Interrogation Page Archetype — Stage Zero Spec

The Problem (Evidence Recap)

Audited Claude + Codex history confirmed the felt regression. In Codex research-skill sessions, mean agent questions fell from a peak of ~28/session (early May) to ~0.3/session (June). The drop is precisely datable to the 2026-05-21 "require alignment pages for all skills" commit.

The root cause is not the alignment page absorbing questions. It is two compounding structural facts:

• Tier classification. ~25+ research/analysis skills are context_intake: scoped, which caps user interrogation at 1–3 questions by design (docs/interview-convention.md:41), and per-skill instructions often make even those optional ("only if unclear", competitive-analysis/SKILL.md:430).
• Cold-start routing. 3 of 4 flagship research orchestrators (competitive-analysis, positioning, journey-map) route a cold start "directly to state E" — the framework-selection page — with no deep interview at all (docs/research-session-loop-convention.md:234–239).

The produced pages then make it feel like rubber-stamping: across all 77 alignment pages, the only controls are approve / thumbs-down / clarification triads plus multiple-choice routing gates whose options the agent pre-authored. There is not a single open free-text field that shapes the research before it is produced.

Decisions Already Made

These were settled in conversation before this page. They frame the spec below; the gates let you reopen any of them.

Proposed: Interrogation Page Archetype

A new HTML page archetype, distinct from the alignment page, governed by its own convention doc (docs/interrogation-page-convention.md, to be authored) and its own audit. Its purpose is elicitation before research, the inverse of the alignment page's approval after research.

Defining contrasts vs. the alignment page

What it must NOT become

The audit for this archetype must enforce that it does not collapse back into approve-only UX. Minimum: each round page must contain at least one genuinely open input (free-text or recommend-and-override) that shapes downstream research, not only accept/reject on pre-generated content. This is the structural guard against re-introducing the rubber-stamp problem in a new wrapper.

Stage Model: 0 / 1 / 2

Interrogation is inserted as a new state before the existing first interaction. It does not replace stage one; for scoped skills it will feel like the new first screen because framework selection used to be first.

stage 0  →  interrogation/{skill}-r1..rN.html   [ELICITATION: open input drives research]
              loop until confidence gate passes
stage 1  →  alignment framework/scope page       [SELECTION: options informed by stage 0]
              research runs
stage 2  →  alignment/{skill}-{topic}.html        [REVIEW/APPROVAL of deliverables]

This preserves the convention's firewall semantically — intake → selection → review — and only changes where intake happens (HTML, not terminal). It generalizes across tiers: deep skills get a rich stage-zero loop; scoped skills get a lighter one. No tier needs terminal interrogation anymore.

The Round Loop & Confidence Gate

Each round is one heavy phase of the Research Session Loop: the agent builds an interrogation page, stops, the user answers, the agent reads the answers and runs its confidence gate.

round = 1
build interrogation/{skill}-r{round}-{branch}.html   (manifest if round==1, else follow-ups)
STOP → user answers → compiled answers YAML
agent reads answers, runs CONFIDENCE GATE:
    confident enough to produce useful research?
        NO  → round += 1; build next round of deeper/follow-up questions; repeat
        YES → write coverage checkpoint; advance to stage 1

Confidence gate (loop exit)

The gate is a contract, not a fixed round count. It restores the "relentless until confident" feeling. Proposed enforcement: the agent may not advance to stage one until (a) at least one interrogation round is completed, and (b) every interview area defined by the skill is covered or explicitly waived by the user. The coverage checkpoint (today's deep-tier Phase 4, interview-convention.md:31) becomes the loop's exit gate — the agent's "here is everything we established, confirm completeness or flag gaps" page.

Volume expectation

Fewer artifacts than it sounds. A 30–40-question terminal interview was ~12 chunked turns; rounds of 6–8 questions each land at ~3–6 rounds for a deep skill, 1–2 for a scoped one.

Round 1 = Assumptions Manifest

The assumptions manifest (interview-convention.md:27) is not a separate concern from interrogation — it is the ideal first round of it. It bans lazy "tell me about..." questions by forcing the agent to do its homework first and ask only where it has no basis to assume. Incorporating it beats replacing it: delete it and round 1 regresses to open filler prompts.

Every correction you make to an assumption seeds round 2's follow-ups — that is the mechanism that makes the loop adaptive. So the manifest bookends the loop: assumptions manifest = entry (round 1), coverage checkpoint = exit.

File Naming, Location & Answer Capture

Location and naming

• Directory: interrogation/, a sibling of alignment/. Keeping these out of alignment/ matters concretely — scripts/audit-alignment-pages.mjs audits alignment/*.html against the alignment-page contract, which elicitation pages won't and shouldn't satisfy.
• Filename: interrogation/{skill}-r{N}-{branch}.html. The {branch} stub handles research forks (multiple product paths → parallel question threads); in the flat common case {branch} == the topic slug, matching today's {skill}-{topic} naming.

Answer capture (what drives "repeat over and over")

Each round page needs a captured-answers sidecar — the same "Compile YAML" mechanism the alignment page already uses — so the agent reads round N's answers back, runs the confidence gate, and only then decides whether to emit round N+1 or write the coverage checkpoint. Proposed sidecar path: research/_working/interrogation-{skill}-r{N}.yaml (or the product-path equivalent), preserving the round-by-round audit trail.

Convention Integration Map

Where the change lands if approved (report-only — none of these are edited yet).

Open Tradeoffs

Evidence Matrix

Assumptions, Alternatives & Gaps

Confidence Register

Alternatives Considered

Gaps

Glossary Additions

Proposed terms introduced by this spec. Only approved terms are appended to research/glossary.md during the confirmed-page write step.

Review Gates