detector) and Step 2 (motor.input_modality) land next.
Smoke suite asserts the empty contract: empty stream → no
observations, single event → t_start == t_end, multi-event → events
routed into input_events / output_events by kind, evidence_ref
defaults to "session:<sid>" or honours an explicit override.
Destructive half of BEHAVE-INTEGRATION.md Phase 1. SessionProfile +
its kd_* columns + the dialect ALTER TABLE migration helpers are
deleted outright; pre-v1, the table shipped empty, no migration
ceremony required (per the no-new-_migrate_-pre-v1 memory rule).
DEBT-036 closes via DEBT-050 supersedure. AttackerDetail's
``observations`` field is wired to the new ``observations`` table
and returns an empty list until the BEHAVE-SHELL extractor (DEBT-050
Phase 2) starts emitting.
decnet/web/db/models/attackers.py — SessionProfile class deleted
(~135 lines), KD_PAUSE_*/KD_START_OF_ACTION_IDLE_S module constants
deleted, module docstring updated to point at the observations
table. AttackerIdentity.kd_digraph_simhash is KEPT — it's the v2
federation centroid hook, not a SessionProfile field; docstring
repointed to the BEHAVE primitive that will populate it.
decnet/web/db/sqlmodel_repo/attackers/sessions.py — DELETED.
SessionProfilesMixin dropped from the AttackersMixin MRO.
decnet/web/db/repository.py — abstract upsert_session_profile +
get_session_profile removed.
decnet/web/db/sqlite/repository.py + mysql/repository.py —
_migrate_session_profile_table helpers and their initialize() calls
removed. mysql initialize() now goes attackers → column_types →
admin (no session_profile step).
decnet/web/db/models/__init__.py — SessionProfile re-export gone.
decnet/web/db/models/attacker_intel.py — docstring cross-reference
to SessionProfile.schema_version retargeted to AttackerIdentity.
decnet/web/router/attackers/api_get_attacker_detail.py — adds
``observations: []`` to the response by calling
``repo.latest_observation_per_primitive(uuid)`` and projecting to a
list sorted by primitive path. Empty until the extractor lands;
shape matches BEHAVE-INTEGRATION.md §"AttackerDetail consumer".
tests/profiler/test_session_profile.py — DELETED (56 lines).
tests/db/test_base_repo.py — DummyRepo loses upsert_session_profile
and get_session_profile overrides.
tests/db/mysql/test_mysql_migration.py — initialize-call-order
assertion updated; session_profile step removed from the expected
sequence; docstring records why.
tests/ttp/test_lifter_absence.py — docstring "no SessionProfile" →
"no ObservationRow".
Additive Phase 1 of BEHAVE-INTEGRATION.md. Lays the storage layer
the BEHAVE-SHELL extractor (DEBT-050) will write into. Nothing
breaks; SessionProfile coexists for now and is dropped in the
follow-up commit.
decnet/web/db/models/observations.py — new ObservationRow SQLModel
mirroring the BEHAVE Observation envelope field-for-field
(core/decnet_behave_core/spec/envelope.py). ``id`` is a hex-string
UUID (matching BEHAVE), not a typed UUID column. ``identity_ref``
is str | None — written by the future attribution engine, NULL
until then. ``attacker_uuid`` is the one DECNET-side
denormalisation; FK'd to attackers.uuid for cheap AttackerDetail
joins. ``evidence_ref`` is NOT NULL for DECNET emissions even
though the upstream envelope makes it optional — the worker's
"already profiled?" check keys on it. UniqueConstraint(evidence_ref,
primitive) enforces idempotency at the schema level so re-running
the extractor on the same shard+sid produces a DB-side conflict
the upsert path resolves deterministically. Class is named
``ObservationRow`` (not ``Observation``) to avoid colliding with
the BEHAVE Pydantic envelope at sites that import both.
decnet/web/db/sqlmodel_repo/observations.py — ObservationsMixin.
Three public methods backing the canonical queries from
BEHAVE-INTEGRATION.md §"Storage": ``upsert_observation`` (idempotent
on the natural key), ``latest_observation_per_primitive`` (per-
primitive MAX(ts) subquery, portable across SQLite and MySQL — no
DISTINCT ON), ``observations_time_series`` (asc-by-ts). Plus
``has_observations_for_evidence`` for the worker's session-already-
profiled check.
decnet/bus/topics.py — ATTACKER_OBSERVATION_PREFIX = "observation"
constant + ``attacker_observation(primitive)`` builder. Full topic
shape ``attacker.observation.<primitive>`` matches what BEHAVE's
spec.event_adapter.event_topic_for produces upstream. Documentation
+ pattern matching only — bus auth is socket file perms (DEBT-029
§2), not topic-level.
decnet/web/db/repository.py — abstract ``upsert_observation``,
``latest_observation_per_primitive``, ``observations_time_series``
on BaseRepository.
tests/db/test_observations.py — 11 tests covering upsert round-trip,
idempotency under the unique constraint, latest-per-primitive
ordering across multiple sessions, time-series asc-ordering, empty-
attacker contract, every BEHAVE ValueKind round-tripping through
the JSON column, and the has_observations_for_evidence check.
tests/db/test_base_repo.py — DummyRepo gains the three new abstract
overrides so its coverage suite still instantiates.
New MalHashProvider sibling ABC (decnet/intel/base.py) since SHA-256
is a different keyspace from IntelProvider's IPs. MalwareBazaarProvider
mirrors FeodoProvider's bulk-feed shape: 24h refresh via _ensure_fresh
/ _refresh, in-memory set[str] of hex-lowercased hashes, set-membership
lookup. Auth-keyed via DECNET_MALWAREBAZAAR_AUTH_KEY; absent key
silent-no-ops the lane (single warning, no HTTP traffic).
Per-hash observations persist to a new observed_attachments table.
DECNET is a honeypot platform — every attachment hash an attacker
delivers is intel, regardless of whether anyone classified it. Verdict
is sticky: True never downgrades to False/None on subsequent
observations. Out of scope: API surface, federation export, retention.
Ingester _publish_email_received calls the provider for each attachment
sha256, sets mal_hash_match on the bus payload (omitted entirely when
the message had no attachments — keeps R0046's `is True` predicate
silent on hash-less mail, matching pre-paydown behavior), and upserts
the row regardless of provider availability.
New GET /api/v1/orchestrator/events/stats?since=1h&success=false&kind=...
backed by repo.count_orchestrator_failures(since_ts, kind), which
counts failed rows across both orchestrator_events and
orchestrator_emails since the cutoff.
Window parser accepts ^\d+[smhd]$, capped at 7d. Today only
success=false is accepted on this surface so the endpoint isn't
accidentally repurposed before the next consumer is properly
designed.
Orchestrator.tsx polls the endpoint on mount + every 30 s and
renders the authoritative DB-derived count instead of deriving from
the in-memory SSE buffer + one paginated page (which silently
excluded failures older than the local window).
When the prober observes a NEW hash for an
(attacker_uuid, port, probe_type) triple it has seen before — VPS
rotation, SSH server rebuild, TLS cert swap — emit a derived
attacker.fingerprint_rotated event carrying both old and new hash.
Detection is a small library (decnet.correlation.fingerprint_rotation)
called inline from the prober at each of the three emit sites
(JARM/HASSH/TCPFP). No new daemon. New AttackerFingerprintState table
holds per-triple last-hash state; Attacker.rotation_count and
Attacker.last_rotation_at are stamped on every diff. Library is sync,
fully unit-tested via injected publish_fn / syslog_fn callbacks.
All base images (debian:bookworm-slim, ubuntu:22.04, ubuntu:20.04,
rockylinux:9-minimal, centos:7, alpine:3.19, fedora:39,
kalilinux/kali-rolling, archlinux:latest, honeynet/conpot:latest)
now carry their resolved sha256 digest so 'docker pull' is
deterministic. :tag retained for human readability; @sha256 is what
Docker actually resolves. Refresh procedure documented at the top of
decnet/distros.py.
When service_cfg["email_seed"] is absent, compose_fragment now falls
back to $PROJROOT/bait/ if that directory exists on the host. Lets
operators drop a deployment-wide bait corpus into one place without
threading email_seed through every decky's config. Missing dir keeps
old no-op behavior.
IMAP_EMAIL_SEED / POP3_EMAIL_SEED accept a directory (rglob *.eml +
*.json) or a single .json/.eml. Loaded entries CONCATENATE with the
hardcoded _BAIT_EMAILS — additive to the realism-engine emailgen
output rather than replacing it. JSON dicts require from_addr /
to_addr / subject / body; bare bodies are wrapped into RFC 5322 on
load. compose_fragment reads service_cfg["email_seed"] and bind-mounts
the host path read-only at /var/spool/decnet-emails/seed.
The TTP-tagging worker is now safe to run on agent hosts: EmailLifter
disk-reaches body-aware predicates from the local artifacts tree
(DEBT-035 unblocked filesystem access; DEBT-047 added the helper).
Drop `ttp` from MASTER_ONLY_COMMANDS in cli/gating.py and remove the
defence-in-depth `_require_master_mode("ttp")` call in cli/ttp.py.
`ttp-backfill` walks the master DB and stays master-only.
R0047 (BEC) and the encoded-payload predicate substring-match against
the email body. Shipping raw body text on the abstracted service bus
is the wrong privacy stance — the bus transport may swap from UNIX
socket to networked at any time, and "loopback today" is not a license
to put PII on the wire.
EmailLifter now opens the .eml lazily from
/var/lib/decnet/artifacts/{decky_id}/smtp/{stored_as} when a body-aware
predicate runs and parses the body in-process via stdlib email +
policy.default. The decoded body is memoized into the payload dict so
multiple body-aware predicates on the same event open the file once.
Bus envelope only carries the artifact pointer (decky_id + stored_as);
raw body bytes never cross the host disk boundary on the agent → master
hop. Filesystem access on agents is unblocked by DEBT-035 (setgid +
group-readable artifacts root, paid 2026-05-02).
The legacy inline body_text path is preserved — when the producer ships
body_text on the bus the helper short-circuits without opening the file.
Move artifact path validation + symlink-escape check out of the
admin-gated download endpoint into decnet/artifacts/paths.py so the
TTP EmailLifter can disk-reach .eml files at tag-time without
duplicating regex/root logic (DEBT-047).
The router now catches ArtifactPathError and re-raises HTTPException(400);
behavior is unchanged.
ANTI flagged two regressions in the existing command-event capture:
1. **Tell**: PROMPT_COMMAND lived in /root/.bashrc, the FIRST file
an attacker greps after landing root. The logger invocation
sitting there is plain-text honeypot signage.
2. **Bypass**: even when missed, `export PROMPT_COMMAND=""` silently
disables capture. ANTI personally bypasses this on engagements.
Reshape:
* Move the assignment to **/etc/environment** — read by pam_env at
session open (sshd via /etc/pam.d/sshd, telnet via
/etc/pam.d/login), before any shell rc file fires. Far less
obvious than .bashrc; a casual `cat .bashrc` no longer surfaces
the capture.
* Define the helper as a function `__bash_history_sync` in
**/etc/bash.bashrc** (system-wide bashrc, sourced by every
interactive bash). Function name reads as generic bash
housekeeping; no DECNET branding in the symbol.
* Pin both the function and PROMPT_COMMAND **readonly** so
`export PROMPT_COMMAND=""` fails with "readonly variable"
instead of silently winning. Mitigation, not airtight —
`bash --norc` still bypasses — but the passive `export`
bypass is closed.
The actual `logger --rfc5424 --msgid command ... CMD ...` invocation
is preserved exactly; only its location and the readonly guard
change. R0001–R0030 (command-rule pack) consume the same syslog
shape as before.
Three new tests assert: the value lands in /etc/environment, the
function body lives in /etc/bash.bashrc, no PROMPT_COMMAND line
remains in /root/.bashrc, and `readonly PROMPT_COMMAND` /
`readonly -f __bash_history_sync` are both present. Mirror
assertions added on the Telnet Dockerfile via
test_config_schema.py.
Real Linux deployments (especially Ubuntu cloud images) ship a non-
root admin user; honeypots that only accept root logins are a tell.
Add a second account on both SSH and Telnet decoys, configurable
via service_cfg keys `user` / `user_password`, defaulting to
`ubuntu` / `admin` so the lure is live on every fresh deploy.
* `decnet/services/{ssh,telnet}.py` — two new ServiceConfigFields
(`user` string, `user_password` secret) and matching env vars
(`SSH_USER` / `SSH_USER_PASSWORD`, mirror for telnet) propagated
via the compose fragment.
* `decnet/templates/ssh/entrypoint.sh` — runtime `useradd -m -s
/usr/libexec/login-session -G sudo "$SSH_USER"` so the new user
inherits the same sessrec pty-recording shell as root and lands
in the sudo group. Privesc attempts (`sudo`) flow through the
existing sudo-log capture; network-enum from the user's shell
rides the recorded transcript.
* `decnet/templates/telnet/entrypoint.sh` — same useradd pattern
(no sudo group — busybox+login telnet image has no sudo
package; privesc rides `su -` which itself flows through the
existing PAM auth-helper at /etc/pam.d/login).
* New tests for default + custom user / password + independence
from root password. Updated the schema-keys assertion to match
the four-field shape.
The new account is ALSO the natural home for the body-aware
predicates that were previously gated on root-only sessions —
attackers who land on `ubuntu@host` and run network-recon /
privesc commands now generate the same structured TTP-rule
events as root sessions did, captured via the same auth-helper
+ sessrec + sudo-log pipes.
DEBT-035 step 2. Today the artifacts subtree is auto-created by
Docker as root when a decoy container's bind-mount fires for the
first time. The resulting permissions are root:root 0o755 — the API
process (running as the decnet user) hits PermissionError trying to
read transcripts written by the container, and the soft-fail 404
path gets exercised on every fresh deploy.
Add `/var/lib/decnet/artifacts` to init's dirs list with mode 0o2775:
* 0o2000 — setgid bit. New files inherit the directory's group
(decnet), regardless of which uid created them. This is the load-
bearing bit for cross-container reads.
* 0o0775 — owner+group rwx, world rx. Group-write lets the API
process and the local TTP worker read each other's outputs
without a manual chown.
`_ensure_dir` already respects the full mode word via `os.chmod`,
no helper change needed.
Test asserts the resulting directory carries exactly 0o2775 after
a fresh `decnet init --prefix`. Defence-in-depth: this works even
if the per-decoy compose `user:` directive (next commit) misses a
template — files still land in the decnet group.
DEBT-035 step 1. The composer needs to know which uid/gid to inject
into each compose fragment's `user:` directive at deploy time. Today
the resolved `--user` / `--group` values reach systemd unit
rendering (init.py:349–354) but are not persisted anywhere the
composer can read them.
Persist as **names** (not numeric ids) under `[decnet] api-user` /
`api-group` in the rendered decnet.ini placeholder. Resolution to
uid/gid happens at deploy time on whichever host runs the deploy,
via `pwd.getpwnam(...)` / `grp.getgrnam(...)` — so the same user
name can have different uids on master vs agents (heterogeneous
/etc/passwd) without breaking artifact ownership. The existing
config_ini auto-translates kebab→DECNET_API_USER / DECNET_API_GROUP
at load time; no domain-map changes needed.
Two new tests: one asserting the rendered ini carries the
`api-user` / `api-group` keys for the values passed to `--user` /
`--group`; one round-tripping through `load_ini_config` to confirm
the env vars land in `os.environ` for the composer to pick up.
A previous agent (and several of my own commits) wrote to a top-level
DEBT.md without seeing the existing development/DEBT.md — the
canonical register since DEBT-001. Resulted in two parallel files,
inconsistent numbering schemes, and references that resolved to the
wrong place.
Migrate the six entries that landed in the rogue file into the
canonical register as DEBT-044 through DEBT-049, preserving their
status (resolved / partial / open) and cross-references. The
TTP_TAGGING.md references to "DEBT.md" already resolve to
development/DEBT.md by virtue of being in the same directory; only
the comment in decnet/ttp/impl/intel_lifter.py needed disambiguation
to "development/DEBT.md DEBT-048".
* DEBT-044 — `attacker.email.received` producer wiring (✅ RESOLVED 2026-05-02)
* DEBT-045 — EmailLifter heavyweight feature extraction (PARTIAL PAID 2026-05-02)
* DEBT-046 — EmailLifter mal-hash feed integration (open)
* DEBT-047 — EmailLifter R0047 BEC unblock (open, gated on DEBT-035)
* DEBT-048 — TTP intel provider mapping review (recurring quarterly)
* DEBT-049 — TTP Sigma adapter — post-v1 (open)
Summary table extended; "Remaining open" line updated; root file
removed. The DEBT-047 entry now explicitly cross-references DEBT-035
as the gating dependency for the R0047 BEC unblock.
The decky's Layer-2 extension (commit 291b78c1) emits body_simhash /
body_base64_bytes / html_smuggling on the message_stored log and adds
macro_indicator / encrypted booleans to each attachments_json
manifest entry. Lift them all onto the email.received bus payload:
* body_simhash — passes through as-is (16 hex chars or "")
* body_base64_bytes — coerced to int (0 on absent / malformed)
* attachment_macros / attachment_password_protected — OR-reduced
across the per-attachment manifest booleans; matches R0046's
matched_trigger semantics where a single positive lane fires the
rule
* html_smuggling — coerced bool from the decky's 0/1 int
Pre-Layer-2 message_stored events (older deckies, malformed log
rows) project to safe defaults: empty simhash, zero base64-bytes,
all booleans False — the EmailLifter then stays silent, never
fires a false positive on missing data.
R0042 (mass-phish) / R0046 macro / R0046 password / R0046 smuggling
/ R0048 (encoded payload) all fire end-to-end after this commit.
R0046 mal_hash_match and R0047 BEC remain deferred per their
respective DEBT entries (filed in the next commit).
Heavyweight Layer-2 extractors land alongside the cheap projections
shipped in commit e9324aca, so the EmailLifter R0042 / R0046 (macros
/ password / smuggling lanes) / R0048 fire from the bus payload
without the lifter having to reach back to disk.
Extractors:
* body_simhash — inlined 64-bit Charikar simhash (md5-keyed,
frequency-weighted) over word tokens of the union of text/* body
parts. Inlined rather than pulling the `simhash` PyPI dep, which
transitively brings numpy ~50 MB into a slim decky container; the
algorithm is ~15 lines and identical in extraction quality.
* body_base64_bytes — largest decoded base64 chunk's byte count,
scanning text body parts with the same `_BASE64_RE` the lifter's
`_p_encoded_payload` fallback uses. R0048 fires from this scalar
alone; the lifter's body_text fallback becomes dead in normal
operation.
* attachment_macro_indicator — stdlib zipfile sniff for
`vbaProject.bin` inside OOXML containers. Catches modern .docm /
.xlsm / .pptm and macro-injected .docx; legacy .xls (CFBF) is a
follow-up.
* attachment_encrypted — flag_bits & 0x01 on any ZIP / OOXML entry's
central directory; magic-byte match for 7z / RAR / CFBF (encrypted
Office wrap).
* html_smuggling — structural lxml parse first: fires when an `<a
download>` element coexists with a `<script>` referencing
`Blob` / `Uint8Array` / `URL.createObjectURL`. Regex pair-check
fallback on lxml parse failure (real-world phish HTML is often
malformed). Cuts the FP rate that pure-regex would produce on
legitimate "click to download" links.
Add `python3-lxml` (~5 MB Debian package, C-extension, no transitive
Python deps) to the SMTP decky's Dockerfile. simhash stays inline.
Per the dependency rule: lxml earns its weight by cutting R0046's
OR-combined FP rate; a heavier macro-detection lib (oletools ~5 MB
pure-python with msoffcrypto) would not measurably improve the
boolean signal we need, so stdlib stays for that lane.
Wires the EmailLifter (R0041–R0048) producer that DEBT.md item #3
deferred. After the existing add_bounty() call in _extract_bounty
(line 615), call _publish_email_received() which:
* resolves the attacker_uuid via repo.get_attacker_uuid_by_ip; drops
the publish if unresolved (the TTP worker can't anchor orphan
events)
* projects the message_stored fields onto the EmailLifter wire
contract: from_domain / mail_from_domain / return_path_domain
parsed via _domain_of, rcpt_count + rcpt_domains via
_rcpt_projection, attachment_sha256s + attachment_extensions
derived from the existing attachments_json manifest, urls from
urls_json, dkim_signed/spf_pass coerced from 0/1 ints to bool
* mirrors _publish_probe_pending's bus-per-call pattern and
swallows all exceptions (the bus is the notification layer, not
the source of truth)
Fires for both relay and non-relay SMTP services. R0041 / R0043 /
R0044 / R0045 are now live end-to-end; R0046 partial (extension
lane). Heavyweight predicates (R0042 simhash, R0046-deep, R0047 /
R0048 body_text) stay deferred per the EmailLifter heavyweight
DEBT entry.
The EmailLifter (R0041–R0048) keys on header-derived signals that the
v0 _summarize_message did not extract. Add cheap Layer 2 projections
inside the existing single-pass parse:
* return_path / x_mailer — direct header reads, decoded RFC 2047
* dkim_signed / spf_pass — booleans derived from any
Authentication-Results header (multiple lines tolerated; positive
verdict on any line wins)
* urls — http(s) URLs lifted from text/* body parts via a tight
regex, deduplicated first-seen-wins, capped at 64 in the wire
payload to bound the syslog SD value
Heavyweight extraction (body simhash, office-macro detection,
HTML-smuggling, password-protected archives, mal-hash-match,
body_text projection) stays deferred per the EmailLifter heavyweight
DEBT entry — those rules need privacy / extractor decisions before
they ship.
Three bug classes uncovered by the 2026-05-02 ship-time audit:
* AbuseIPDB code/name mismatch in v1: cat 10 was treated as DDoS (it's
Web Spam — DDoS is cat 4, intentionally unmapped per A.10) and cat 17
as VPN IP (it's Spoofing — VPN IP is cat 13). Both typos mirrored in
code AND the design doc Appendix A.10. Code now matches the AbuseIPDB
taxonomy exactly; cat 17 retargets to T1566 (email-spoofing as a
phishing precursor), and cats 7 (Phishing) and 16 (SQL Injection)
pick up T1566 / T1190 emissions that v1 didn't cover.
* ThreatFox dispatch keyed on `ioc_type` in v1, but `ioc_type` is the
indicator format (url / domain / hash variants) and carries no ATT&CK
signal. The canonical taxonomy field per ThreatFox's API is
`threat_type` (botnet_cc / payload_delivery / payload / cc_skimming).
Repoint dispatch through the new `threatfox_threat_types` payload
field; `ioc_type` rides as evidence only. Also adds the missing
cc_skimming -> T1056 (Input Capture) mapping and registers T1056 in
attack_catalog.py.
* GreyNoise bare-malicious lane: a `classification == "malicious"` row
with no recognised tag used to emit nothing. Now lights T1071 at a
half multiplier, suppressed when a tag already fires T1071 to avoid
double-stamping at conflicting confidence levels.
The TTP worker forwards the bus payload verbatim to the IntelLifter as
TaggerEvent.payload. The pre-audit publish payload only carried
{attacker_uuid, attacker_ip, aggregate_verdict, providers}, so even with
the new AttackerIntel taxonomy columns populated the lifter still saw
nothing. Lift the relevant fields (categories / tags / threat_types /
malware family / score / classification) into the bus event and decode
JSON-string list columns back to native lists at the boundary.
The 2026-05-02 ship-time audit of the R0054-R0058 intel rule pack found
that AbuseIPDB / GreyNoise / ThreatFox stored only the aggregate verdict
(score / classification / listed-bool) plus the raw response blob. The
TTP IntelLifter expects per-provider taxonomy fields (categories, tags,
threat_types) that were never populated, so R0054 / R0055 / R0057
emitted zero tags in production despite passing unit tests.
Add typed columns: abuseipdb_categories, greynoise_tags, greynoise_name,
feodo_malware_family, threatfox_threat_types, threatfox_ioc_types,
threatfox_malware_families. Each provider now parses the relevant
taxonomy out of the upstream response and writes it through
column_updates. JSON-list columns ride as TEXT with default "[]" to
keep the SQLite/MySQL backend split honest, deserialised back to native
lists by the repo on read.
The inspector was dumping the whole `CMD uid=0 user=root src=… pwd=…
cmd=nmap -p- 192.168.1.0/24` syslog body into a single ``command_text``
blob. ANTI: "I'd like to separate the fields." Done — three layers
work together:
1. Collector session aggregator: new `_parse_cmd_msg` splits the bash
PROMPT_COMMAND msg into `{uid, user, src, pwd, command}`. The
session-ended envelope's per-command dict now carries the
structured fields, with `command_text` set to just the cmd= value
(preserving embedded whitespace — `nmap -p- 1.2.3.0/24` etc.).
2. Rule engine: per-source_kind auxiliary evidence list
(`_AUX_EVIDENCE_FIELDS`). For `command` events the engine
automatically promotes uid/user/src/pwd into the persisted
`evidence` dict on top of the rule's explicit `evidence_fields`.
Engine-controlled, not per-rule — adding a new aux field is one
line here, not a 30-rule YAML sweep, and rule authors can't
accidentally drop it.
3. TTPInspector frontend: evidence renders as a structured
`kvs` grid (UID / USER / SRC / PWD / CMD rows) instead of
pretty-printed JSON. Primary-order list keeps shell fields at
the top; everything else falls below alphabetically so unfamiliar
evidence shapes still surface predictably.
Tests:
- session_aggregator pins the structured-fields emit (uid/user/src/
pwd/command_text without "CMD" prefix, embedded whitespace
preserved).
- rule_engine_tagger pins the aux-field auto-promotion + the
no-`None`-leakage path when payload doesn't carry an aux key.
"T1595" alone is opaque; "T1595 — Active Scanning" tells you the
story at a glance. The names come from a backend-side static catalogue
pinned to the same ATT&CK release as the rule engine
(_ATTACK_RELEASE = "v15.1") — names are the canonical MITRE labels,
not author-supplied strings on rules, so a rule author can't typo a
name and the entire fleet sees the typo.
- New `decnet/ttp/attack_catalog.py` with `TECHNIQUE_NAMES` covering
every technique_id + sub_technique_id emitted by `rules/ttp/`
(R0001..R0058 → 69 IDs in the v0 pack).
- `IdentityTechniqueRow` / `TechniqueRollupRow` / `CampaignTechniqueRow`
/ `TTPTagDetailRow` gain optional `technique_name` /
`sub_technique_name` fields. Repo + router populate them from the
catalogue at row-construction time. None when an ID isn't in the
catalogue — UI falls back to the bare ID.
- Coverage test (`tests/ttp/test_attack_catalog.py`) walks every
YAML rule and asserts every emitted ID has a catalogue entry, so
a future rule author who forgets to update the catalogue gets a
loud failure rather than a silent UI fallback.
Frontend:
- `TTPsObservedSection` shows "T1595.002 — Active Scanning:
Vulnerability Scanning" instead of just the ID, with overflow
ellipsis + tooltip for narrow viewports. Inspector header /
TECHNIQUE row also surface the names.
The TTPsObservedSection rollup tells the operator "we saw T1059" but
not why. Click any technique row → side drawer opens listing every
ttp_tag row in scope with the persisted evidence JSON, firing
rule_id / rule_version, source_kind / source_id, confidence, and
created_at. Mirrors the CredentialReuseInspector / BountyInspector
pattern (drawer-backdrop + bd-head/bd-body + kvs grid).
Backend:
- New `GET /api/v1/ttp/tags/by-{scope}/{uuid}/{technique_id}`
(`scope ∈ {identity, attacker, session}`, optional
`?sub_technique_id=`, `?limit=` capped to 1000). Returns raw
TTPTag rows newest-first.
- New `TTPTagDetailRow` Pydantic model + re-export.
- New repo method `list_tags_by_scope_and_technique` on
TTPMixin (+ abstract on BaseRepository) — single query branched
on scope; identity scope projects through `Attacker.identity_id`
the same way `list_techniques_by_identity` does.
- Tests: evidence round-trips, sub_technique filter, JWT-required,
empty scope, unknown scope rejected.
Frontend:
- New `TTPInspector.tsx` + `TTPInspector.css` (violet accent, slide
animation, focus-trapped panel matching the existing inspector
family).
- `TTPsObservedSection`'s TechniqueBar is now click+keyboard
activatable; clicking opens the inspector for that
(technique, sub_technique) tuple.
mypy clean. 532 passed in the targeted sweep.
The collector's `attacker.session.ended` envelope carries
`attacker_uuid: null` and `attacker_ip: <ip>` because the collector
doesn't talk to the DB. The TTP worker passed that null straight
through, and `TTPTag.__init__` raised the documented invariant:
ValueError: ttp_tag requires at least one of attacker_uuid /
identity_uuid; both NULL is not a valid anchor.
The worker now resolves `attacker_uuid` from `attacker_ip` via
`BaseRepository.get_attacker_uuid_by_ip` before fanning out the
event. When the IP isn't in the DB yet (profiler hasn't ingested
the row), the event is dropped with one log line — better than
exploding mid-tag.
- New `get_attacker_uuid_by_ip(ip) -> str | None` on the repo
(BaseRepository abstract + AttackersCoreMixin impl).
- `_resolve_attacker_uuid` helper in `decnet/ttp/worker.py` runs
before `_build_events`. Short-circuits when the payload already
has either anchor; drops the event when neither anchor is
resolvable.
- Tests pin: short-circuit on existing uuid/identity, repo lookup,
drop on unknown IP, drop on "Unknown" sentinel, drop on
no-anchor payload, drop on repo failure.
The TTP worker subscribes to attacker.session.ended but no upstream
component published it — the rule pack (R0001–R0030) therefore never
fired on live SSH traffic even after the consume-side wiring landed
in E.3.18a/b/c.
The collector now hosts a per-attacker_ip command index
(_SessionAggregator) that watches the same parsed-event stream as
_publish_log. Shell `command` events are appended to a per-IP list;
on `session_recorded` the aggregator slices the list to commands
inside the [ended_at - duration_s, ended_at] window and publishes
attacker.session.ended with the session metadata + commands list.
The TTP worker's _build_events fan-out (E.3.18b) turns each command
into a source_kind="command" TaggerEvent that the RuleEngineTagger
(E.3.18c) matches against R0001–R0030.
Memory bound: per-IP entries TTL-evict at DECNET_COLLECTOR_SESSION_AGG_TTL_SEC
(default 3600 s). Publish failures are swallowed in the aggregator —
a misbehaving bus cannot stall the per-container stream threads.
Honeypot SSH containers run `PROMPT_COMMAND` that calls
`logger --rfc5424 --msgid command -t bash "CMD …"`. The Docker-stdout
reader prepends an outer RFC5424 envelope (HOSTNAME=<decky>,
APP-NAME=1, MSGID=NIL) around that inner syslog line. Both the
collector parser (`parse_rfc5424`) and the correlation parser
(`parse_line`) saw the outer NIL MSGID and emitted `event_type="-"`
for every shell command — which:
- kept `Attacker.commands` rows missing `command_text`
- left R0001–R0030 (the pattern rule pack that matches shell
commands) with no haystack
- made `decnet.collector.log` show `event written … type=-`
for the very lines that should be `type=command`
Both parsers now detect the inner-RFC5424 shape (`<TS> <HOST> <APP>
<PROCID> <MSGID> <rest>`) when the outer MSGID is NIL and the SD-arm
is also NIL, and re-extract HOSTNAME / APP-NAME / MSGID / remainder
from the body. The collector parser also recovers the post-SD msg
tail when the SD block isn't `relay@55555` (the bash CMD line carries
a `[timeQuality …]` block) so the kv-fallback can find `src_ip`.
Mirroring tests in tests/collector and tests/correlation pin both
the unwrap and the regression guard for non-double-wrapped lines.
The endpoint was a contract-phase stub returning `[]` even though the
RuleStore loaded all 58 YAML rules at worker startup. UI saw an empty
table; operators couldn't tell whether anything was wired up.
- `api_list_rules` now calls `get_rule_store().load_compiled()` and
serializes each CompiledRule + its operational state into a
RuleCatalogueRow. Sorted by rule_id for stable golden snapshots.
- Add `description: str` to RuleSchema (pydantic) and CompiledRule
(NamedTuple, defaulted) + propagate through `_compile_one` so the
catalogue surfaces the human-readable YAML description, not just
the slug-style `name`.
- Update `tests/ttp/test_rule_engine.py` _fields assertion for the
new column; new `tests/api/ttp/test_rules_catalogue.py` pins the
catalogue contents (R0001/R0014 presence, row shape, sort order).
Worker behaviour is unchanged: it was already loading rules
correctly. This is purely a read-side wiring fix on the operator API.
The TTP worker entry moved out of decnet/cli/workers.py into its own
module so the TTP CLI surface (worker + admin verbs) is colocated,
mirroring decnet/cli/canary.py / webhook.py / swarm.py.
- New `decnet/cli/ttp.py` with `decnet ttp` (worker, ExecStart-stable
for decnet-ttp.service) and `decnet ttp-backfill --since-days N`.
- `decnet ttp-backfill` walks Attacker.commands and CanaryTrigger
history, dispatches each row through the live CompositeTagger,
persists tags via repo.insert_tags (idempotent INSERT OR IGNORE).
--dry-run / --source command|canary|all / --batch-size supported.
- Backfill deliberately bypasses bus publish — historical replay
must not re-trigger SIEM/webhook fan-out per TTP_TAGGING.md
§"Bus topics" loop-prevention invariant.
- Added `iter_attacker_commands_since` / `iter_canary_triggers_since`
read-only iterators on TTPMixin + abstract bindings on
BaseRepository.
- Master-only via gating; both `ttp` and `ttp-backfill` listed in
MASTER_ONLY_COMMANDS.
The canonical rule-based engine from §"Tagging engines, layered §1"
of TTP_TAGGING.md was fully implemented but never instantiated as a
composite child — pure pattern rules (R0014/R0017/R0023/... 23 rules
total) had no tagger to dispatch them.
- Add `RuleEngineTagger(Tagger)` adapter in rule_engine.py wrapping
`RuleEngine.evaluate()`. `HANDLES = {command, http_request,
auth_attempt, payload}` — the source kinds whose rules typically
live outside any per-source lifter.
- Adapter's `watch_store()` filters via `_is_engine_owned` so the
engine's dispatch index excludes lifter-claimed rules
(`match.kind: lifter:*`) and stays disjoint from per-lifter ownership.
- Prepend `RuleEngineTagger` to the `CompositeTagger` lifter list so
generic pattern rules dispatch before per-source cross-event logic.
- Composes with E.3.18a (worker hydrates `watch_store`) and E.3.18b
(worker fans session payloads into per-`command` events) — together
these three commits make R0001–R0030 actually fire at runtime.
R0001–R0030 declare `applies_to: [command]` and match per command, not
per session. The worker now translates one `attacker.session.ended`
payload carrying a `commands: list` into:
- one source_kind="session" event (behavioral / cross-event lifters)
- one source_kind="command" event per command (RuleEngineTagger)
Both string and dict command shapes are accepted; dicts contribute
their `id` / `uuid` / `command_id` as the per-command source_id so
the deterministic `compute_tag_uuid` keeps replays idempotent. Tags
from session + per-command dispatch are aggregated into a single
`ttp.tagged` envelope per upstream session.
Each per-source lifter holds its own RuleIndex and exposes an
`async watch_store()` that loads the corpus and drains store change
events forever. Until this commit nothing called `watch_store()` in
production — every dispatch index stayed empty and no rule fired.
- Add `WatchableTagger` runtime-checkable Protocol in `decnet.ttp.base`.
- `CompositeTagger.iter_watchables()` yields lifters that satisfy it.
- `run_ttp_worker_loop` fans out one task per watchable, cancelled
and awaited alongside pump/heartbeat/control in the existing finally.
- Watch failures log and exit the watch task without taking the
worker down — mirrors the pump-task tolerance contract.
Wires decnet-ttp as a first-class worker:
* `decnet ttp` CLI command (master-only via MASTER_ONLY_COMMANDS)
* deploy/decnet-ttp.service.j2 systemd unit (After= identity / intel
/ reuse-correlator workers; ProtectHome=read-only since
FilesystemRuleStore only reads ./rules/ttp/)
* deploy/decnet.target Wants= chain extended with decnet-ttp.service
* `ttp` was already in web/worker_registry.KNOWN_WORKERS
tests/api/test_schemathesis_ttp.py: TTP-routes-only schemathesis
suite, filtered via the OpenAPI tags=["TTP Tagging"] annotation
shared by the eight TTP routes. Reuses the live uvicorn subprocess
the wider test_schemathesis spawns; max_examples=400 keeps the
focused gate fast for E.3.13–E.3.16 iteration.
wiki-checkout/Service-Bus.md committed in its own repo: ttp.tagged
and ttp.rule.fired.<id> flipped from "reserved (TTP worker)" to
"decnet.ttp.worker" now that the worker publishes them.
Add BaseRepository.list_ttp_decky_phases(identity_uuid) returning
per-decky tag observations as (decky_id, tactic, created_at_ts) rows
ordered by creation time. Rewrite from_identity_row() to project
tactic → UKCPhase via tactic_to_ukc_phase and populate the four
phase-handoff maps (first/last_phase_per_decky,
first/last_seen_per_decky) so combined_campaign_weight finally lights
up on real DB rows — not just synthetic fixtures.
ConnectedComponentsCampaignClusterer.tick() pulls each active
identity's per-decky phase observations before projecting features.
Repo failures are non-fatal: a partial repo falls back to the empty
phase-handoff signal (legacy behavior) so the worker stays up.
tests/clustering/test_ttp_phase_handoff.py pins the production-row
pair clearing CAMPAIGN_EDGE_THRESHOLD on a C2 → DISCOVERY hand-off —
the trip-wire that says the whole project paid off.
commands_by_phase_on_decky itself stays empty on the production path:
it is consumed only by the synthetic-fixture similarity surface, and
the phase-handoff edge does not use it. Synthetic fixtures still
populate it directly via from_synthetic_identity.
Inner loop drains a per-process asyncio.Queue populated by one pump
task per topic in _TOPICS, dispatches each event through
CompositeTagger, persists via repo.insert_tags(), and publishes
ttp.tagged + per-technique ttp.rule.fired.<id> only when the insert
returned a non-zero rowcount.
CompositeTagger seeded with all six lifters (Behavioral, Intel,
CanaryFingerprint, Email, Identity, Credential).
Loop-prevention invariant from TTP_TAGGING.md §"Bus topics" enforced:
N replays of the same upstream event publish exactly one ttp.tagged
event. test_worker_bus covers both the direct invocation path and
the idempotency replay path.
Intel catch-up via attacker.session.ended is intentionally deferred
to E.3.14b — needs a session→intel join the repo doesn't expose yet.
IdentityLifter owns lifter:identity_* — currently R0003 (password
spraying). CredentialLifter owns lifter:credential_* — R0001 generic
auth brute, R0002 password guessing, R0004 credential reuse, R0005
valid-account use, R0006 default credentials.
YAMLs R0001/R0002/R0003/R0005/R0006 had their match.kind normalised
to fit the lifter prefix scheme — the design doc's promised "YAMLs
normalised in a separate refactor commit" lands here.
Identity-rollup tags null out attacker_uuid on emit so the worked-
example invariant holds (the tag belongs to the Identity, never to
one member IP).
Tests: test_identity_lifter.py + test_credential_lifter.py cover
each predicate's positive/negative path, state modulation
(disabled/clipped/expired), source-kind gating, and idempotent
replay. test_lifter_absence and test_lifters updated for the new
ctor signature.
SMTP message-level technique tagger per Appendix A.6: open relay abuse
(rcpt_count + foreign From), mass phishing (rcpt_count + body simhash),
phishing-kit X-Mailer, IDN/punycode URL, sender masquerade composite
(From/Return-Path/DKIM/SPF), malicious attachment (macro/.lnk/.iso/.img/
hash match), BEC subject+body composite, encoded payload in body.
PII discipline (TTP_TAGGING.md §'Hard parts §6') is enforced at the
lifter layer via _filter_evidence(): emitted TTPTag.evidence is
restricted to the EmailEvidence-allowed allowlist (body_sha256,
matched_headers — names only, rcpt_domain_set — domains only,
attachment_sha256s, rcpt_count) plus PII-safe match discriminators
(matched_kit, matched_trigger, matched_url_host, etc). Raw addresses,
raw body bytes, full URLs, and decoded base64 previews NEVER appear in
evidence — defense-in-depth over the YAML evidence_fields hint.
Tests: tests/ttp/test_email_lifter.py per-rule positive + negative +
PII allowlist guard + state modulation. tests/ttp/rule_precision/
test_email_rules.py xfail flipped to real precision (R0041-R0048
H-band ≥95%). Corpus rows updated to acknowledge that R0045 (masquerade)
co-fires with R0041 / R0047 when the sender-masquerade signals are
present alongside open-relay or BEC patterns — overlap is by design,
not a precision bug.
