Interlock

A reference monitor for agent side effects. An agent proposes a plan. It holds no credentials and cannot execute. Interlock applies the plan against a real but uncommitted substrate, measures what actually changed, evaluates deterministic predicates against that measurement, and only then commits or rolls back.

uv run python demo.py

  ACT 2  The identical plan, through Interlock.
    MEASURED by the database, not predicted from the plan:
      rows updated    63   inserted    63   deleted   103
      blast radius   229 rows across 4 tenants
      orders.total $21,350.25 -> $425.25

    ADJUDICATED by 7 deterministic checks:
      BLOCK   blast_radius          plan would mutate 229 rows, limit is 8
      BLOCK   tenant_isolation      plan spans 4 tenants, limit is 1
      BLOCK   no_delete             plan deletes 103 row(s) from ungranted table(s): order_audit
      BLOCK   column_value_guard    orders.total falls 98.0% (21,350.25 to 425.25), limit is 30%
      advise  stated_footprint      agent stated 6 row(s); substrate measured 229 (38.2x)

    OUTCOME  ROLLED BACK
      Byte-identical to before the plan ran. Nothing reached disk.

Why authorization does not catch this

A prompt-injected agent is a fully authorized agent. It is doing something it has permission to do, so an authorization check passes. The injected instruction arrived from the customer's own database, inside a tool result, after every provider-side filter had already run, so a provider-side filter does not see it either. The only remaining place to catch it is between the effect being computed and the effect becoming durable.

That position requires write-path credentials to the database the effects land in, which is why this runs inside the deployment rather than at the model provider.

The demo is also a control for model quality: the agent loops zero times, costs $0.0012, and emits valid SQL on the first attempt. The damage is not a model-quality failure.

demo.py's agent is scripted, not live. injected_plan (a function local to demo.py, not part of the package) and the rest of Act 2's input are hardcoded EffectPlans in demo.py itself; no model is called anywhere in that file. That is deliberate -- it isolates one variable (does the escrow stop this plan?) from another (would a real model produce it?) -- but it means demo.py cannot answer the second question, and an adversarial audit of this project found that a real, current Claude model declines the exact injected instruction shown above, across several realistic framings of the same prompt injection. For the first question answered against a real model instead of a scripted stand-in, see scripts/live_stress_test.py: eight scenarios, a real anthropic.Anthropic client, real governed spend, and verdicts the script itself says are not reproducible, because what the model chooses to do is not.

What it measures

The diff is read from the substrate, never reconstructed from the plan. The two are routinely different:

Agent stated	1 row
SQL rowcount	1 row
Measured	2 rows across 2 tables

A compliance trigger nobody remembered doubled the footprint. No plan inspection, SQL parser or policy engine reading the statement could know that. The database knew, because it had just done it.

Lifecycle

PLANNED -> STAGING -> STAGED -> VERIFIED  -> COMMITTED
                             \-> REJECTED -> ABORTED

These properties hold by construction:

1. No path from STAGED to COMMITTED skips adjudication.
2. REJECTED is not overridable in-process. Overriding means submitting a new plan that carries an explicit waiver, which is then itself a recorded artifact.
3. A committed effect is discoverable from the chain. COMMIT_INTENT is written, and forced to stable storage, before the substrate is told to commit, and COMMITTED after it returns, so a crash in that window leaves an intent with no terminal record after it.
4. That crash is resolved exactly, not guessed. An intent says only that a commit was attempted, so the substrate also writes a marker row, in _interlock_commits, inside the stage's own transaction: it exists if and only if the effects do. On restart, EscrowEngine.recover() reads the marker for every intent a crashed process left open and appends COMMITTED or ABORTED, noted as recovered. EscrowRuntime runs it at startup whenever it has a chain_path. An intent written before v0.1.2, or by a substrate without markers, is left open and logged for an operator: its marker's absence proves nothing. So is one the server has not decided yet (a client that died mid-COMMIT leaves PostgreSQL still committing); a later recover() resolves it.
5. A commit whose answer is lost is resolved the same way, while the process lives. If the connection drops with COMMIT in flight, the engine reads the marker at once: the plan is reported committed if it landed, and rolled back only once the server has ended the transaction without it. Until the server decides, execute() raises CommitUnsettledError and writes no terminal record. Do not retry the plan then: the intent stays open, and recover() settles it.

Quickstart

EscrowRuntime owns the substrate, the record chain, the AgentGov anchor and the engine. Five lines to a governed write:

from interlock import EscrowRuntime, TableSpec

runtime = EscrowRuntime(
    "prod.db",
    tables=[TableSpec("orders", columns=["id", "tenant", "total"], tenant_column="tenant")],
    scope_id="support-agent",
)
result = runtime.execute_sql(
    "UPDATE orders SET total = :total WHERE id = :id",
    {"total": 480.0, "id": 1},
    tenant_id="acme",
    stated_rows=1,
)
print(result.committed, result.blocked_by)

That stages the statement in a real transaction, measures what the database actually did, adjudicates the measurement, and commits or rolls back.

Placeholders must be named. Effect.parameters is a Mapping, so the driver binds by name: write :id, never ?. A positional placeholder is refused when the plan is built, before any lock is taken.

Building a plan

For anything past one statement, PlanBuilder mints the identifiers and the UTC timestamp and wires the dependency DAG:

from interlock import PlanBuilder

plan = (
    PlanBuilder(scope_id="support-agent")
    .update(
        table="orders",
        statement="UPDATE orders SET total = :total WHERE id = :id",
        parameters={"total": 100.0, "id": 1},
        tenant_id="acme",
        stated_rows=1,
    )
    .build()
)

Effects are sequential by default: each depends on the one added before it. Pass independent=True for no ordering constraint, or after=[effect_id] to fan out. builder.last_effect_id hands back the id to depend on.

Under the runtime

The primitives are unchanged and the runtime hands them back (runtime.engine, runtime.chain, runtime.anchor). Wire them yourself when you need to:

from interlock import (
    BlastRadius,
    EscrowChain,
    EscrowEngine,
    PlanBuilder,
    SqliteSubstrate,
    TableSpec,
    TenantIsolation,
)

substrate = SqliteSubstrate(
    "prod.db",
    tables=[TableSpec("orders", columns=["id", "tenant", "total"], tenant_column="tenant")],
)
engine = EscrowEngine(
    substrate,
    checkers=[BlastRadius(50), TenantIsolation(1)],
    chain=EscrowChain("escrow.jsonl"),  # omit this and the chain is in memory only
)

plan = (
    PlanBuilder(scope_id="support-agent", intent="refund order 1")
    .update(
        table="orders",
        statement="UPDATE orders SET total = :total WHERE id = :id",
        parameters={"total": 480.0, "id": 1},
        tenant_id="acme",
        stated_rows=1,
    )
    .build()
)

result = engine.execute(plan)
if not result.committed:
    print(result.blocked_by, result.diff.blast_radius)

execute() takes a per-plan settle_cost= that overrides the engine's default, because what a plan cost to produce is a property of the plan.

The chain file. EscrowChain(path) resumes an existing file: every record is read and verified before the first append, so a restarted process continues the chain rather than starting a second one at sequence 1. One file has one writer. The chain claims <path>.lock for as long as it is open, and a second writer, in this process or another, gets ChainInUseError: two writers would fork the chain, and recovery would resolve one writer's in-flight commits as if it had died. EscrowChain.load(path) reads and verifies a live chain without claiming it, and cannot append. A final line left torn by a crash mid-append is cut off by the next writer, because that append never returned. Intents and outcomes are fsynced before append returns; pass fsync=False to give that up for speed.

Typing

PlanId and EffectId are NewTypes over str. PlanBuilder mints both, so most code never names them. Under mypy --strict, constructing an EffectPlan or Effect by hand means importing them:

from datetime import UTC, datetime

from interlock import Effect, EffectId, EffectKind, EffectPlan, PlanId

plan = EffectPlan(
    plan_id=PlanId("plan-0001"),
    scope_id="support-agent",
    trajectory_id="traj-0001",
    created_at=datetime.now(UTC),
    effects=(
        Effect(
            effect_id=EffectId("eff-0001"),
            kind=EffectKind.UPDATE,
            target="orders",
            statement="UPDATE orders SET total = :total WHERE id = :id",
            parameters={"total": 480.0, "id": 1},
            tenant_id="acme",
            stated_rows=1,
        ),
    ),
)

Invariants

Checkers are pure functions of (plan, diff). No I/O, no clock, no RNG, no model call. A checker that reads the world cannot be replayed, and a verdict that cannot be replayed is not evidence. Where a judgment genuinely needs a model, the model is a sensor whose output is passed in at configuration time, never consulted from inside check. A checker that raises becomes a blocking violation: fail closed.

Checker	Bounds
BlastRadius	captured mutations in one plan
TenantIsolation	distinct tenants one plan may span
ColumnValueGuard	fall in a column total, table-wide
TenantDrawdownGuard	fall in a column total, per tenant
NoDelete	deletions outside granted tables
NoSchemaChange	effects declared as DDL
TruncationGuard	a diff that hit the row cap
StatedFootprint	the agent's claim against the measurement

TenantDrawdownGuard exists because ColumnValueGuard sums the column across the whole diff, and a sum is the wrong denominator on a multi-tenant substrate. A plan that zeroes one tenant while inflating another by the same amount moves that sum by 0%, so a table-scoped guard passes it — with a tenant wiped. Grouping by tenant first is the only way to see it. Column totals are Decimal, computed through Decimal(str(value)), so a guard permitting a 30% drawdown does not admit 30.000000000000004%.

Every threshold in default_checkers is an uncalibrated placeholder. Measure your own diffs and set the limits from the measurement.

Errors

Every failure on the write path is an InterlockError. The AgentGov seam is an implementation detail of the anchor, and its exceptions are translated at the boundary, so this is sufficient:

from interlock import EscrowRuntime, InterlockError, TableSpec

runtime = EscrowRuntime(
    "prod.db",
    tables=[TableSpec("orders", columns=["id", "tenant", "total"], tenant_column="tenant")],
    scope_id="support-agent",
)

try:
    runtime.execute_sql(
        "UPDATE order_audit SET note = :note WHERE id = :id",
        {"note": "x", "id": 1},
        table="orders",  # the label lies; the statement writes elsewhere
        tenant_id="acme",
        stated_rows=1,
    )
except InterlockError as exc:
    print(type(exc).__name__, exc)

That example is the case worth knowing about: table= is a label the agent writes, and the statement goes somewhere else entirely. The substrate denies the write inside SQLite before it executes, and it surfaces as a ForbiddenStatementError like everything else on this path.

Refusals: two audiences

A refusal has two readers, and they must not get the same record. The operator needs everything: which tenants the plan reached, the exact totals, the checker's own message. The agent must get none of it. A guard's message is data about other tenants, and an injected agent reading refusals is reading a side channel.

So every result carries both. StageResult.refusal.evidence is the operator's record. StageResult.feedback (and exc.feedback on any error execute raises) is what the agent may be told:

from interlock import EscrowRuntime, TableSpec, TenantIsolation

runtime = EscrowRuntime(
    "prod.db",
    tables=[TableSpec("orders", columns=["id", "tenant", "total"], tenant_column="tenant")],
    scope_id="support-agent",
    checkers=[TenantIsolation(1)],
)
result = runtime.execute_sql("UPDATE orders SET total = 0", table="orders", tenant_id="acme")

assert not result.committed
print(result.feedback.render())  # for the agent
# The plan was refused and rolled back; nothing changed.
# - tenant_isolation: the plan changed rows of more tenants than one plan may
#   (at most 1); its declared tenants are acme. Confine every statement to its
#   declared tenants.
assert "globex" in result.refusal.evidence.tenants_touched  # for the operator

Feedback is built so that it cannot carry more than the agent already knew:

• it names only tables the plan's effects name, and tenants they declare. Anything else is "a table the plan does not name", never named or counted;
• row counts are buckets (0, 1, 2-9, 10-99, 100-999, 1000+);
• it carries no aggregate at all: no column total, no fraction of one. The only other number is a built-in checker's configured limit, as a whole percentage;
• its text comes from fixed templates, and its constraints come in a canonical order, never the order the violations came in.

A checker contributes a typed hint per violation, and the hint is sanitized against the plan field by field. A custom checker's hint can name the plan's own tables and tenants and pick a kind of guidance. Its numbers and columns are dropped, because nothing can tell a row count from a total. Errors are mapped by type, never by message, since a database error can quote another tenant's row. Property tests check, over arbitrary multi-tenant diffs and adversarial checkers, that renaming anything the plan did not name, scaling every amount, or changing other tenants' values leaves the feedback byte-identical (see tests/test_feedback.py).

Repair: the part that would pass

A refused plan is often one step away from an acceptable one: nine corrections to one tenant's orders and a tenth that reaches another tenant. Refusing all ten is right; telling the agent which nine would pass turns a halt into finished work. repair() finds the largest part of a plan that would be admitted, by experiment. It stages the plan once and, inside that one stage, tries candidate sub-plans in savepoints: each is run, measured and adjudicated by the same checkers, then rolled back.

from interlock import EscrowRuntime, TableSpec, TenantIsolation

runtime = EscrowRuntime(
    "prod.db",
    tables=[TableSpec("orders", columns=["id", "tenant", "total"], tenant_column="tenant")],
    scope_id="support-agent",
    checkers=[TenantIsolation(1)],
)
plan = (
    runtime.plan(intent="apply ticket 9001 corrections")
    .update(
        table="orders",
        statement="UPDATE orders SET total = 450 WHERE id = 1",
        tenant_id="acme",
        effect_id="reprice_1",
        independent=True,
    )
    .update(
        table="orders",
        statement="UPDATE orders SET total = 250 WHERE id = 2",
        tenant_id="acme",
        effect_id="reprice_2",
        independent=True,
    )
    .update(
        table="orders",
        statement="UPDATE orders SET total = 0 WHERE id = 3",  # globex's order
        tenant_id="acme",
        effect_id="reprice_3",
        independent=True,
    )
    .build()
)
assert not runtime.execute(plan).committed

repair = runtime.repair(plan)  # advisory: nothing is committed here
print(repair.feedback.render())  # for the agent
# Resubmit the largest part of the plan that would be admitted: keep reprice_1,
# reprice_2; drop reprice_3.
# - reprice_3: tenant_isolation: the plan changed rows of more tenants than one plan
#   may (at most 1); its declared tenants are acme. Confine every statement to its
#   declared tenants.
assert runtime.execute(repair.proposal).committed

• Advisory, and re-adjudicated. The repair stage is always rolled back. The proposal is a new plan whose repair_of names the refused one, and it commits only through execute(), staged and adjudicated again from scratch, because the database may have moved in between.
• Exactly as proposed, and once. A plan naming itself a repair is admitted only if this engine's chain holds a REPAIR_PROPOSED record with its exact content hash, and only until it has committed. A widened or forged "repair" is refused.
• No monotonicity assumed. A debit alone fails a drawdown guard that the debit and its balancing credit pass together. A search that dropped whatever fails on its own would keep a lone credit; this one keeps both and drops only the step that is wrong.
• Runnable candidates, largest first. A step is kept only with everything it depends on. The whole plan is tried, then every such sub-plan one step smaller, and so on, later steps dropped first within a size; the first admitted one is a largest, and repair.exhaustive says so. max_trials (default 32) bounds the search; when it runs out, a greedy pass keeps each step whose addition is admitted, which is sound but may not be largest. The stage's time bound ends the search too.
• Explained without leaking. Each dropped step says why, from its own trial: the constraints that refused it, sanitized like every other feedback, a dropped dependency, or a statement the substrate refuses outright.

Both substrates have savepoints. A repair's trials all run inside one stage, so the stage's locks are held for the whole search.

AgentGov integration

Interlock imports agentgov as a read-only audit dependency. The dependency runs one way (interlock -> agentgov) and requires no change to AgentGov.

Every adjudication is written to a hash-linked chain whose records carry the AgentGov head hash observed at the time. That proves a lower bound on a record's time. It does not prove an upper bound on its own, because a record could name a stale head, so verify_anchors() additionally requires the observed AgentGov sequence to be non-decreasing across an append-only chain.

The breaker is read as it is now, immediately before commit. Attached read-only (audit_path=, the separate-process deployment), Interlock refreshes its view of the governor before every record and before the pre-commit breaker check, verifying everything the governor wrote since, so a halt written after Interlock attached stops the commit. A view that fails verification refuses to operate. With a governed manager in the same process, the governor's own lock is held from that check through the substrate's commit, so a trip lands strictly before it (and the commit is refused) or strictly after it. Across processes no lock spans both databases: a trip the governor commits between the check and the commit cannot be excluded, and when one happens the COMMITTED record says so.

Where Interlock is co-resident with a write-capable governor it also writes its own chain head into AgentGov's chain: into a zero-value ANCHOR entry, or, when the plan carries a cost, into the memo of the spend that settles it. memo is inside AgentGov's hash payload, so the reverse anchor cannot be altered without breaking AgentGov's own verification. The two chains then interlock in both directions.

Reverse anchoring is free. settle_cost defaults to "0", which writes a zero-value ANCHOR entry and charges nothing. Pass the plan's real cost and the anchor rides on the spend that settles it instead:

from agentgov import BudgetManager, money

from interlock import EscrowRuntime, TableSpec, TenantIsolation

governor = BudgetManager()
governor.open_root("support-agent", money("8.00"))

runtime = EscrowRuntime(
    "prod.db",
    tables=[TableSpec("orders", columns=["id", "tenant", "total"], tenant_column="tenant")],
    scope_id="support-agent",
    checkers=[TenantIsolation(1)],
    governed=governor,  # enables reverse anchoring
    settle_cost="0.0002",  # what this plan cost to produce
)

result = runtime.execute_sql(
    "UPDATE orders SET total = :total WHERE id = :id",
    {"total": 480.0, "id": 1},
    tenant_id="acme",
    stated_rows=1,
)
print(result.committed, bool(result.anchored_to))

scope_id must name a scope AgentGov knows about. That is checked at admission, not at commit, because the reverse anchor is written after the substrate commits and a failure there would arrive on top of durable effects.

Every failure on this seam is an InterlockError. agentgov exceptions are translated at the anchor boundary, so a caller never has to import agentgov.exceptions to catch a write-path error.

AgentGov's latching breaker is re-read immediately before commit, not at admission. The staging window is where a trip has to be observed, because a halt that does not stop in-flight side effects does not stop the effects.

Settling with the commit: claim and settle (PostgreSQL)

Across processes, the governed mode above leaves two gaps: a trip another process commits between the breaker check and the commit, and a crash between the commit and the reverse anchor, which leaves effects with no settlement. Both close when the ledger lives in the same PostgreSQL database as the observed tables (BudgetManager.open_postgres, agentgov v0.3):

from agentgov import BudgetManager

from interlock import EscrowEngine, LedgerAnchor, PostgresSubstrate, TableSpec, default_checkers

governor = BudgetManager.open_postgres("postgresql://owner@db/app")  # the tables' owner
governor.store.grant_join("interlock_agent")  # once: the stage role may claim, and no more
engine = EscrowEngine(
    PostgresSubstrate(
        "postgresql://interlock_agent@db/app",
        tables=[TableSpec("orders", columns=["id", "tenant", "total"], tenant_column="tenant")],
    ),
    checkers=default_checkers(row_limit=8),
    anchor=LedgerAnchor(governed=governor, same_transaction=True),
    settle_cost="0.25",
)

Each plan is settled in three steps:

1. Reserve. Before the stage opens, the governor places a hold for the plan's cost, in its own transaction. A scope that is halted or cannot pay refuses the plan here, before anything is staged, and an overdraft trips its breaker.
2. Claim. At commit the stage takes the ledger's writer lock inside its own transaction, the governor catches up with every other governor and checks the breaker, and a settlement claim keyed to the hold is written into the stage. The claim commits in the stage's COMMIT, with the effects and the commit marker, or not at all, and no trip in any process can land between the check and the commit.
3. Settle. After the commit the claim is redeemed into AgentGov's chain: the hold's release and the spend, which names the plan's COMMIT_INTENT record.

The claim reads nothing of the chain, so the stage's REPEATABLE READ snapshot, however many ledger commits old, conflicts with nothing: a busy ledger never stops a plan from committing. Measured with four engines racing for one account while a governor wrote the ledger back to back, every plan committed, and every retry was the engines' own race for the account's row, as many as with no ledger traffic at all.

No crash can leave effects without their claim, or a claim without its effects. A process that dies after its commit leaves the claim pending, and recover() books it (any governor may book any pending claim: one exists only for a stage that committed). A plan that did not commit leaves at most its hold, which recover() releases once the chain says the plan can no longer commit. tests/test_crash_consistency.py runs every kill point against this mode, including after the hold, mid-claim, and after the commit but before the claim is booked.

• The stage's role cannot write the ledger or claim anything. It gets no privilege on the ledger's tables, only EXECUTE on one SECURITY DEFINER function that records a claim, and only with a token derived from a secret the role cannot read and from the transaction's own id. A statement in an agent's plan runs as that role, and can neither write the ledger, forge a claim, nor reuse a governor's token.
• Booked, a settlement stands whatever happened since. A claim whose hold an operator voided is booked as a spend on its own; one that overdraws the scope is booked and trips the breaker. Its effects already happened.
• A refused plan is charged against its hold, after its stage, as before, naming its terminal record. A plan whose stage fails releases its hold.
• A stage stuck in COMMIT holds the ledger until the server decides: no governor can write meanwhile, bounded by the stage's lock_timeout.

Receipts

With a ReceiptIssuer, every plan the engine adjudicates, committed or refused, gets a signed ARC1 receipt in an agentgov ReceiptLog (the format is agentgov's docs/RECEIPTS.md). Anyone holding the issuer's key can then check offline what the agent was allowed to do, what it said it would do, what it measurably did, what was decided and what it cost:

from agentgov.receipts import HmacKey, ReceiptLog, verify_bundle

from interlock import EscrowRuntime, ReceiptIssuer, TableSpec, TenantIsolation

key = HmacKey.generate()
log = ReceiptLog("support-receipts", key, path="receipts.jsonl")
runtime = EscrowRuntime(
    "prod.db",
    tables=[TableSpec("orders", columns=["id", "tenant", "total"], tenant_column="tenant")],
    scope_id="support-agent",
    checkers=[TenantIsolation(1)],
    receipts=ReceiptIssuer(log),
)
result = runtime.execute_sql(
    "UPDATE orders SET total = 480 WHERE id = 1", table="orders", tenant_id="acme"
)
receipt = result.receipt
print(receipt.outcome.status.value, receipt.effect.row_count)  # committed 1

bundle = log.bundle(log.index_of(receipt.receipt_id), log.checkpoint())
assert verify_bundle(bundle, issuer_key=key).passed

What goes into one:

• Authority: the AgentGov scope path when an anchor is attached, the trajectory, and a capability naming the observed tables and the tightest BlastRadius limit.
• Intent and effect: the plan's content hash and stated footprint; the diff hash the chain's DIFF_COMPUTED record carries; a salted commitment over every measured row change, from which one row can be disclosed later without the rest; and a hash of the observed schema, foreign keys included.
• Coverage: the observed tables, and every gap in words: an acknowledged cascade, a substrate that did not read the foreign keys or does not refuse unobserved tables.
• Decision: the verdict hash, each checker with a digest of its configuration, the issuer's policy epoch, and, for a repair, the refused plan's receipt as repair_of.
• Cost: the AgentGov transaction that settled the plan. The receipt is issued last, after the reverse anchor, so agentgov's verifier can check the cost against the ledger.

The receipt and the escrow chain name each other: the stage's terminal record carries the receipt id, and the receipt's anchors.escrow names that record by sequence and hash. A receipt that cannot be issued after a commit is logged, not raised; the effects stand, and the terminal record names the receipt that should exist. An HMAC receipt verifies only for a holder of the key: for third parties, sign with agentgov's Ed25519Signer (agentgov[sign]) and publish the public key.

Recovery

A halt is correct, and it used to be the end of the task. RecoveryRuntime gives a halted task a bounded way to finish instead: a few more model calls, each under tighter constraints than the one before, paid from a reserve set aside for it, and each recorded and signed before it is made.

from decimal import Decimal

from agentgov import BudgetManager
from agentgov.cognitive import CognitiveBreaker, CognitivePolicy
from agentgov.exceptions import AgentThrashingError
from agentgov.receipts import HmacKey

from interlock import (
    Directive,
    RecordLog,
    RecoveryPolicy,
    RecoveryRuntime,
    ToolRevokedError,
    Trip,
    TripKind,
)

governor = BudgetManager()
governor.open_root("org", "10.00")
governor.delegate("org", "support-agent", "5.00")
policy = RecoveryPolicy(
    reserve=Decimal("0.50"),  # carved out of support-agent's envelope
    step_estimate=Decimal("0.05"),  # held for each recovery call
    directives=(
        Directive(
            "no-repeat",
            "Do not repeat a tool call that has already returned; use the results you have.",
            frozenset({TripKind.THRASHING}),
        ),
    ),
)
records = RecordLog(HmacKey.generate(), log_id="support")
runtime = RecoveryRuntime(
    governor, "support-agent", policy, records, tools=["lookup", "apply_plan"]
)

transcript = [
    {"role": "user", "content": "What is the total of order 1?"},
    {
        "role": "assistant",
        "content": [{"type": "tool_use", "id": "toolu_02", "name": "lookup", "input": {"id": 1}}],
    },
]
breaker = CognitiveBreaker(
    policy=CognitivePolicy(max_identical_repeats=2), observer=None, manager=governor
)
try:
    for _ in range(2):  # the same lookup, twice: halted before the second runs
        breaker.observe_call("support-agent", "lookup", kwargs={"id": 1})
except AgentThrashingError as exc:
    trip = Trip.of(exc, tool="lookup")

step = runtime.recover(trip, transcript, max_tokens=4096)
print(step.rung, step.granted_tools, step.revoked_tool)  # revoke_tool ('apply_plan',) lookup
# Send step.messages (through the beta endpoint with step.betas when non-empty,
# the plain one otherwise) and step.max_tokens, the system prompt and the tools
# as before; settle what the call cost; refuse a revoked tool where you run it:
runtime.settle(step, "0.03")
try:
    runtime.check_tool("lookup")  # the tool the halt named, and this step revoked
except ToolRevokedError:
    pass  # refuse the call here instead of running it
runtime.close()
records.verify()
print(governor.is_halted("support-agent"), len(records))  # True 4

The ladder is fixed and deterministic, one rung per step: revoke the tool the halt names, then an operator directive from the policy's allowlist, then a lower token ceiling, then guidance in fixed words. Constraints come before conversation. Each step takes the first rung that can still do something for this kind of halt, and tightening only accumulates: a revoked tool stays revoked and the ceiling stays down. When no rung is left, or max_steps is reached, recover() raises RecoveryExhaustedError and the halt stands. The same halts always take the same steps. A refused plan is left to guidance by default, so the agent can correct it; its guidance is the refusal's sanitized feedback, or a repair's (see Repair).

History is never edited. A step returns the transcript it was given, as the same objects, with messages appended: error results for tool calls the halt stopped, then the rung's notice. The system prompt and the tool definitions are never touched. On Claude Opus 5, Opus 5.5, Opus 4.8, Fable and Mythos, a directive goes in an appended {"role": "system"} message and a revocation in a tool_removal block (beta inline-tools-2026-09-15, in step.betas), the operator channel a user turn cannot forge; Channel.USER uses user-turn notices for every other model. Either way, check_tool() refuses a revoked tool where the harness runs it. Each step checks that the transcript it is handed extends the one the last step returned, so an edit in between is refused, and its record pins the transcript before and after by a running hash (transcript_head).

A non-empty step.betas has to go through the beta endpoint, not the plain one: with the anthropic Python SDK that is client.beta.messages.create(..., betas=step.betas), never client.messages.create(betas=...) (which raises a plain TypeError — that constructor does not take a betas keyword at all). step.betas is empty under Channel.USER, so a harness that only ever runs on a Channel.USER model never needs the beta endpoint. TOOL_CHANGES_BETA (interlock.recovery) names the one exact string this release sends; treat it as a snapshot of Anthropic's current API surface, not a permanent constant — an audit against live traffic already caught it drifting once (see the docstring), and nothing here detects the next rename automatically.

The reserve sits beside the scope, not under it. A trip halts the tripped scope's whole subtree, and recovery exists for a tripped scope, so {scope}/recovery is carved out of the scope's own envelope (released to its parent and delegated from there). The halted scope stays halted; every call made during recovery is billed to the reserve, a step's through recover() and settle(), the task's own through hold() and capture(). A root scope's reserve is a new root funded from the treasury, or RecoveryPolicy(funding=...) names a scope to delegate it from. close() returns what is left to the parent.

Every act is signed first. Opening, each step, each settlement and the close are RecordLog entries: ILOK1, agentgov's canonical JSON under an ILOK1/record/v1 signing prefix, hash-linked, and anchored into the AgentGov ledger as ANCHOR entries, so check_anchors() catches a log truncated or rewritten after the fact. A step's record names the halt (and a refused plan's ARC1 receipt), the rung and exactly what it did, the hold that pays for it, the policy's digest, and the transcript before and after, and is written before the step is returned. Persist the log (RecordLog(..., path=...)): a runtime restarted over it adopts the recovery scope and replays what was revoked and lowered, so a restart never loosens a constraint.

Extension quotes

AgentGov's backstop is blunt on purpose: an authorization that would overdraw a scope trips its breaker, and so does a capture that spends it to zero. For a runaway loop that is the answer. For a task that is only more expensive than its envelope, it throws away the work done. BudgetGuard.authorize() asks first: when a call will not fit, it returns a signed ExtensionRequest instead of placing a hold that would trip the breaker.

from agentgov import BudgetManager
from agentgov.receipts import HmacKey

from interlock import BudgetGuard, ExtensionRequest, Milestones, RecordLog

governor = BudgetManager()
governor.open_root("org", "10.00")
governor.delegate("org", "support-agent", "5.00")
guard = BudgetGuard(governor, RecordLog(HmacKey.generate(), log_id="support"))

held = guard.authorize("support-agent", "4.60")  # fits: an ordinary hold
governor.capture(held, "4.60")

done = Milestones(done=3, total=5, unit="tickets")
quote = guard.authorize("support-agent", "0.50", milestones=done)
assert isinstance(quote, ExtensionRequest)  # did not fit: quoted, and nothing tripped
print(quote.requested, governor.is_halted("support-agent"))  # 3.44 False
print(quote.render())
# Extension requested for support-agent: 3.44000000.
# Spent 4.60000000 (support-agent 4.60000000); held 0.00000000; 0.40000000 left; ...
# Proof of work: 0 plan(s) committed (0 rows); declared 3 of 5 tickets done.
# Estimate: 1.53333333 per unit x 2 remaining = 3.06666667, plus 25% margin = 3.84000000 ...

grant = guard.grant(quote, approved_by="ops@example.com")
print(grant.scope_id, governor.available(grant.scope_id))  # support-agent/ext-1 3.84000000

A quote has three parts, and is a signed extension.quoted record anchored in the ledger:

• Spend to date, from the ledger: what the task's scopes (the scope, its recovery scope, its extensions) settled net of refunds, what they hold, and what is left.
• Proof of work: with BudgetGuard(receipts=...), the ARC1 receipts of the plans the task committed and the rows they measurably changed, with a checkpoint the receipt log signed for the quote, so each receipt is provable by inclusion (log.bundle(log.index_of(id), checkpoint)); the recovery steps it took; and any milestones the harness declares, which are carried as declared, not proven.
• An estimated completion cost, by a named method: cost per declared milestone times those remaining, or with none declared, the call that did not fit; plus a margin, rounded up to the cent. requested is that, less what the scope has left.

A quote is answered once, before it expires, by the guard that issued it: grant() or decline(), each a signed record naming the quote. A root scope is topped up in place, and if the money running out is what tripped it, its breaker is reset. A delegated scope cannot be topped up, so the grant delegates {scope}/ext-N beside it, carries along what the old scope had left, and the task bills there from then on. A grant the parent cannot make changes nothing. A scope halted for any other reason is not quoted: authorize() raises AgentGov's CircuitOpenError, because more money is not the answer to a safety halt.

A RecoveryRuntime given the guard quotes for its own reserve the same way: when a step will not fit, RecoveryExhaustedError.quote carries the request, and extend() takes up the grant. Quotes are operator evidence; nothing in them is sent to the agent. Granting is the operator's act: the guard records approved_by but cannot authenticate it, so put grant() behind the approval you already trust, and never within the agent's reach as a tool.

PostgreSQL

PostgresSubstrate stages each plan in a REPEATABLE READ transaction and measures it with row triggers installed once, so no stage ever alters a busy table. Install as the tables' owner, then stage as a separate role granted DML on the observed tables and nothing else:

uv pip install 'interlock[postgres]'
interlock install --config interlock.toml --database postgresql://owner@db/app
interlock check   --config interlock.toml --database postgresql://interlock_agent@db/app

substrate = "postgres"
schema = "public"
stage_roles = ["interlock_agent"]   # granted the interlock schema's stage functions
acknowledge_cascades = []

[[tables]]
name = "orders"
columns = ["id", "tenant", "total"]
tenant_column = "tenant"

from interlock import EscrowEngine, PlanBuilder, PostgresSubstrate, TableSpec, default_checkers

substrate = PostgresSubstrate(
    "postgresql://interlock_agent@db/app",
    tables=[TableSpec("orders", columns=["id", "tenant", "total"], tenant_column="tenant")],
)
engine = EscrowEngine(substrate, checkers=default_checkers(row_limit=8))
plan = (
    PlanBuilder("support-agent")
    .update(
        table="orders",
        statement="UPDATE orders SET total = %(total)s WHERE id = %(id)s",
        parameters={"total": 100, "id": 1},
    )
    .build()
)
result = engine.execute(plan)

What interlock install puts in the database: an interlock schema holding the stage table and four functions, and on each observed table one AFTER row trigger plus one TRUNCATE trigger, both ENABLE ALWAYS so session_replication_role does not switch them off. For a transaction that opened no stage, the trigger logs the write to interlock.unmediated (see Unrecorded writes). For one that did, it writes before and after images into a temporary table that exists only in that session for that transaction. Rows are keyed to the stage by pg_current_xact_id(), read from a row only the stage-opening function can write, not from the interlock.stage_id setting, which any statement could change; the setting is still set, and the trigger refuses to go on when the two disagree.

Every stage checks, before its first effect and with the observed tables locked ROW EXCLUSIVE so none of it can change underneath:

• each observed table carries Interlock's triggers, enabled always, capturing exactly its TableSpec's columns, and has no inheritance children or partitions (a statement on the parent writes their rows too, through no trigger, and PostgreSQL checks privileges on the parent alone);
• the stage's role is not a superuser, owns no observed table (an owner can disable a trigger), and can write no other table, directly, through a column grant, or through any role it belongs to. PostgreSQL has no statement authorizer, so the grant is the table boundary; the substrate checks it rather than trusting it. enforce_table_access=False trusts it;
• the cascade check, read from pg_constraint.

What differs from SQLite:

• Placeholders are psycopg's: %(name)s, and %% for a literal percent sign.
• Only row statements stage: SELECT, INSERT, UPDATE, DELETE, MERGE, WITH, VALUES, TABLE. Each is sent as a prepared statement, which the server will not split, so UPDATE ...; COMMIT fails instead of committing before adjudication.
• A cascade gate fires per row. A delete of a gated row, or an update that changes a gated key, is refused inside the statement and PostgreSQL rolls the statement back, cascade included. SQLite refuses when it prepares the statement.
• NUMERIC is read as Decimal, exactly.
• Bounds are the server's: statement_timeout, lock_timeout and idle_in_transaction_session_timeout are set for the stage and re-set before every effect, so a statement cannot lift them, and a stage whose client died cannot hold locks past max_stage_seconds.
• Crash recovery reads the transaction. The stage's row is its commit marker, and pg_current_xact_id() is written into the commit intent, so recovery can tell a transaction the server still has open (a crashed client's, until its session times out, or one still committing when its client died) from one that rolled back, and leaves the first open. A connection lost with COMMIT in flight is answered from the same marker. PostgreSQL stops statement_timeout before running a commit's deferred triggers, so work there is bounded by the stage's lock_timeout and not by its statement bound.

What grants cannot see, so the substrate cannot either: a SECURITY DEFINER function the role may call that writes elsewhere, an extension such as dblink that opens another connection, and large objects. Do not grant them to the stage role. Requires PostgreSQL 14 or later; CI runs 16.

Unrecorded writes

The monitor only sees what goes through it. A cron job, a migration, a DBA at a prompt, or the agent's own credentials used directly all write observed tables and leave nothing in the escrow chain. interlock reconcile-effects is the check that closes the loop: nothing changes an observed table without a record.

interlock install           --config interlock.toml   # once; SQLite needs it for this too
interlock reconcile-effects --config interlock.toml --chain escrow.jsonl [--after N]

It fails (exit 1) on:

• an unmediated write: a row change no stage made. On PostgreSQL the installed trigger logs each one to interlock.unmediated inside the writer's own transaction, with the session user, application_name and transaction id; a TRUNCATE is logged too. On SQLite, interlock install adds permanent journal triggers that record every row change in _interlock_journal, and each committed stage records the range of journal rows it produced, which is exact because BEGIN IMMEDIATE admits one writer at a time. A cascade from a table nobody observes is found row by row;
• an unrecorded stage: one the database committed that no chain records, such as one opened by hand with the stage role's credentials, or by an engine on the default in-memory chain;
• a contradicted stage: recorded as aborted, or under another plan;
• an unresolved stage: committed with only its commit intent in the chain, after a crash in the commit window. Recovery resolves it; EscrowRuntime runs recovery at startup;
• on SQLite, an observed table whose journal trigger is missing.

Each run prints last entry: N; pass it back as --after N to check only what is new. Exit code 5 means a chain failed verification and proves nothing. On PostgreSQL, run it as a role named in audit_roles at install, which may read Interlock's logs and write nothing.

What it cannot see: anyone who can disable the triggers or edit the logs. On PostgreSQL that is the tables' owner or a superuser, which logical decoding would close; on SQLite it is anyone who can write the file.

Scope

Two substrates: SQLite and PostgreSQL. One connector at the row-level-diff standard is worth more than several reporting only row counts, because a row count is the summary this design exists to replace.

What this is not yet a boundary against

Read this before pointing it at a production database. Each item below was reproduced against the shipped code, not inferred.

• Writes outside TableSpec are denied, by SQLite, at prepare time. Effect.target is a label the agent writes and is never compared against the statement, so it cannot be the boundary. SqliteSubstrate installs a sqlite3.Connection.set_authorizer callback that denies INSERT/UPDATE/DELETE on any table not in TableSpec, plus ATTACH/DETACH. A statement whose label claims an observed table while writing elsewhere never executes; it raises ForbiddenStatementError naming the table it reached for. Pass enforce_table_access=False to lift the unobserved-table rule for a migration; the rules below stay on.

  The same callback refuses a statement that writes the capture table (which would erase the measurement), the commit-marker table, or transaction control: a COMMIT inside a stage used to make every effect before it durable before any checker ran. Grant the connection exactly the observed tables anyway; a database-enforced grant is a boundary, and this is a second one in front of it.

  DDL and transaction control are refused twice over: SqliteSubstrate.reject_reason() reads the statement's leading verb at admission, so ALTER/DROP/CREATE/TRUNCATE /VACUUM/ATTACH/PRAGMA/COMMIT/SAVEPOINT and friends are rejected whatever Effect.kind claims.

• A foreign-key cascade into an unobserved table is refused unless you acknowledge it. A DELETE on orders that cascades into an unobserved order_notes used to destroy those notes unmeasured. Every stage now reads the foreign-key graph under its write lock and refuses, before a row changes, a delete or a referenced-key update whose CASCADE, SET NULL or SET DEFAULT actions reach an unobserved table, naming the path. The analysis follows chains of actions to any depth and is column-precise: updating orders.status is never refused because of a key on orders.id. EscrowRuntime runs the same check at startup and logs every gated operation. To let a cascade run unmeasured, name the table in acknowledge_cascades=[...]; every stage that runs with that gap says so in its STAGE_OPENED record. The refusal is enforced three ways, so no single layer is load-bearing: the authorizer refuses the parent operation, the authorizer refuses any foreign-key action SQLite compiles into the table (which is what catches INSERT OR REPLACE), and temporary triggers on the table abort any row change there while the stage is open.

• One engine stages one substrate. A plan whose effects name another substrate is refused at admission; nothing coordinates two substrates, so a plan spanning SQL and a non-transactional sink needs the outbox pattern below.

• blast_radius counts captured mutations, not distinct rows. A row mutated twice in one plan contributes 2, so the limit is conservative and is not a row tally.

• ColumnValueGuard and TenantDrawdownGuard are one-directional. Both fire on a fall in the column total. An inflation passes, and so does any diff whose pre-image total is zero, which includes every insert-only plan. TenantDrawdownGuard re-denominates per tenant; it does not make the guard symmetric.

• StatedFootprint is a cooperative diagnostic. It reads Effect.stated_rows. An agent that omits the field on any effect turns the checker into a no-op.

• Chain durability is opt-in. With no chain_path the chain lives in memory and dies with the process. With one, the intent is write-ahead and a crashed commit is resolved exactly on restart; see Lifecycle. The _interlock_commits table gains one row per committed stage and is not pruned.

• The chain is keyless. Anyone who can write the chain file can recompute a SHA-256 chain from start to finish and it will verify. A reverse anchor in a governed AgentGov is one copy of the head outside the file. With receipts on, each adjudicated plan also gets a signed receipt naming its terminal record, under a key the chain file does not hold; the chain itself stays unsigned.

• Lock footprint. A stage holds write locks for its whole life, a repair search's included. max_stage_seconds bounds it. Human review must not happen inside an open stage; abort, present the recorded diff, and re-stage on approval, because the substrate may have moved.

• Recovery trusts the harness to route its calls. The runtime holds and records its own steps, refuses a revoked tool when asked, and catches an edited transcript at the next step. It cannot see a harness that bills a recovery call to another scope, runs a tool without calling check_tool(), or sends the model something other than step.messages.

• An extension estimate is arithmetic on what the harness declares. Cost per declared milestone times those remaining assumes the rest of the work costs what the done work did; with none declared it prices only the call that did not fit. The receipts in a quote's proof of work are verifiable; its milestones are the harness's claim.

• Latency. Staging roughly doubles write-path round trips. Plans below a blast-radius threshold should bypass escrow entirely.

• Non-transactional sinks cannot be staged. An email has no shadow. Those belong in a transactional outbox written inside the stage and delivered by a separate relay, with the compensation serialized before the outbox row commits. EffectKind.ENQUEUE is defined and carries no behaviour.

• Every threshold in default_checkers is a placeholder. They are uncalibrated. Measure your own diffs and set them from the measurement.

docs/ESCROW_SPEC.md is the interface contract, with a conformance section listing implemented, partial and unimplemented requirements.

Install

agentgov resolves from git, so a clone needs no sibling checkout:

git clone https://github.com/crimsondevil0929/interlock
cd interlock && uv sync

Or install the package directly:

uv pip install git+https://github.com/crimsondevil0929/interlock

The agentgov dependency is a PEP 508 direct reference, which is what makes the second form work. Two consequences worth knowing before you depend on this:

• Interlock cannot be published to PyPI as-is. PyPI rejects direct-URL dependencies. Publishing means putting agentgov on PyPI and pinning a version range instead.
• The pin is an agentgov release tag, never a branch. A resolver cache is keyed on name and version, and @main is a moving target: interlock 0.1.1 locked an agentgov commit that reported itself as 0.1.0 and lacked APIs this README relied on. Interlock 0.1.2 pinned the agentgov tag v0.1.2; 0.2.0 pins v0.2.0, the first release with agentgov.receipts. That tag's package metadata still reports version 0.1.2, so tell the two apart by the commit, 6d3cac2, which uv.lock records. 0.2.1 pins agentgov's v0.2.1 (35788dd), whose package metadata correctly reports 0.2.1. Pin a tag or a commit for anything reproducible, and use uv sync --refresh-package agentgov when you suspect a stale build.

Developing on both repos at once. Because the dependency is git-pinned, uv sync (or uv add --editable) in a plain interlock checkout always fetches agentgov from GitHub -- a sibling ../agentgov checkout on disk is not consulted on its own, so edits there have no effect until you say so explicitly. To point uv at it instead, add to interlock's pyproject.toml:

[tool.uv.sources]
agentgov = { path = "../agentgov", editable = true }

then uv lock && uv sync. This is workspace configuration, not written into built distribution metadata (the direct-reference dependencies entry above is what ships), so it is safe to keep on a local branch while iterating and drop before merging. Remember to revert it (and rerun uv lock) once you are done, or a later uv sync on that branch keeps resolving against your local checkout instead of the pinned tag.

Development

uv sync
uv run pytest -q
uv run ruff check . && uv run ruff format --check .
uv run mypy src/ && uv run mypy --strict tests/ demo.py

The PostgreSQL tests run against a live server: set INTERLOCK_TEST_POSTGRES_DSN to a role that may create databases and roles, and INTERLOCK_REQUIRE_POSTGRES=1 to fail rather than skip without one. Two suites are release gates. tests/test_crash_consistency.py kills a real process with SIGKILL at every step of a PostgreSQL commit, and at random instants, then checks exactly what recovery makes of it. tests/test_tamper_evidence.py alters every part of the audit trail and checks that verification fails exactly there. INTERLOCK_CRASH_SEED and INTERLOCK_CRASH_ROUNDS replay or lengthen the random-instant run.

Apache-2.0.