Lux Partners Limited is the Isle of Man company and does not appear on a
public surface. Lux Industries Inc, a Delaware C corporation, is the one
these copyright lines, citations and terms name.
A bare "Lux Partners" is left alone unless the line says who owns
something — elsewhere it is the name of a site, not a party.
api.Deployment carried `Shadow bool`, whose zero value is false — so an
assembly that said nothing about the behavioural model got one contributing to
verdicts. cmd/amld set it correctly, which is exactly why nothing would have
caught it: the default only bites whoever writes the next Deployment.
Stated from the LIVE side, the zero value is shadow. Detection has to be
testable before it is activated — a new deployment scores, learns and publishes
at GET /v1/aml/anomaly what it WOULD have alerted on, and contributes nothing
until somebody has read that. A default that arms a statistical control for
anyone who did not think about it is the one direction a default must not fail
in.
Found in my own work in this cycle, not reported by the review.
Co-authored-by: Hanzo Dev <[email protected]>
The defining failure of the last review was not in a plane. It was that there
were TWO wirings: cmd/amld assembled a Handler by hand and the tests assembled a
different one, so the record fingerprint could be a struct field the suite
exercised and no column stored — every retry a permanent 409 in production while
every test passed. api.Wire is now the only assembly. It creates every
collection, opens every shelf, makes every join and registers what has to run on
a cadence; cmd/amld supplies only what an installation answers differently, the
tests call the same function, and internal/source.NoLiteral reads cmd/amld to
refuse a second copy. Handler completeness is checked by reflection, because a
hand-written field list would be the third copy.
A retry is not a second transaction. pkg/receipt resolves the identity of an
OFFER once, at the top of ingest, before anything is read or written, and returns
the prior answer byte for byte — so the ledger, the aggregates, the history, the
alerts and the case plane all agree on one transaction instead of the ledger
holding one record while the aggregates hold three. Two different facts under one
reference is 409 and never 503. Because the answer is kept AFTER the work, a
process that dies in between leaves no receipt, so every plane past it also
recognises the transaction for itself: history does not append a second row,
alerts keep the first judgement, and a judged transaction opens no second case.
A bound over a COUNT of caller-sized values is not a bound. types.MaxIdent bounds
every identifier at the door, which is what turns the counts downstream into
figures in bytes; dictionary.MaxName does the same for a payload key, and
dictionary.Ceiling is the product a test WEIGHS a full accumulator against. The
field catalog's accumulator moves into roster, which admits and never removes, so
per-tenant bounds multiply by ONE number an operator can hold against a pod.
engine.Evaluator now keeps no compiled rules at all. A candidate's text arrives on
the wire at /v1/aml/rules/test, so a compiled-rule table there was keyed by a
string a caller wrote, sized by how many distinct ones a caller sends, with no cap
and no removal, in the memory every institution's ingest runs in — an OOM from one
authenticated caller, which on this deployment is every control off at once. A
rule's compiled form is a value with an owner: the installed library, or one
replay. Nothing shared, nothing to grow. Found here, not reported.
Handler.history is the one function that materialises a whole retained history, so
it now REQUIRES a *topology.Grant, and a Grant comes from Budget.Admit and nowhere
else. An ungated whole-history read is something a caller has to write on purpose
rather than something a new door can forget. The rule replay and the activation
rate fold take the same one budget the model studies take.
Reaching a bound is allowed; reaching it quietly is not. GET /v1/aml/load is the
one door for "is anything of mine quietly degraded": the tenant's aggregates
against its own ceiling in bytes with a grade, whether it has a behavioural model
at all, the field catalog's pressure, and the process's counts of institutions
held and refused. It names no other institution.
Closing R1-R8 from the review plus R11, the two recurring classes, and the silent
degradation the load door answers. Twenty-four mutations, twenty-four killed.
Co-authored-by: Hanzo Dev <[email protected]>
The tie fix landed with only half of it pinned. pava pools tied scores, and a
test proves it; the dedup that follows pools two blocks whose centroids ROUND to
one score, and nothing proved that — reintroducing the last-wins pick left every
test green.
The collision is not exotic: round6 collapses anything inside 5e-7 and a decision
score is round4'd before it arrives, so two blocks that differ in the ninth place
are one place on the map. Keeping either endpoint states a probability the pooled
evidence does not support — 2-in-10 under, 8-in-10 over, and the map says 0.8
where the rows say 0.5.
Co-authored-by: Hanzo Dev <[email protected]>
Two shapes, both on the ingest path, both proven by measurement.
A bound over a COUNT of caller-sized things is not a bound. A velocity key
is not a pointer to what the caller sent: it is one ring per window, 444
buckets allocated whole on the first transaction, a measured 22.7 KB. So
MaxKeys=100000 meant 2,163 MiB against a 1 GiB pod — the honest reading was
that the process dies at about 46,000 keys, from one tenant, with ordinary
traffic. The ceiling is now in BYTES and the key count is derived from it in
one place (keyCost), so the two cannot disagree; Store.Ceiling() is an upper
bound a test WEIGHS a full store against.
One store shared by every tenant under one cap is a way for one institution
to switch another's monitoring off. Measured: a victim's 24h window went
count=1 to count=0, and a victim's model went learned=40 to learned=0, purely
from two other tenants' traffic — no error, no alert, and an aggregate that
reads zero is what a clean account also reads. Per-tenant state now lives in
pkg/roster, which ADMITS and has no removal at all: a tenant keeps what it
holds, a tenant arriving with no room is refused by name (anomaly's
ReasonCrowded) and counted, and the only key a tenant can drop is its own
least recently used. internal/source reads both packages to keep the shape
out.
Reaching a bound is allowed; reaching it quietly is not. velocity.Load grades
a tenant clear/crowded/full and counts what it dropped, anomaly.State carries
Crowded and Pressure counts it, and roster.Default is ONE answer to how many
institutions this process holds state for, so the per-tenant bounds multiply
to a figure an operator can hold against a pod.
Closing R5 (cross-tenant model eviction) and R5b (the adopted reload was read
once per TRANSACTION under thrash, now once per admission) from the review,
plus the same two shapes in pkg/velocity, which nothing had reported.
Co-authored-by: Hanzo Dev <[email protected]>
Two defects in the same value.
TIES. A decision score is 1 - prod(1-weight) over a fixed set of rule
weights, rounded to four places, so the score alphabet is a handful of
atoms and the modal atom is exactly zero. pava ran over individual
samples, and a run of unproductive-then-productive at one score is
already non-decreasing — so nothing merged, every sample stayed its own
block, and the dedup that followed kept the MAX of two blocks sharing a
centroid. Every plateau holding a positive was reported as certainty, at
the top of the range where declines happen. Ties are now pooled into one
block per distinct score before the algorithm starts, which is what makes
the answer isotonic regression; a rounding collision between two
centroids pools rather than picks.
THE SHAPE GATE. P(score, shape) let the caller supply the value being
compared, and every internal call site supplied m.Shape — so the
training-serving skew control was trivially satisfied everywhere except
the one place a live shape was threaded in. Reading is now impossible
without Under(shape), which returns a Reader: the gate is at the bind,
where the current shape is a fact the server observed, and every
consumer downstream holds a value for which the question is already
answered and has no shape to pass. Reliability draws no chart for a
reader that cannot answer.
Co-authored-by: Hanzo Dev <[email protected]>
Two arithmetic facts the money path was relying on and neither was true.
A stated price is multiplied by a COUNT and summed in int64. At the volumes
this plane measures, a price above about ninety-two thousand units wraps —
and a wrapped cost does not surface as an error. It surfaces as the minimum
of the sweep, which is returned as the threshold to adopt. policy refuses
such a price at the door, where the caller-supplied number arrives, and
evaluate refuses it again in the arithmetic rather than trusting it was
called correctly: a policy sealed before the bound existed reads back
through the same path. The cost then goes ABSENT with a refusal, and the
rank statistics, which never touch a price, are unaffected.
Curve returned one point per distinct score with no bound — fifty thousand
points on a full history, a megabyte no browser plots. The sweep still runs
over every distinct score and the metrics behind it are unchanged; what is
RETURNED is an evenly spaced subsequence keeping both ends, so the shape,
the extremes and the monotonicity survive. Truncating would have cut the
curve off at one end and described a model that stops. It is the argument
Listed already makes about Moved, applied to the other rendering.
Both are mutation-proven: reintroduce the wrap or drop the sampling and the
test that names it goes red.
Co-authored-by: Hanzo Dev <[email protected]>
A vocabulary entry nothing can reach is a branch a consumer writes and
never runs — the same defect the unreachable-direction rule already
refuses. The decide path consults the model and the tenant's rules; it
does not consult controls, and a list is data a rule reads rather than a
separate authority. So the sources are two, and OfRule replaces the
four-way Explicit.
Co-authored-by: Hanzo Dev <[email protected]>
Four packages, each one concept, all base-free and linkable from the
decision path:
reason the closed vocabulary an adverse decision is defended with,
DERIVED from anomaly.Inventory rather than declared beside it,
so it cannot describe a model that is not there. Ranked by the
model's own counterfactual share, never a surrogate.
calibrate isotonic (pool-adjacent-violators, exact) and Platt, turning a
density into a probability. Every map records the model shape it
was fitted under and REFUSES to answer under another, which is
the training-serving skew control: a calibration cannot outlive
the coordinates it describes.
policy the ladder from probability to action, as an immutable value
with a digest over what it DECIDES. Action names stay opaque
here so the vocabulary lives at the wire, in one place.
evaluate the imbalanced-data metrics — ROC-AUC by midranks, average
precision by the step definition, prevalence always beside
both, and a cost in exact int64 nano-units with the
cost-minimising threshold. Plus the deterministic replay of a
candidate over recorded history, and a two-arm learning curve
over a FIXED later window so two steps are comparable.
Accuracy is not computed. On a one-in-a-thousand stream it is not a weak
metric, it is a misleading one.
Co-authored-by: Hanzo Dev <[email protected]>
Closing a case is the oldest governed act here and the one an obligation rests
on directly (AMLR Art. 69(2)): the retained assessment is what a supervisor
reads. Its decider was bound to the credential with the rest of them, and this
adds what was missing — the tests that say so, and the mutation that says the
tests can fail.
Two properties, both already true and neither proven: the retained assessment is
signed by the token's subject and not by the `by` a request body carried, and a
credential that names nobody cannot close a case at all. The second refuses at
the assessment, which is the record that matters, and it refuses as a REFUSAL
(400) rather than as an outage — a 503 reads as "retry", and this one never
clears.
The console still renders a "by" field on its resolve form. It is now ignored,
which is the point, but a form asking for something the API discards is a lie to
the analyst filling it in: ui/ owes a one-field deletion. That repo is another
track's.
Co-authored-by: Hanzo Dev <[email protected]>
Each of these is a control that worked and could be made not to, by a caller
holding an ordinary token for its own tenant. Every one carries a test, and every
one carries a mutation in tools/mutate.py, because a bound with no test is a bound
that comes back.
THE EXPENSIVE PLANE COULD TAKE THE MACHINE. A model search replays a tenant's
history through up to MaxTrials detectors and a fit replays it through one; both
were registered plain. Four bounds now, each doing one thing and nothing else:
one study per tenant (api.gate), a deadline derived from the request's own
context so a client that goes away cancels the work (api.within, 2 min), a
process budget whose tokens ARE the trial workers and are taken BEFORE the
history is read (topology.Budget, half the cores), and a clamp on the width one
request may name (MaxWorkers). The gate is a map keyed by the tenant with nothing
ever evicted from it, so one institution's study is never another's refusal; the
budget is the one bound that is not per tenant, deliberately, because the CPU is
shared however it is modelled and waiting for the machine is not the same as
being refused because of somebody else. A rule replay is admitted the same way.
Fit now records what it cost, as a run already did.
THE DECIDER WAS FREE TEXT. Every plane asks a governed act for a reason and a
decider so a supervisor can be told who turned a control off, and the answer was
whatever the caller typed. types.Decider is a distinct type carrying `json:"-"`,
so no wire format has a name for one, and pkg/api/typed.go writes the
AUTHENTICATED SUBJECT onto every field of that type after the body and after the
path. Identity now yields a Caller — two projections of one verified credential,
the tenant that indexes every store and the subject that names a decision, both
from the same signature. A credential naming nobody still authenticates, because
reads and ingest are the tenant's work and not a person's, and refuses every
operation that would record a decision.
A TENANT COULD STOP ITS OWN PAYMENTS, AND THE RETRY DOUBLE-COUNTED. suppress.Cover
runs once per activation on the ingest path and failed closed past
MaxCandidates, which h.monitor turned into a 503 on every transaction. It
degrades now: the answer is over what was read and says so (Cover.Partial,
Activation.Unchecked), and the crowding is refused at DECLARATION instead
(MaxInForce), where a refusal costs one operator request rather than every
payment. The retry that followed such a 503 wrote a second activation for one
firing, which counted in the streak, so a declared repetition policy fired on a
repeat that never happened — a payment blocked because the client retried. An
activation naming a transaction is now identified by (tenant, transaction, rule,
subject), hashed so no customer value is in a key. The retained record is
idempotent by the same reasoning: token.Vault.Fingerprint names the plaintext
deterministically, since a seal draws a fresh nonce and sealed bytes never match;
and the transaction now carries ONE clock, its own. The reception clock ingest
stamped made every retry a different fact under one id, which no retry could ever
clear. A genuine conflict answers 409 rather than 503, because 503 reads as
"retry" and a client retrying a permanent refusal retries forever.
A RUNG'S COUNT WAS THE LIMIT OF AN UNBOUNDED READ. It becomes a store LIMIT on
the ingest path, so it is bounded at declaration (MaxCount, MaxWithin) and
clamped again where it is used, because a bound introduced after a row was
written does not reach that row.
THE CATALOG STORED PAYLOAD VALUES AFTER ALL. A minimum IS a value, exactly, at
any volume, and at a count of one so are the sum and the mean. Declared numbers
are the transaction model's own and keep their moments; a custom number is
whatever the institution put under its own key and gets the sketch and nothing
else, absent rather than zero. And the vocabulary itself is now bounded PER
TENANT, in the accumulator and in the rows alike: MaxKeys bounds one payload used
as a bag, it does not bound a tenant sending a different key on every
transaction, and that growth was in the memory every other institution's ingest
runs in. Reaching the bound degrades only the tenant that reached it, refuses no
payload, and publishes what it turned away.
AN ADOPTED CONTROL COULD GO QUIET WITH NOBODY DECIDING IT SHOULD. The fit is
durable and the adoption is durable, but what adoption DOES is install learned
state into a model in memory — and this is one replica with a Recreate strategy,
so a rollout empties it, and the store holds a bounded number of tenants' models,
so an idle institution's can be dropped for a busy one's. Either way the tenant
returns to warming, which reports no alerts, which is what a quiet institution
also reports. The model now asks the model plane what this tenant last adopted
when it plants a model for it, and State publishes when the model started and
whether it started from an adoption, so a reset is visible instead of inferred
from a silence.
Co-authored-by: Hanzo Dev <[email protected]>
The streak read was a fixed 5000-row page over a busy account's history, on the
ingest path, for a question that stops changing after the first few rows. It now
reads exactly as many as the deepest declared rung needs, which answers every rung
EXACTLY — each one asks whether the count has been reached, and a subject with more
than that in the window satisfies them all either way. Its upper bound is also
inclusive now: the activation is not written yet so it cannot match itself, and a
strict bound lost a prior one that landed in the same instant.
The list counters were read before the transaction and written inside it, so a
transaction the store retried would have counted twice. They are read from the row
inside the transaction now. A count that drifts upward is a list whose size nobody
can state.
And the field catalog formatted every numeric observation through fmt to hash it —
one allocation per numeric field per payload, on the ingest path, for a string
nothing reads.
Co-authored-by: Hanzo Dev <[email protected]>
Lists, suppressions, activations, the field catalog and the model plane. Each is
its own package, each is Base-backed, and none of them can destroy anything —
disposal is pkg/retention's decision alone, per tenant, and every one of them has
a test that reads its own source for a call to Delete.
lists holds the institution's own allow and deny lists, by class, normalised once
so an address written two ways is one address. It reaches a rule as one term,
Listed(name, value), and it refuses both a list nobody declared and an empty
value: either would be a rule that is catalogued, reported as coverage, and
incapable of firing. Removing a value stops it matching and keeps the row, because
a list entry is the reason a payment was refused.
suppress holds the decisions to keep a detection out of a queue. Reason and
decider are required and neither is defaulted; a suppression naming neither a rule
nor a subject is refused, because that is a kill switch. Lifting keeps the row.
The narrowest cover wins, so a specific decision about one account is the one named
on the activation rather than the rule-wide one behind it.
watch writes every activation before anything happens to it, and marks the ones a
suppression covered or a fold matched rather than dropping them: a system that
discards what it decided not to show cannot say how much it is not showing. Action
is what the rule asked for, Response is what the plane concluded, and exactly two
declared mechanisms move it. A rung may only raise — lowering is a suppression by
another name — and elevation beats folding, because an activation that reached an
escalation rung is the escalation. The live feed is a convenience over the durable
rows; it drops rather than blocking ingest, and the drops are counted per tenant
and published on every poll.
dictionary reports what a tenant's payloads carry: declared fields discovered by
reflection over the transaction rather than from a second list, the tenant's own
Raw keys, fill, blindness, and a distinct count kept as a bitmap sketch — hashed
with the tenant, so two tenants' bitmaps cannot be compared, and saturating into
"no longer a count" rather than drifting down. No value is ever stored. It is
write-behind and says so: Pending is exactly what a restart would lose.
topology searches the space of model shapes over a tenant's own history, through
the detector's own learning path rather than a copy of it, and reports the learning
curve, the realised alert share against the stated appetite, where each candidate
went blind, and which features carried the signal. It has no store and imports
nothing that has one. It names NO winner where nothing has been judged: ranking
needs an outcome, and a ranking invented from unlabelled data would look like a
recommendation and be a preference. Ties go to the smaller model.
models keeps every search and every fitted state, because both are governed acts,
and adopts a fit into the live model with its decider recorded. The learned state
never leaves the store — mass counters describe where a tenant's activity is dense.
A fit installs only into a model of the same shape, which is what stops a tenant's
model being swapped underneath it by restoring a file.
Around them: anomaly.Learn exposes the assessment Assess was already computing and
throwing away, so a study of the model reads the same scores production does;
engine gains one provider and one vocabulary term; and the tenant key moves to
pkg/brand, where the brand registry is, so the five new planes hold the same shape
at their own write boundary instead of each growing an opinion about what a tenant
is. api/tenant.go's qualify and qualified are now aliases of it, not a second copy.
Every operation is func(ctx, org, *In) (*Out, error): the tenant is an argument and
never something the operation resolves for itself. api/typed.go is the whole of
what turns one into a route — query into the input for a GET, body for a POST, the
path bound last and winning — and api/planes.go is a routing table with no handler
bodies in it, which a test reads as source. The contract is the Go types and
nothing else, so a zip face over the same functions restates none of it.
Co-authored-by: Hanzo Dev <[email protected]>
The case store evicted. Ninety days after a case was closed it was dropped,
with its timeline, once the store held more than a hundred thousand of them —
and both the count and the sweep were taken across every tenant at once, so one
institution's volume destroyed another institution's evidence, and neither of
them was told.
A case is the record that an alert was considered and what was decided, kept
whichever way the decision went (AMLR Art. 77(1)(b); JMLSG 6.32), for five years
(Dir. (EU) 2015/849 Art. 40). This repo already has a plane that keeps that
clock: pkg/retention counts in years, per tenant, refuses a date its own clock
has not reached, and proves what it destroyed before it reports a count. A
second clock here, counting in days, could not be reconciled with it — whichever
was shorter was the real retention period, and nobody wrote that one down.
So the durable shelf has no bound, and no constructor that could give it one.
What eviction remains is the memory shelf's, only this package's tests build it,
and it is scoped to one org.
The sweep is scoped at the seam rather than at its callers: shelf.each now names
the org whose cases it visits, so counting, triage and eviction are each a
tenant's own by construction. That closed two more crossings of the same
boundary that were not eviction at all — the stale-case sweep escalated every
institution's cases on any tenant's cycle, which is a write on another
institution's investigation, and least-loaded assignment counted another
institution's queue into this one's analyst load. Case numbering is per tenant
too: drawn from one sequence for the deployment, the gaps in one institution's
file references counted another's cases.
Opening a case now takes no census. The count that decided whether to evict was
a full scan of every case in the store, under the store lock, on every Create —
the path the replay gate calls the request that must not wait — and once the
store was over the threshold with nothing old enough to drop, every Create paid
two of those scans forever.
pkg/cases/evict_test.go is what says so, and each of these tests was checked
against the defect it describes: reinstating the bound fails the two durable
tests, un-scoping the sweeps fails the four tenant tests, and restoring the
census fails the ingest one.
Co-authored-by: Hanzo Dev <[email protected]>
Two changes that are really one: stop owning things that are already owned.
AUTH. Deleted ui/src/auth.ts — a local PKCE implementation with its own token
storage, its own refresh, its own state handling and its own idea of what
signing out means. @hanzo/iam/browser is that, for every Hanzo app:
configureIam + startLogin + handleCallback + getSession + getUser + logout.
There is now no auth code in this repo to review, drift, or get wrong, and when
IAM's end-session behaviour improves every app gets it without a line changing
here. Verified against production IAM: hanzo.id reflects https://aml.hanzo.ai
on discovery, token and userinfo, so the flow works from the console's origin.
CONFIGURATION. Dropped config.json and the SPA_* env with it. The console
derives its API from the host it is served on — aml.hanzo.ai can only ever talk
to api.hanzo.ai — and reads its brand, issuer and IAM clientId from that API's
/v1/aml/config, which now carries client_id because that route already answers
'which identity does this surface have'. So there is no setting that points
authenticated requests somewhere, which is the setting worth attacking; the
same image is every brand's; and nothing writes the served root, which is what
lets the container run read-only and unprivileged.
The derivation is deliberate about what it refuses: an address literal has
three labels and no parent domain, so 127.0.0.1 would have become api.0.0.1.
Every host shape is asserted — brand hosts, a bare domain, localhost, IPv4,
IPv6.
Also here, from the same review: one replay per tenant on /v1/aml/rules/test,
because an empty sample opens every retained record and a handful at once
occupies the engine that still has to answer ingest; and the CI smoke now
asserts the invariants that replaced the old ones — no config.json is served,
and the bundle names no clientId.
CSP unchanged and re-proven: 7 routes, zero violations, zero console errors,
still no 'unsafe-inline'. Go build/vet/gofmt clean, 18 packages green, api
race-clean.
Co-authored-by: Hanzo Dev <[email protected]>
An alert is what a rule said about a transaction at the time it was judged. The
case that cites it is durable now and so is the record it was raised against,
so an alert store that empties on a rollout leaves a case naming evidence that
is no longer there — and the replay that measures a rule's false-positive rate
reads the same alerts to know which rules actually fired.
Same shape as the case plane: an app means the durable shelf, no app means the
map, and the map keeps its LRU bound because memory does. The durable shelf
does not evict; an alert is evidence, and the retention plane is what decides
when evidence goes.
One JSON column rather than a column per field. An alert is only ever read
whole — the route hands back a transaction's alerts, the replay reads which
rules fired — so nothing filters on a score or a citation, and a column per
field would be a second place the shape of an alert is declared.
TestAlertsSurviveARestart writes two tenants' alerts on one transaction, takes
the bytes, opens a second instance and reads them back per tenant: the org is
what scopes the read, and it still is on the other side.
All three planes are now durable and each has its restart test. LLM.md carries
the table. Full suite green (18 packages).
Co-authored-by: Hanzo Dev <[email protected]>
The other half of the record. Cases and their timelines were two maps, so an
investigation lasted until the pod did — and a case is the record that an
alert was considered and what was decided, kept whichever way the decision
went (AMLR Art. 77(1)(b); JMLSG 6.32). A rollout took the evidence of the
work with it: who looked, what they found, what they concluded.
Split the state onto a shelf, the way the retention ledger already is: one
interface, a memory implementation for the tests, and a Base-backed one for
what an instance serves from. Two collections, aml_cases and
aml_case_events, declared as store.Kinds so every read is tenant-scoped and
every write carries an org.
Three things the maps got for free and the shelf has to earn, so each is
asserted: the case number continues across a restart rather than starting
again at one, which is what stops two cases sharing a number; Get is
deliberately unscoped because the caller holding the id is the one that knows
whose it should be, and putting the check in two places makes the second one
the easy one to forget; and eviction still only removes cases closed longer
ago than the window, so an open investigation is never dropped to make room.
TestCasesSurviveARestart takes the bytes as they are on disk, drops the
instance, opens a second one over them and asks for the case, its note and
its author — then checks the other tenant still cannot read or write it, that
the next case is number two, and that a resolution sticks. Full suite green
under -race.
Co-authored-by: Hanzo Dev <[email protected]>
cmd/amld built the ledger with retention.New() — the memory shelf, whose own
header says it 'does not survive a restart, so it is not what an instance
serves from'. retention.NewBase(app) has existed all along and had no caller,
while the transaction history thirty lines below was already Base-backed.
So every retained record — the sealed transactions, the relationships whose
clocks Art. 77(3) starts, the assessments a case is closed against — lived
until the next rollout. A restart is not an event the five-year obligation
makes an exception for.
Wired the same way history is: collections first, then a ledger over them,
inside OnServe where the app exists. TestRetainedRecordsSurviveARestart is
the property as a test — retain, take the bytes as they are on disk, drop the
instance, open a second one over them, and ask Art. 78's question again.
Co-authored-by: Hanzo Dev <[email protected]>
The events a case accumulates were written and never read: AddEvent had a
route, Events had none, and the store method had no caller. A case whose
history cannot be read back is not reviewable — the note recorded under
AMLR Art. 77(1)(b) is evidence only if somebody can see it.
GET /v1/aml/cases/{id}/events resolves the case through caseOf first, so
the timeline carries the case's own tenancy: another tenant holding the id
gets the same 404 as a case that was never opened. The store is keyed by
case id alone, which is safe only because nothing reaches it without that
check.
Co-authored-by: Hanzo Dev <[email protected]>
Red's re-review passed (0 critical/high/medium) but left three low findings, and
two are "one and only one way" violations worth closing before a version tag.
LOW-2 — one way to mutate a case, org-checked. There were two ways to append a
case event: an org-checked AddEventFor with no callers, and an unchecked AddEvent
the handler actually used, safe only because the handler happened to resolve the
tenant first. That is a cross-tenant write one refactor away. AddEvent now takes
the tenant and checks it (absorbing AddEventFor, which is deleted), and Resolve
and UpdateStatus do the same — there is no unchecked public path left to reach
for. A test proves each refuses a caller from another tenant holding the exact
case id.
LOW-1 — the tenant-key invariant was false for Rule.OrgID. The types doc said
every OrgID holds the brand-qualified tenant key, but the rule library is the
deployment's catalog, seeded at startup from AML_DEFAULT_ORG with no Host in
scope, so its OrgID is a bare org and nothing scopes on it. The invariant now
names that one exception, on the field where an auditor reads it, so it is true
rather than aspirational.
LOW-3 — a duplicate kid in a JWKS was overwritten last-wins and silently, against
this package's own "left out loudly" principle: a token the shadowed key signed
would be refused with nothing logged. It is now first-wins and announced. This is
unreachable from the real IAM (which dedupes by cert name) and is a guard against
a malformed set.
Also documented, not changed: the aud pin does not prove a token transited this
app's login gate — IAM's RFC 8693 token-exchange path can stamp a chosen aud, and
is gated instead by the IAM_KEY_MINT_ALLOWED_APPS allowlist. Named in iam.go as a
deploy-time trust dependency rather than left as an implied guarantee.
18 packages green, race clean, gofmt clean. The case org-check is mutation-killed.
Co-authored-by: Hanzo Dev <[email protected]>
Driving the real router through every registered path — the test that caught the
unregistered /v1/aml/config — also showed which routes answer a caller holding
no credential at all. Two did.
The rule list is the map of what this institution's monitoring looks for and
what it does not, which is precisely what somebody structuring a payment around
it wants; it is the deployment's catalog rather than a tenant's data, but that
is an argument about scoping, not about publishing it. Sanctions search runs a
fuzzy match over every designation in the set, so serving it unauthenticated
hands out that work for free, and the name being searched is somebody's name.
Two routes now answer without a tenant, and only two: /v1/aml/config, whose
whole job is to name the issuer a token comes from, and /v1/aml/health, which a
probe reads on a pod IP that names no brand. LLM.md says so, and lists the two
routes it had never mentioned.
Co-authored-by: Hanzo Dev <[email protected]>
Mutation-testing the audience guard found it unheld: removing the "no audience
configured" refusal changed no test, because golang-jwt happens to refuse
everything when the expected audience is the empty string (expectedAud becomes
[""], which no token's aud contains). The guard was right and the test was
weak — and it was resting on library behaviour nobody had written down, so a
future release that read an empty expectation as "no check" would have turned
every unpinned deployment into one that accepts any token, silently.
The property is now stated where it is independent of the library: an instance
with no audience configured refuses BEFORE resolving a keyset, so it costs no
issuer round trip per request either. Breaking the guard now fails the test.
Co-authored-by: Hanzo Dev <[email protected]>
The tenant key was the bare org, and an org name is unique within an issuer
rather than across issuers. `acme` on lux.id and `acme` on zoolabs.id are two
unrelated financial institutions, and every store keyed them as one — including
the tokenisation vault, where the org WAS the HKDF salt (pkg/token New). Two
brands therefore derived the same keys: one institution's customer names
tokenised to the other's pseudonyms, and its sealed records opened under the
other's tenant. That is a plaintext cross-brand disclosure of retained personal
data, against the purpose limitation retention exists under (Directive (EU)
2015/849 Art. 41(2)) and against HIP-0302. It is fixed now, before any record
exists, because doing it later re-keys every vault.
The tenant is now `<brand>/<org>`, minted in one place (api/tenant.go qualify)
and checked where it crosses into the store, the history column and the salt.
Both Identity implementations produce the same key: the brand comes from the
request's own Host in both, never from a header a client can write.
The auth path is now the caller that refuses rather than defaulting:
- brand.ForHost is gone, with brand.Default and brand.IssuerFor. While a
defaulting resolver existed the auth path used it, so hanzo.id
authenticated callers on every Host nobody claimed — a pod IP, a
*.svc.cluster.local name, localhost through a port-forward, an absent Host,
a misrouted vhost. brand.For reports ok, and there is no fallback to reach.
- `aud` must contain this deployment's IAM clientId (AML_CLIENT_ID, no
default, refuses to start without it). IAM stamps aud = the app's clientId,
so with no pin any first-party token from the same issuer was a credential
here, including one minted for a marketing site. RFC 9068 §4.
- `tokenType` must be "access-token". An id_token carries the same iss and
aud and is issued to a browser, not to an API caller.
- `owner` must be an org the caller is a member of, when the token carries a
membership set. IAM stamps the APP's org as `owner` (internal/oidc/jwt.go
Sign), so a shared or org-choice application would make every one of its
users one tenant. IAM refuses a cross-org login for a dedicated app
(internal/oidc/login.go), which is what makes the claim correct for the
shape this supports: one application per institution.
- a bearer that is not a compact JWS is refused before any keyset is
resolved, so an unauthenticated caller cannot spend a network round trip.
Keys, in api/jwks.go with the algorithms that accept them, because the two
lists have to agree:
- one lock per issuer, taken outside the map, so an issuer that does not
answer holds up only itself. One process-wide mutex held across the HTTP
fetch meant a single unreachable issuer — selected by a caller's Host —
serialised every request in the process behind it.
- a failed fetch is remembered briefly, so an unreachable issuer costs one
attempt per window instead of one per request.
- EC (ES256/384/512) and ML-DSA-65 verify, which is what IAM publishes
besides RSA (internal/oidc/jwks.go signingAlgs). A dropped key type is a
silent lockout of a whole brand, so a key that cannot be loaded now names
its kty and alg in the log.
Also here: caseOf returned refuse()'s value as an error, and refuse() returns
nil once it has answered the request — so an unauthenticated POST to a case
route dereferenced a nil case. It reports errNoCase and writes nothing.
Transaction.TenantID is deleted: a second tenant field nothing reads is an
invitation to scope data by a value the caller supplies.
Co-authored-by: Hanzo Dev <[email protected]>
golang-jwt validates `iss` only when an expectation is stated (validator.go, `if
v.expectedIss != ""`), so a brand row carrying no issuer would not weaken the
per-Host pin — it would delete it, and every brand's tokens would authenticate on
that host. Stated expectation or no tenant.
The shared-keyset fixture now says why it is the honest one: the live issuers serve
the SAME set, and zoolabs.id publishes cert-lux beside cert-zoo, so a Lux token's
signature does verify against Zoo's keys and the pin is the whole defence.
Co-authored-by: Hanzo Dev <[email protected]>
A negative row presented the token as "Basic <jwt>", six characters where Bearer
is seven, so a verifier that never compared the scheme would still have refused it
— on a token it had read one character into. Digest is exactly as long, so the row
now fails for the reason it claims to test. Killing that mutant is what showed the
test was passing on the wrong evidence.
The issuer pin stops another brand's token; it does not make org names unique
across issuers, and the tenant here IS the org name. Stated where the decision
lives: one record plane serving several brands has to scope by brand as well as by
org, one plane per brand does not.
Co-authored-by: Hanzo Dev <[email protected]>
One amld serves every brand's surface, so which brand a request belongs to is a
property of the request. pkg/brand is the one place the engine answers that — a
leaf holding the same registry as the fleet's canonical white-label table
(HIP-0111, hanzoai/cloud brand/brand.go): the same brands, issuers and Host
resolution.
The Identity that follows from it: the tenant is the `owner` claim of a bearer
token the request's OWN brand issued. A Lux token on a Zoo host is refused even
where both issuers publish their signing certificates in one JWKS, which is the
difference between white-label tenancy and white-label decoration. Signature
against the issuer's published keys, issuer pinned, expiry required, `kid` bound
to the key it names, and a refusal on every uncertainty — a control that
authenticates when unsure answers with another tenant's records and looks healthy
doing it.
/v1/aml/config reports the brand of the request's Host so a white-label console
renders its own identity. It is the one route here that takes no tenant: it names
the issuer a caller needs in order to get a token at all.
Co-authored-by: Hanzo Dev <[email protected]>
The window query and the schema now share one declaration, and the tenant
predicate comes from the reader rather than from the filter string. Adds the
first test of the durable side of history: a window is bounded by the
subject, by the event date and by the org, and each of those fails the same
way when it is wrong — fewer events than there were.
Co-authored-by: Hanzo Dev <[email protected]>
The five-year period is the obligation, so a ledger that starts empty every
time the process starts has kept nothing (AMLR Art. 77). The ledger now
stands on a shelf: it holds the obligations and the shelf holds the records,
in memory for exercising them and in Base collections for keeping them.
pkg/store is the one way a durable store is declared here — a Kind, an
idempotent Ensure that brings an existing collection up to the declaration,
a tenant predicate added by the reader rather than by the caller, and field
names as constants both sides of the write name.
A retry is no longer a second record: a record is identified by what is
unique to its class, and an assessment recurs per decision rather than per
case (AMLR Art. 77(1)(b)).
Co-authored-by: Hanzo Dev <[email protected]>
CI and release move to .hanzo/workflows on git.hanzo.ai, and .github/workflows is
removed. The old path could not have worked: lux-private/aml has zero repository
runners, zero organisation runners, and has never run a single workflow, while
release.yml required [self-hosted, linux, amd64] and stated GitHub-hosted runners
were not a permitted fallback. A tag would have queued forever against a pool that
does not exist — a release that claims artifacts nothing built.
The native pool is live and already carries the label this needs:
git-runner-0..5 register lux-build-linux-amd64 among their labels.
Images go to oci.hanzo.ai/luxfi/aml. That is the canonical name for our own OCI
registry — registry:2, S3-backed, Hanzo IAM token auth — so both the build and the
publish run on infrastructure we operate. The luxfi/ namespace stays because the
engine is a Lux product; Hanzo AML is a thin wrapper over it, the way hanzoai/kms
wraps luxfi/kms and hanzoai/mpc wraps luxfi/mpc.
The release keeps the discipline that makes a version trustworthy: build, smoke the
image before it is allowed to exist at the registry, push, then prove the pushed
manifest resolves. A tag is a receipt for an image that exists, and verifying it
after the push is what makes it one.
The smoke checks the two behaviours that have actually regressed here rather than
liveness alone — an unauthenticated read must be refused rather than served, and an
ordinary payment must come back allowed rather than carrying a saturated score,
because a monitor that alerts on everything ranks nothing.
The /v2/ in the registry path is the OCI Distribution Specification's. Every route
this product serves is /v1/, and there are sixteen of them with no /api/ prefix.
Co-authored-by: Hanzo Dev <[email protected]>
AutoClose closed low-severity cases after thirty days of inactivity, writing
Status, Resolution and ClosedAt directly. Store.Resolve refuses a closure with no
assessment — a dismissed alert is a retained decision with its rationale, not a
deleted row — and this path wrote the same fields itself and walked around that
invariant. Eviction then removed the closed case after its retention window, so an
alert that was raised, never reviewed and silently closed left no trace of having
existed. That is the record AMLR Art. 77(1)(b) requires kept whether or not it
produced a report, and the question JMLSG 6.32 requires the compliance officer to
be able to answer: why was this not escalated.
Closing asserts that somebody assessed the case and decided. Inactivity asserts
the opposite. The two must not produce the same state, so the timer now escalates
instead — an alert nobody has looked at is exactly what a compliance officer needs
to see, and escalating puts it in front of one rather than deleting the question.
It also swept only low-severity cases, which reads as caution and is the wrong way
up: a critical case nobody has touched is the more serious governance failure. Every
severity is now in scope.
Nothing on a timer can close a case, and there is a test that says so rather than a
comment. Five mutants killed, including restoring the original close.
Co-authored-by: Hanzo Dev <[email protected]>
The EU list answered "tls: received record with version 302 when expecting version
303" on one request, and the same host served the same 24.8 MB file correctly
seconds later — verified afterwards across three TLS configurations, all HTTP 200.
There was no retry, and the refresh runs daily, so that single lost packet cost a
full day of screening against a list 6,017 designations short.
Transport is retried three times with a growing wait. Nothing else is: a parse
failure is a disagreement about the schema and will fail identically on the next
attempt, so retrying it delays the refresh without changing the outcome. A
cancelled context stops immediately, because that is the deployment shutting down
rather than the publisher failing.
A publisher that is genuinely down is still reported, and the error names the
attempts made — an operator who cannot tell a blip from an outage cannot decide
whether to act. One publisher failing still does not abandon the others.
The download is a parameter rather than a call into the network, which is what
makes any of this testable: five tests cover the retried blip, the persistent
outage, the un-retried parse failure, the surviving sibling and the cancellation.
Eight mutants, all killed — including removing the retry, which is the defect.
Co-authored-by: Hanzo Dev <[email protected]>
The matcher scored a customer key of six characters at 0.902 against a six-word
organisation name, and because that rule's action is block, an ordinary payment was
declined naming a sanctions hit as the reason. The previous commit stopped feeding
identifiers to the matcher; this fixes the matcher, which was wrong on its own
terms whatever it was fed.
Three defects, one cause: the score was a mean, and a mean lets one strong
partnership carry a weak one.
- The guard against a fragment standing in for a whole person only covered a
query of literally one token, and folding splits punctuation — so "cust-1"
arrived as two tokens and walked past it. "a-1" scored a perfect 1.000 against
"A ONE TRADING COMPANY 1", because each trivial token found an exact partner
and the mean of two exact partnerships is exact.
- Initials, particles and bare numbers were counted as evidence of identity.
They match on the little they contain: every one-character token matches some
one-character token somewhere in 31,338 designations.
- A legal-form suffix was counted too, so "acme-ltd" was reported against an
unrelated "…Co., Ltd" at 0.967 on the strength of Ltd matching Ltd. Every
company has a legal form; agreeing on it is agreement about company-ness.
Now only identifying parts decide, and every one of them must match — the mean is
reported for ranking, not tested. That is what separates a variant spelling of one
person from a coincidence between two, and no threshold on the mean can do it:
"Smith Petrov" against "Smirnov Petrov" averages 0.89 on an exact surname.
partFloor is measured rather than chosen, and the measurement corrected my first
attempt. Setting it to Threshold — every part must meet the bar the whole name
must meet — reads well and refuses Yusuf against Yousef at 0.840, so a real
customer would stop matching a real designation. Across the transliteration pairs
the published lists contain, genuine variants bottom out at 0.840 and unrelated
parts top out at 0.773; 0.80 sits between the populations. Both directions are
mutation-tested.
Measured against all 31,338 live designations. Customer keys producing any match:
5 of 17 before, 4 after, and the residue is name-shaped — "wallet-1" contains
Wallet and there is a designated SMILE WALLET; "ivan-petrov-42" contains a real
Russian name. No matcher can know those are keys, which is why the input fix
matters. True positives lost: 0 of 10 — Putin, bin Laden, Kim Jong Un, al-Assad,
Kadyrov, Maduro, Khamenei, Lukashenko, Prigozhin and Sechin all still match at
1.000. Twelve mutants, all killed.
Co-authored-by: Hanzo Dev <[email protected]>
History.Append had no caller. The collection existed after the previous commit and
stayed empty, so every window query returned nothing and every aggregate rule
evaluated to false — not to an error. That is the worse failure of the two: a rule
reporting "no velocity" for a window it never read looks exactly like a customer
with no velocity, and nothing anywhere says otherwise.
Measured on nine deposits of 9,400 against a 10,000 limit, before and after.
Before, every one of the nine returned an identical verdict — score 0.075, four
alerts — because the only rules that could fire were per-transaction checks. The
ninth deposit was indistinguishable from the first. After:
tx 1-2 score 0.525 near-threshold only
tx 3 score 1.0 structuring-day fires at three
tx 5-9 score 1.0 structuring-week fires at five, seven alerts
with currency-day at 0.45 on the day's aggregate of 84,600. That is the typology
this product exists to detect, and it was undetectable end to end until now.
The append happens before evaluation, and the order is the point: a transaction
absent from the window it is judged against is judged on a history that stops one
short of it, so the deposit completing a pattern is the one excluded from seeing
it. A write failure refuses the transaction rather than evaluating it against an
incomplete history.
Co-authored-by: Hanzo Dev <[email protected]>
Ingest defaulted the customer's name from its identifier, so an opaque database
key was screened against the designation lists. Measured against the live lists,
"cust-1" scored 0.902 against "BRANCH 1 OF THE SHIRAZ REVOLUTIONARY COU" — above
the 0.85 reporting threshold — and an ordinary $100 payment came back blocked.
A handle cannot produce a true positive against a list of people's names, so every
match it yields is false, and this one was the most expensive kind of false: the
rule's action is block, so the engine declined a legitimate payment and named a
sanctions hit as the reason.
The customer now comes from the caller, who is the only party that knows it. With
no name supplied the customer-screening rule's own guard holds and the rule does
not fire, which is the honest outcome — a customer whose name this service has
never been given has not been screened, and the way to screen them is to send the
name rather than to invent one from a key. The catalog already publishes this as a
gap.
Measured end to end on a benign $100 payment, against 31,338 loaded designations:
six alerts at a saturated score before this arc, then two, then blocked, and now
allow / score 0 / no alerts. A designated party still matches — "Vladimir Putin"
returns 28 hits, exact on OFAC, EU and OFSI.
Also here: a second defect this exposed and did not fix. The matcher scoring a
six-character single token at 0.902 against a six-word organisation name is wrong
on its own terms, whatever is fed to it. The single-token asymmetry does not hold
on this path. Left as a named finding rather than a hurried threshold change,
because raising the threshold to hide it would suppress genuine transliterations,
which score lower than this false positive did.
Co-authored-by: Hanzo Dev <[email protected]>
The catalog serves the rule set actually installed, not a document describing one,
so the coverage claim is the running configuration and cannot drift from it.
The published gap list is in the same response deliberately. A catalog that shows
only what is covered invites the reader to assume the rest is covered too, and the
reader here is a supervisor. Naming what is not detected, with the citation that
requires it, is the difference between a coverage claim and a coverage assertion.
Co-authored-by: Hanzo Dev <[email protected]>
The collection every window query reads did not exist, and nothing created it.
No migration ran, no bootstrap existed, and the schema in migrations/0001_core.sql
was referenced from nowhere.
The effect was the whole aggregate plane. A window query against a missing
collection fails, a rule that cannot read history reaches no verdict, and twelve
of twenty rules therefore faulted on every transaction — currency reporting,
three structuring rules, pass-through layering, velocity, deviation, dormancy,
concentration and round-amount detection among them. Measured on a benign $100
payment before and after: twelve alerts, then two.
Ensure creates it and is idempotent, so it runs on every start. It lives beside
the reader rather than in a migration file because the field names are shared
between them, and a migration that drifts from them does not fail — it returns an
empty window, which reads as a customer with no history rather than as a broken
deployment.
The indexes follow the query, not the fields: every window is read as
(org, subject, time), so there is one index per subject axis. Without them each
velocity rule is a table scan per rule per transaction, which is the cost model
the sliding-window design exists to avoid.
The two remaining faults on a benign payment are the sanctions rules, which
correctly report that no list has loaded yet rather than reporting a clean screen.
Co-authored-by: Hanzo Dev <[email protected]>
Two branches had independently rebuilt the same two packages, which is one
implementation too many. This resolves them into one.
Their side is the base for detection and screening, because it is better:
- Rules name evidence, and SetRules refuses a set whose evidence this
deployment cannot supply — so the eleven stub helpers that silently disabled
thirteen rules cannot exist. A rule that cannot fire is worse than an absent
one, because it still appears in the catalog a regulator reads.
- One screening store, filled by the refresh and read by the endpoints. Two of
them is why sanctions search answered "no match" for every name while
designations were being loaded somewhere else.
- Parsers for all four publishers against their real files, recovering what the
single-parser path could not read at all: OFAC's digital-currency addresses,
weak aliases held apart from screening candidates, and OFSI's row-per-name
layout collapsed to subjects.
- Name matching over runes, not bytes. Two distinct Han names scored 0.867
byte-wise, above the 0.85 threshold, and 0.700 correctly.
- Deviation from the median and MAD rather than the mean, because a monitoring
baseline contains the outliers it is used to hunt.
- A rule vocabulary with providers behind every term, in place of expressions
reaching into a struct.
Kept from this side, because their branch did not have it:
- The tenant comes from an authenticated identity, not from a request header.
- A rule that could not run carries no weight. Their action() already capped a
failing rule at review, which is right, but the hit still entered the score
at full weight — so a broken rule saturated every transaction's score and the
queue stopped ranking. Faults are now excluded from the aggregate while still
raising their own alert and still forcing review.
- Per-source readiness: which list, how many designations, how long ago, and
the reason if the last attempt failed. Their store answers whether screening
can be relied on at all; this answers what is behind it, which is what an
operator can act on.
- Behavioural scoring, retention, tokenisation, replay and reference data.
Conversion now refuses an unknown currency instead of passing the amount
through. Passing it through understates every currency worth more than a dollar,
so a payment three times over a reporting limit tested as sitting under it.
One ordering table for actions (types.ActionRank), one EvalAll signature, one
screening store, one readiness surface. 15 packages, tests and race clean.
Co-authored-by: Hanzo Dev <[email protected]>
The earlier note said ConList answered 404. Measured across all three paths that
is only true for two of them: publishlive/ConList.{xml,csv} are 404, but the
legacy publishlive/2022format/ConList.csv still answers 200 and still carried a
Last Updated of 03/06/2026 — months after OFSI closed the list.
That distinction matters to anyone reading this file to decide a source. A live
blob the publisher has stopped documenting is worse than a dead one, because it
can stop at any time and nothing downstream would say so.
Co-authored-by: Hanzo Dev <[email protected]>
The refresh filled a Base-backed sanctions store and the search endpoint queried
a separate empty in-memory one. Neither was reachable from the other, so
/v1/aml/sanctions/search returned no matches for every name, permanently, and
nothing in the system reported it. There is now one store, shared by the
endpoints and by the engine's screening term.
Screening refuses rather than reports clean. An unloaded or stale list set, a
list that parsed to zero entries, or a class the deployment cannot answer all
produce an error, which the engine turns into a review. Political exposure is
refused outright: there is no authoritative public list of politically exposed
persons and none of the four sanctions publishers is one, so answering from
sanctions data would report every customer as not politically exposed.
Each publisher is fetched and parsed by its own parser, and a failure for one
leaves the other three loaded rather than clearing them.
Transactions carry a direction and the customer's name and jurisdiction. The
previous ingestion path synthesised a customer from the identifier alone, so
every screening and jurisdiction rule evaluated against an empty name and an
empty country and could not fire whatever the engine did.
Case events check that the case belongs to the requesting organisation. A case
identifier is returned to whoever submitted the transaction, so possession of it
was authority to write on another tenant's investigation.
Health reports function rather than process: whether the lists are ready, which
loaded when, and how many terms this deployment can answer. Currency rates and
the higher-risk jurisdiction listing are loaded, never compiled in.
/v1/aml/catalog publishes the rules with their citations, the obligations claimed,
and the obligations outstanding.
Screening 31,338 real designations across the four published lists produced two
false positives, both a one-word list entry matching an unrelated three-part name
on a shared prefix. Raising the threshold could not separate them from genuine
transliterations, which score lower; the asymmetry could. A single-token name now
has to match a multi-token name near-exactly. Zero false positives after.
Co-authored-by: Hanzo Dev <[email protected]>
The memory bound in Config was an arithmetic estimate; it is now the figure a
test measures (336 KB per tenant at the defaults, 84 MB at MaxOrgs=256), so the
number an operator sizes against and the number the code enforces are the same
one.
LLM.md gains the reasoning a later reader needs and cannot recover from the
code: why half-space trees rather than isolation forest or robust random cut
forest, the two places this implementation departs from the paper and what each
departure buys, why attributability rather than accuracy decided the choice, and
what the appetite governs. The departures matter most — averaging clamped
per-tree votes instead of summing raw masses is the difference between a score
that ranks and a score a quantile can be cut from, and a reader who assumes the
paper's aggregation would reintroduce the fault.
Co-authored-by: Hanzo Dev <[email protected]>
The test asserted that two domains mint different tokens, which the domain prefix
guarantees on its own — so it passed with the per-domain key derivation removed
and one key minting every domain. It now compares the mac behind the prefix,
which is the thing the key schedule is responsible for.
Found by mutating the HKDF info to a constant: the mutant survived, which is the
only way to learn that a test passes for the wrong reason.
Co-authored-by: Hanzo Dev <[email protected]>
The defect that hid two dead endpoints for months was not that they were dead. It
was that nothing could be asked whether they were alive. Screening failure is
silent by construction: an empty list matches nobody, and neither does a clean
party, so "no designations matched this payment" is the same answer either way.
Monitor keeps the outcome of the last refresh of each list — entry count, load
date, payload digest, and the reason if the last attempt failed. A failed attempt
keeps the previous successful load rather than erasing it, because that data is
still what screening runs against; but it ages out instead of staying fresh
forever on the strength of one old success. A list that has never loaded is
reported as unfit, not omitted.
Ready is all-or-nothing on purpose. Screening against three of four lists is not
three-quarters of a control: a party designated only on the missing list passes
completely, and the result is indistinguishable from a clean one.
GET /v1/aml/sanctions/sources serves this and answers 503 when any list is
unfit, so a readiness probe and the dashboard read one truth instead of each
deciding for itself. The dashboard panel leads with whether the instance can
screen at all, above the case counts — a queue length means nothing if the lists
behind it are empty — and states in words that a result against unfit lists
cannot be read as clean.
Six mutants killed, including the two that matter most: treating a zero-entry
load as fresh, and letting three-of-four report ready.
Co-authored-by: Hanzo Dev <[email protected]>
pkg/api/records_test.go is where the three packages that do not know about each
other get checked together. A transaction that cannot be recorded is refused
rather than processed, and nothing is stored on the way to refusing. A blocked one
is retained as a refusal running from its own five-year clock, naming the rule
that refused it. A retained record reconstructs the transaction and carries no
identifier in the clear. A case cannot close without its retained decision. A
replay over an empty or unreadable history refuses instead of reporting zero
alerts, and writes nothing either way.
It is also the only place *engine.Evaluator is put behind replay.Evaluator, so
the sandbox and the live path cannot drift apart without this failing to compile.
cases: the closure invariant on its own — no retained assessment, no closure, no
closing date.
ui: the rule tester reports the replay (alerts over transactions, and the
false-positive proportion when the period held judged alerts) rather than one
match against a synthetic transaction.
LLM.md: the record plane, what AML_TOKEN_KEY is for, the routes it added, and the
open gaps — persistence first, because a retention ledger that does not survive a
restart is a record-keeping breach.
Co-authored-by: Hanzo Dev <[email protected]>
An ordinary $100 payment came back with six alerts and a saturated score, so the
effective false-positive rate was 100% and the score ranked nothing.
The mechanism was one line of policy. A rule whose evaluation errored was
recorded as a hit with Match=true, on the reasoning that a broken rule should not
silently let a transaction through. The safety instinct is right; the encoding
conflated two different findings. "This transaction exhibits this typology" is
evidence about the transaction. "This rule did not run" is a fault in the system,
which says nothing about the transaction — and because a rule that fails once
fails on every transaction, recording it as a match alerts on all of them.
Evaluate now separates them. A matching rule produces a typology alert that
carries its score. Rules that could not be evaluated produce one alert naming
them, with no score, because a rule that did not run is not evidence of risk. The
transaction is still not cleared: the fault forces review, since a transaction
assessed by only part of the rule set has not been assessed.
Tests cover the reported case directly — six unevaluable rules against a benign
payment, expecting one fault alert, zero score, and a refusal to allow — plus a
benign transaction under a working rule set alerting not at all, and a fault
alongside a real match where only the match scores. Five mutants, including
restoring the original Match=true, all killed.
Co-authored-by: Hanzo Dev <[email protected]>
Twenty rules across twelve typologies, each naming the requirement it implements
with a locator a reviewer can open. Citations were read from the primary text.
Two figures that a coverage claim commonly gets wrong are handled explicitly:
Recommendation 16 is cited by its current title, Payment transparency. FATF
revised and renamed it in June 2025 and still hosts the pre-revised text at a
separate path; a test rejects a citation carrying the old title or that URL.
The suspicious-reporting section carries one threshold, 5,000 dollars, applying
whether or not a suspect is identified. The familiar 5,000-with-suspect and
25,000-without split belongs to the banking agencies' rules in 12 CFR, not here.
A test rejects attributing 25,000 to this section.
Currency reporting cites both the threshold section and the aggregation section,
because the first states the figure and only the second states the business-day
period. Citing only the first is how a point-in-time check comes to be described
as implementing currency reporting.
The crypto transfer rule carries no value floor, because the recast transfer
regulation applies regardless of amount; a test rejects adding one.
Gaps() publishes the requirements located in the primary text that this engine
does not meet — report workflow, filing clock, retention, tipping-off
confidentiality, correspondent banking — each with why and what would fix it. A
test rejects claiming a requirement as both covered and outstanding.
Co-authored-by: Hanzo Dev <[email protected]>
Five gaps that a monitoring product cannot be sold without.
anomaly scores whether a transaction is unusual for the entity that made it,
which is the statutory question — consistency with the customer's own profile —
and one no fixed threshold can express. It complements the rule library rather
than replacing it: rules encode typologies someone has already named, and are
blind to anything nobody has written down. The design is deliberately
attributable: a score can name the features that drove it, which is what a
supervisor asks for and what a neural scorer cannot answer.
retention keeps records for their statutory period and destroys them when it
runs out, including the retained assessment whether or not it produced a report.
token protects direct identifiers at rest while keeping the engine able to
correlate: one key schedule, a deterministic org-scoped pseudonym to index on
and a sealed record for reads that are entitled to it, so a record can be found
by party without holding the party's name in the clear.
replay runs a rule over real history and reports what it would have done before
anyone activates it — the same question as "how many alerts would this raise",
which is what makes a rule change reviewable rather than a live experiment.
reference holds the published data a rule needs but cannot derive: currency
conversion and jurisdiction risk listings, each carrying the date it was current
at, each refusing to answer rather than guess. Passing an unknown currency
through unchanged moves every threshold by the exchange rate — the structuring
rule compared raw notional, so 9,500 units of a minor currency read as a
near-threshold cash amount. A jurisdiction list that answers "not listed"
because it was never loaded is indistinguishable from a clean world.
Co-authored-by: Hanzo Dev <[email protected]>
zeekay
