Reference tables
Every table in the guide that restates something the products hold is generated from it at each build, so
none can describe a mode, a key or a rule the products do not have. Most stand in the chapter that explains
them; this page holds the ones a reader looks up rather than reads, and says where the others are.
The modes
A job is one mode with its operands (chapter 3). Every mode, what it does,
the grant it needs on your key and the profile it runs under unless you name one:
| Mode |
What it does |
Grant it needs |
Default profile |
guardrail |
Ask whether one action can fire without one safety context, and report a repair when it can |
guardrail |
fast |
batch |
Run a committed table of properties, one record per row, with a certificate per row and a patch gate on the stalling ones |
batch |
fast |
enumerate_properties |
Emit the candidate action and context pairs the circuit admits, as a proposal for a committed table |
guardrail |
takes none |
derive_pins |
For each row of a committed table, compute the weakest condition under which the obligation holds |
batch |
fast |
derive_expectations |
Answer a sweep's questions and write one expectation per row, each carrying where its answer came from |
batch |
fast |
graph |
Verify an exported agent graph before it is deployed: every tool reached without its check, every write without its confirmation, every order rule, every privileged tool under taint and every saga rule decided, each one broken returned with a path through the graph as its witness |
graph_check |
fast |
replay_certificate |
Confirm a certificate independently: parse its leaf equations, execute its operations, and check the withheld final step closes |
guardrail |
takes none |
forest |
Verify a random forest against a table of robustness, monotonicity, fairness and reachability properties: a counterexample checked on the model for each one violated, a refutation for each one that holds |
forest |
forest |
forest_enumeration |
Search a random forest for counterexamples around real records: every set of at most k named moves, around every record searched, decided exhaustively, with each minimal finding verified on the model and returned as a proposal |
forest |
forest |
hunt |
Hunt a team's rules for every way through them to a forbidden action, and report each as a finding with a severity, a reproduction and a checked fix |
hunt |
fast |
Each mode's fields: what goes in a job's inputs and operands, and what each means.
guardrail
| Field |
Type |
Required |
Meaning |
inputs.policy |
a file |
one of the alternatives |
|
inputs.rulebook_ir |
a file |
one of the alternatives |
|
operands.action |
string |
yes |
the tool action the obligation is about |
operands.context |
string |
yes |
the safety context the action must not fire without; may carry a rail as name@phase, where phase is input, output or dialog |
operands.pins |
pin_list |
no |
further conditioning pins, each a name and a 0 or 1 |
operands.enumerate |
bool |
no |
report every context that defeats the obligation and a proof there is no other |
operands.whole_policy |
bool |
no |
compile the whole policy rather than the live cone under the pins |
operands.prohibitions |
bool |
no |
read the policy's prohibitions as well as its obligations: propose forbidden candidates, honour a row's direction, and flip a single question's target polarity |
operands.session_steps |
int |
no |
chain length to verify at; absent verifies a single turn |
operands.frat_proof |
bool |
no |
log a FRAT proof of a Phase 0 refutation, checked by frat-rs and returned as proof artefacts the certificate's proof_log names |
profile |
string |
no |
the profile; fast where not given |
config |
object |
no |
keys that override the profile's, one by one |
batch
| Field |
Type |
Required |
Meaning |
inputs.properties |
a file |
yes |
|
inputs.policy |
a file |
one of the alternatives |
|
inputs.rulebook_ir |
a file |
one of the alternatives |
|
inputs.rails_config |
a file |
no |
|
inputs.concepts |
a file |
no |
|
inputs.expectations |
a file |
no |
|
inputs.declared |
a file |
no |
|
operands.whole_policy |
bool |
no |
compile the whole policy rather than the live cone under each row's pins |
operands.prohibitions |
bool |
no |
read the policy's prohibitions as well as its obligations: propose forbidden candidates, honour a row's direction, and flip a single question's target polarity |
operands.intent_axis |
bool |
no |
propose and decide each action against the intents in its cone, as user:<intent> rows; a required intent row rests on the classifier and is refused where the intent is uncalibrated |
operands.exactness |
bool |
no |
say on every row whether every route to the action was read exactly, and name the routes that were not |
operands.sequence |
bool |
no |
ask every row with no depth of its own at session_steps, in the cumulative form: the obligation held false at every step of the chain |
operands.session_steps |
int |
no |
chain length to read the policy at; absent reads a single turn |
operands.transfer |
bool |
no |
propose, for review, the controls necessary for a sibling action and not for this one, in a leads block that is never a verdict |
operands.allow_oracle_refutations |
bool |
no |
accept a refutation the engine did not derive, and record on every receipt that it was accepted |
operands.frat_proof |
bool |
no |
log a FRAT proof of a Phase 0 refutation, checked by frat-rs and returned as proof artefacts the certificate's proof_log names |
profile |
string |
no |
the profile; fast where not given |
config |
object |
no |
keys that override the profile's, one by one |
enumerate_properties
| Field |
Type |
Required |
Meaning |
inputs.policy |
a file |
one of the alternatives |
|
inputs.rulebook_ir |
a file |
one of the alternatives |
|
operands.prohibitions |
bool |
no |
read the policy's prohibitions as well as its obligations: propose forbidden candidates, honour a row's direction, and flip a single question's target polarity |
operands.intent_axis |
bool |
no |
propose and decide each action against the intents in its cone, as user:<intent> rows; a required intent row rests on the classifier and is refused where the intent is uncalibrated |
operands.exactness |
bool |
no |
say on every row whether every route to the action was read exactly, and name the routes that were not |
operands.sequence |
bool |
no |
ask every row with no depth of its own at session_steps, in the cumulative form: the obligation held false at every step of the chain |
operands.session_steps |
int |
no |
chain length to read the policy at; absent reads a single turn |
derive_pins
| Field |
Type |
Required |
Meaning |
inputs.properties |
a file |
yes |
|
inputs.policy |
a file |
one of the alternatives |
|
inputs.rulebook_ir |
a file |
one of the alternatives |
|
profile |
string |
no |
the profile; fast where not given |
config |
object |
no |
keys that override the profile's, one by one |
derive_expectations
| Field |
Type |
Required |
Meaning |
inputs.sweep_rows |
a file |
yes |
|
inputs.rulebook_ir |
a file |
yes |
|
inputs.corpus |
a file |
no |
|
inputs.observed |
a file |
no |
|
inputs.previous |
a file |
no |
|
inputs.model |
a file |
no |
|
operands.rung |
string |
no |
the vocabulary rung the router beside this derivation was generated at, L0 to L4, recorded in the table so the pair can be checked |
operands.axes |
string_list |
no |
which of does, to and needs separate two intents. Three is the relation this design uses; naming one is how a table derived under the older relation is produced for comparison |
profile |
string |
no |
the profile; fast where not given |
config |
object |
no |
keys that override the profile's, one by one |
graph
| Field |
Type |
Required |
Meaning |
inputs.export |
a file |
yes |
|
inputs.catalogue |
a file |
yes |
|
inputs.rulebook |
a file |
no |
|
operands.session_steps |
int |
no |
the turns every obligation is decided again across; the binary's default is 1 |
profile |
string |
no |
the profile; fast where not given |
config |
object |
no |
keys that override the profile's, one by one |
replay_certificate
| Field |
Type |
Required |
Meaning |
inputs.certificate |
a file |
yes |
|
forest
| Field |
Type |
Required |
Meaning |
inputs.model |
a file |
yes |
|
inputs.properties |
a file |
yes |
|
inputs.anchors |
a file |
no |
|
profile |
string |
no |
the profile; forest where not given |
config |
object |
no |
keys that override the profile's, one by one |
forest_enumeration
| Field |
Type |
Required |
Meaning |
inputs.model |
a file |
yes |
|
inputs.search |
a file |
yes |
|
inputs.anchors |
a file |
yes |
|
profile |
string |
no |
the profile; forest where not given |
config |
object |
no |
keys that override the profile's, one by one |
hunt
| Field |
Type |
Required |
Meaning |
inputs.policy |
a file |
one of the alternatives |
|
inputs.rulebook_ir |
a file |
one of the alternatives |
|
inputs.model |
a file |
no |
|
inputs.rails_config |
a file |
no |
|
inputs.dependencies |
a file |
no |
|
inputs.properties |
a file |
no |
|
inputs.intents |
a file |
no |
|
inputs.vocabulary |
a file |
no |
|
inputs.domains |
a file |
no |
|
inputs.clarification |
a file |
no |
|
inputs.expectations |
a file |
no |
|
inputs.route_baseline |
a file |
no |
|
inputs.probes |
a file |
no |
|
inputs.corpus |
a file |
no |
|
inputs.ratchet_baseline |
a file |
no |
|
inputs.dialogues |
a file |
no |
|
inputs.dialogue_baseline |
a file |
no |
|
inputs.graph |
a file |
no |
|
inputs.catalogue |
a file |
no |
|
inputs.baseline |
a file |
no |
|
inputs.quality_profile |
a file |
no |
|
operands.session_steps |
int |
no |
the conversation depth every obligation that holds at one turn is decided again at; the binary's default is 3 |
operands.gate |
bool |
no |
let the quality gate decide the run's outcome: a failed gate exits 7, which the job reports as the gate's result |
operands.count_ways |
bool |
no |
count every breaking context of each open policy finding: its ways through. Needs the hunt_details entitlement |
operands.no_solve |
bool |
no |
decide no obligation and say so, running every stage that needs no solver: the router's findings are counted and cost no solve-seconds. Set by the service on a hunt submitted after the key's month is spent |
profile |
string |
no |
the profile; fast where not given |
config |
object |
no |
keys that override the profile's, one by one |
Where the other tables are
- The HTTP API's 74
operations, by product: chapter 18.
- The package's public names: chapter 17.
- The profiles, and every configuration key: chapter 10.
- The verdicts, the phases and the diagnostics: chapter 6.
- The events a job streams: chapter 9.
- The error codes and the refusals: chapter 15.
- The plans and what each includes: chapter 40.
- The bug hunter's 49
rules, each with its page: the rule pages.
- The words of
agent.yaml, a tool's words and the stalwart command: chapter 30.