---
title: "TLF Interpretation: DRAFT (biostatistician review required)"
params:
  tlf_id: "t_ae_overview"
format: html
---

```{r setup}
source("R/slm_client.R")
source("R/extract_facts.R")
source("R/validate_interpretation.R")
source("R/interpret.R")
suppressPackageStartupMessages(library(gt))
```

```{r run}
# PRODUCTION: read the VALIDATED result dataset for params$tlf_id (the aggregate
# numeric summary the TLF program emitted: n's, %'s, geo-means; no PHI).
# SKELETON: a synthetic AE-overview result so this renders end-to-end offline.
ae_result <- data.frame(
  category = c("Any TEAE","Related TEAE","Severe TEAE","SAE",
               "TEAE leading to discontinuation","Death"),
  n   = c(24L, 15L, 2L, 1L, 0L, 0L),
  pct = c(80.0, 50.0, 6.7, 3.3, 0.0, 0.0))

facts <- extract_facts(params$tlf_id, ae_result, arm = "Active 100 mg", N = 30L)
res   <- interpret_tlf(params$tlf_id, facts)   # uses local SLM, or STUB if none installed

badge <- c(DRAFT_OK = "✅ DRAFT_OK", FLAGGED = "⚠️ FLAGGED", REJECTED = "⛔ REJECTED")[res$status]
callout <- c(DRAFT_OK = "note", FLAGGED = "warning", REJECTED = "important")[res$status]
```

## The validated output (source of truth)

The table/figure shown here is the **validated TLF result**, not anything the model produced.

```{r table}
ae_result |>
  gt() |>
  tab_header(title = "Overview of Treatment-Emergent Adverse Events",
             subtitle = "Active 100 mg (N = 30): Safety Population") |>
  cols_label(category = "Category", n = "Participants (n)", pct = "%") |>
  tab_source_note("Validated result dataset. Reported numbers from validated programs per the study's SOPs.")
```

## Draft interpretation  `r badge`

::: {.callout-`r callout`}
## Draft: biostatistician review and sign-off required

`r res$draft`
:::

```{r status, results="asis"}
m <- res$audit
cat(sprintf(
"- **Status:** %s\n- **Backend:** %s\n- **Model:** %s%s  ·  seed %s · temp %s\n- **Prompt SHA1:** `%s`  ·  **Facts SHA1:** `%s`\n",
  res$status, m$backend, m$model_name,
  if (isTRUE(m$stub)) "  *(STUB, no local model installed)*" else "",
  m$seed, m$temperature, substr(m$prompt_sha1,1,12), substr(m$facts_sha1,1,12)))
```

## Grounding facts + automated checks

```{r checks}
fct <- data.frame(fact = names(facts),
                  value = vapply(facts, function(x) format(x, trim = TRUE), character(1)),
                  row.names = NULL)
gt(fct) |> tab_header("Facts the draft was grounded on (the only numbers permitted)")
```

```{r checks2, results="asis"}
v <- res$validation
cat(sprintf("- **Numeric-consistency:** %s; %d numbers checked%s\n",
            ifelse(v$numbers$ok, "PASS", "FAIL"), v$numbers$n_checked,
            if (length(v$numbers$unmatched)) paste0("; UNMATCHED (not in facts): ",
                paste(v$numbers$unmatched, collapse = ", ")) else ""))
cat(sprintf("- **Claim guard:** %s%s\n", ifelse(v$claims$ok, "PASS", "FLAGGED"),
            if (length(v$claims$flagged)) paste0(": ", paste(v$claims$flagged, collapse="; ")) else ""))
cat(sprintf("- **Template:** %s (%d words, %d sentences)%s\n",
            ifelse(v$template$ok, "PASS", "ISSUE"), v$template$n_words, v$template$n_sent,
            if (length(v$template$issues)) paste0(": ", paste(v$template$issues, collapse="; ")) else ""))
```

## Human gate (sign-off)

Nothing is `APPROVED` automatically. The biostatistician edits the draft as needed and records
sign-off; only an approved interpretation flows to the CSR narrative. The audit record
(`res$audit`), with the model digest, seed, prompt/facts hashes and timestamps, is written to the sign-off store.

```r
# approve_interpretation(res, approver = "P. Iben", final_text = "<edited text>")
#   -> writes res$audit + final_text + approver + timestamp to the audit log,
#      sets approved = TRUE. (Implement against your validated-doc store.)
```