AI research-agent deep dive / PG-015
One Bounded AgentJob In Practice
An AgentJob is a permission envelope for one auditable transaction. It names the objective, inputs, allowed paths, forbidden classes, validators, expected outputs, completion record, and handoff boundary.
Static comprehension diagram
Read the job as an envelope, not an answer machine.
The diagram is a static reader aid generated from the dossier Mermaid source. It does not replace the committed AgentJob, completion, handoff, or registry records.
Concrete committed example
AJ-RT-20260614-249-001 shows one bounded job.
This example is source-approved and committed. It is shown as an operations pattern, not as proof of the physics topic it worked near.
AgentJob surfaces
The job states what is allowed, required, and forbidden.
A bounded job is useful because it is inspectable at transaction scale. The same record names what may be read, what may be written, what must be produced, and what must remain outside scope.
Objective
One transaction
Classify the post-Gate route and select exactly one next bounded packet while preserving adoption and downstream-claim blocks.
An objective is not broad permission to settle the topic.
Allowed reads
Evidence surface
The job names prior tasks, handoffs, artifacts, roles, schemas, ledgers, and registries that may be inspected.
Read access does not adopt every possible claim in a source.
Allowed writes
Change surface
The write list is limited to the task directory, selected control state, handoff records, and specific registries.
Anything outside the allowlist is out of scope.
Forbidden classes
Stop conditions
The job forbids ontology writes, source-law adoption, target imports, role authority, validation authority, and downstream promotion.
A forbidden class needs a new route, stop, or human gate.
Validators
Receipt evidence
Memory, project-control, research-control, generated-derivative, diff, and hygiene checks are named before completion.
PASS is operational evidence, not theorem proof.
Expected outputs
Closeout obligations
The job lists child outputs, a fused artifact, completion YAML, and handoff files that the completion must account for.
Output existence is not claim adoption.
Closeout chain
Completion and handoff close this job rather than widening it.
The completion records what happened. The handoff recommends what may happen next. Neither one turns the past AgentJob into live authority for another task.
Completion
Command results and verdict
The completion records output paths, command evidence, validation status, verdict, uncertainty, and next recommendation.
Handoff
Separate next packet
The handoff recommends one future criteria-formalization packet and repeats the claims that remain blocked.
Registry
Discovery and checkability
The AgentJob registry links the job to the task, decision, completion path, status, timestamps, and concise summary.
Boundaries
One job is intentionally small.
Small scope is the control mechanism. It prevents a workflow receipt from becoming broad project truth, reusable permission, or physics promotion.
Permission
not reusable authority
A historical AgentJob does not grant future write permission.
Truth
not broad project truth
One transaction can record what it did; it cannot settle the whole project.
Validation
not proof by PASS
Validators can accept a checked state without proving physics.
Output
not adoption by existence
An expected output can exist as evidence without becoming an adopted claim.
Privacy
no private local details
The page uses repo-relative committed paths and avoids local runtime context.
Handoff
not current-job extension
The next packet requires its own decision and job.
Safe and unsafe summaries
Safe summary also names the unsafe one.
The RT-249 AgentJob preserved no coupling-law adoption, no matter-coupling derivation or adoption, no MetricData(E), no g_eff, no Einstein equations, no benchmark promotion, and no completed derivation.
Safe summary
One bounded AgentJob makes one transaction auditable.
The job names its objective, inputs, allowed paths, forbidden classes, validators, expected outputs, completion record, and handoff boundary.
Unsafe summary
One AgentJob settles broad project truth.
It would be unsafe to treat one job as reusable permission, theorem proof, benchmark promotion, or adoption of expected outputs.
Reader path
Use related internal routes before source inspection.
Internal routes give the reading path. Source links remain provenance for readers who want to inspect committed records directly.
Research-agent workflowSee the full request-to-decision-to-job-to-handoff record chain.Roles and skillsRead why a role template is not current task authority by itself.Memory and registriesRead how retrieval and registry rows support source inspection.Director and AgentJob lifecycleInspect how AgentJobs fit with decisions, completions, and handoffs.Validator and operator workflowRead why PASS is bounded operational evidence.Source authorityReview why source files remain authoritative over reader surfaces.