Research: Question to Verified Cited Report

# Title: Research: Question to Verified Cited Report
#
# Public workflow: investigate a question from three independent angles in
# parallel, synthesize one grounded report, then run a SEPARATE groundedness
# judge that checks every claim against its cited source and gates the result —
# the synthesizer revises once on failure. Public roles only.

name        = "research"
description = "Investigates a question from background, evidence, and counter-argument angles in parallel, synthesizes a cited report, then loops a groundedness judge until every claim verifies against its sources."

# ── 1. Parallel angle research ───────────────────────────────────────────────
[[steps]]
name     = "angles"
parallel = true

  [[steps.run]]
  name    = "background"
  role    = "assistant:researcher"
  session = "fresh"
  prompt  = """
Research the background and current landscape of this question. Search broad,
then narrow. Prefer primary and official sources over SEO/content-farm pages.

For every claim return: the claim, an exact supporting quote, the source title
and URL, and your confidence. Output a structured evidence list — not prose.
Do not assert anything you cannot quote from a source you actually searched.

<question>
{{input}}
</question>
"""

  [[steps.run]]
  name    = "evidence"
  role    = "assistant:researcher"
  session = "fresh"
  prompt  = """
Research the strongest evidence, data, and expert positions that bear on this
question. Search broad, then narrow. Prefer primary/authoritative sources
(filings, datasets, peer-reviewed work) over secondary summaries.

For every claim return: the claim, an exact supporting quote, the source title
and URL, and your confidence. Output a structured evidence list — not prose.
Do not assert anything you cannot quote from a source you actually searched.

<question>
{{input}}
</question>
"""

  [[steps.run]]
  name    = "counter"
  role    = "assistant:researcher"
  session = "fresh"
  prompt  = """
Research counter-arguments, risks, caveats, and credible dissenting views on this
question — actively seek disconfirming evidence, not reassurance. Prefer primary
and authoritative sources over SEO/content-farm pages.

For every claim return: the claim, an exact supporting quote, the source title
and URL, and your confidence. Output a structured evidence list — not prose.
Do not assert anything you cannot quote from a source you actually searched.

<question>
{{input}}
</question>
"""

# ── 2. Synthesize + groundedness gate ────────────────────────────────────────
# Synthesizer (researcher) and judge (ai:evals) are distinct roles so the writer
# never grades its own work — separation of concerns is the core anti-hallucination
# lesson. The loop caps at 2: one synthesis, at most one targeted refinement.
[[steps]]
name           = "report"
loop           = true
max_iterations = 2
exit_when      = { output = "judge", matches = '(?m)^VERDICT: GROUNDED' }

  [[steps.run]]
  name    = "synthesizer"
  role    = "assistant:researcher"
  session = "continue"
  prompt  = """
Synthesize the three research passes below into one balanced, well-structured
report that answers the question.

Rules:
- Every non-trivial claim carries an inline citation that maps to a source from
  the evidence lists below — never introduce a claim with no backing source.
- Surface disagreements and counter-arguments explicitly; do not average them
  into a bland consensus.
- Mark single-source or weak claims as tentative.
- End with a numbered Sources list (title + URL) that the citations point to.

On a later iteration you will receive the judge's per-claim feedback: fix exactly
the flagged claims — re-ground, downgrade to tentative, or drop them. Do not force
an unsupportable claim.

<question>
{{input}}
</question>

<background>
{{background}}
</background>

<evidence>
{{evidence}}
</evidence>

<counter>
{{counter}}
</counter>

Output only the report — no preamble, no commentary.
"""

  [[steps.run]]
  name    = "judge"
  role    = "ai:evals"
  session = "fresh"
  prompt  = """
You are a groundedness judge in a fresh session. You did not write this report.
Check it against the gathered evidence — never from memory.

<report>
{{synthesizer}}
</report>

<background>
{{background}}
</background>

<evidence>
{{evidence}}
</evidence>

<counter>
{{counter}}
</counter>

Two layers, in order:
1. Citation resolution — every cited claim must map to a source that actually
   appears in the evidence above, and its supporting quote must overlap that
   source. Any fabricated, orphaned, or quote-mismatched citation is a hard fail.
2. Per-claim groundedness — for each non-trivial claim, give one line of
   reasoning then a binary label: Supported or Not-Supported by its cited source.
   Also check the report covers the requested aspects including counter-arguments
   and labels single-source claims as tentative.

PASS only if there are zero unresolved/fabricated citations AND every claim is
either Supported or explicitly downgraded to tentative/single-source. Otherwise
list the specific failing claims with what to fix — this goes straight to the
synthesizer.

Critique first, then end with exactly one line: VERDICT: GROUNDED or VERDICT: REVISE. Nothing after it.
"""