A daily briefing should distinguish new evidence, an unchanged finding, and a correction. Its earlier reports and evidence must remain inspectable after the next run. Tempest produces and validates the research; the host reviews it; Graft retains exact artifacts and decisions.
Ownership
| Component | Responsibility |
|---|---|
| Host | Schedule, topic, credentials, review, access, and retention policy |
| Tempest | Evidence, verification, report, scientific provenance, saved research |
| Graft | Exact artifact selections and explicit host decision history |
| scans | Inspection through Tempest’s validated public trajectory projection |
Start with a retained basis
These are host integration snippets, not a scheduled service. Supply real current eligibility checks and model/search configuration in the application.
library(tempest)
library(graft)
source(system.file("examples", "briefing-basis.R", package = "tempest"))
store <- graft_store("briefing/artifacts", create = TRUE)
basis_path <- "briefing/basis.rds"
knowledge <- NULL
if (file.exists(basis_path)) {
basis <- readRDS(basis_path) # Trusted host-owned checkpoint.
changes <- briefing_changes(store, basis)
historical <- read_briefing_basis(store, basis)
# Inspect changed decisions before choosing the basis for another run.
knowledge <- reuse_briefing_basis(
store, basis, eligible = function(event) host_allows(event)
)
}The checkpoint names one exact decision, selection, stream, and purpose. Reading history verifies the retained evidence and report without granting current admission. Reuse also checks that the decision is the current accepted head, that the purpose matches, and that the host permits consultation now. A stale, withdrawn, or disallowed basis fails; it is never silently replaced by “latest.”
Run and review
result <- tempest_run(
"Battery recycling: evidence that changes today's operating decisions",
knowledge = knowledge,
config = tempest_config(),
output_dir = "briefing/runs"
)
report <- tempest_report(result)
review <- tempest_trajectory_review(result)
selection <- tempest_publish_artifact_research(result, store)Publishing stores a candidate and its exact report, sources, spans, claims, supports, and program provenance. It does not accept the research. The host reviews the complete evidence before making an explicit decision.
previous <- if (file.exists(basis_path)) basis$decision else NULL
accepted <- graft_accept(
store, graft_read_selection(store, selection), "battery-briefing",
expected = previous,
key = "review-2026-09-18", actor = "reviewer",
reason = "Evidence and proposed changes reviewed", purpose = "briefing"
)
basis <- capture_briefing_basis(
store, "battery-briefing", accepted@id, "briefing"
)
saveRDS(basis, basis_path)The predecessor is explicit. If another review has advanced the stream, this request fails and the host reviews the changed state. Retrying the same committed request returns its original event; it cannot restore eligibility after withdrawal.
Unchanged days and corrections
Tempest compares verified statements with active Claim records in the admitted artifact selection. A verified restatement supports a no-change finding; it cannot masquerade as a new observation. The report still needs exact evidence.
An unchanged day can retain its new run and report without replacing the accepted basis. An explicit new review of the same selection creates a distinct decision. A correction accepts a new complete selection, making the old event ineligible for new consultation while keeping the old report and bytes readable. Nothing automatically unions evidence across decisions or accepts model output.
withdrawn <- graft_withdraw(
store, basis$stream, expected = basis$decision,
key = "withdraw-2026-09-19", actor = "reviewer",
reason = "Evidence needs reconsideration"
)
historical <- read_briefing_basis(store, basis)
# reuse_briefing_basis() now fails for this event.Save, inspect, and resume
Saved STORM and Co-STORM products contain their exact artifact selection and materialized evidence, not a native graph object or snapshot sidecar. Offline inspection requires neither a Graft store nor permission to start a new run. Resuming a product with accepted inputs requires a fresh matching knowledge value and current admission. The host controls interruptions and checks during active work; a prior admission is not a lasting permit.
The offline artifact research example exercises acceptance, an unchanged review, correction, and withdrawal. The daily-briefing tests also reopen checkpoints in a fresh process. These synthetic fixtures are contract evidence, not a claim that a live model researched the example topic.