haipipe-application-advice

name	haipipe-application-advice
description	Stage orchestrator for the intervention's 0-lifecycle/1d-advice/1d-advice.md: rung 1d of the venue-FREE evidence ladder (1a-descriptions -> 1b-themes -> 1c-claims -> 1d-advice) and the ladder's DELIVERABLE. Design advice: actionable, evidence-derived guidance entries (W-shaped), each derived from >=1 claim in the 1c ledger. Advice is counsel, not mandate: downstream venue-ALIGNED stages ADOPT or DECLINE entries per venue+audience (declined entries stay for the next round). Distinct from venue Artifact Principles (channel-how); these are content-what. Markdown only. Trigger: advice, advise, design advice, recommendations, what should the message do, principles (legacy), /haipipe-application advice.
argument-hint	[intervention-path] [--deposit <Ann>]
allowed-tools	Bash, Read, Write, Edit, Grep, Glob, Skill
metadata	{"version":"0.1.9","last_updated":"2026-07-19","summary":"Advice stage (rung 1d of the venue-FREE 1a–1d evidence ladder; the W rung and the ladder's DELIVERABLE) — design advice, one A<n> per entry (exploit|explore role) derived from >=1 claim in the 1c ledger; counsel not mandate (venue-ALIGNED stages adopt or decline downstream); content-WHAT, distinct from the venue's channel-HOW Artifact Principles. History: ./CHANGELOG.md."}

Skill: haipipe-application-advice

Rung 1d of the venue-FREE evidence ladder, and the ladder's deliverable rung. It answers: what does the evidence advise this intervention to do, and which claims ground each entry?

1a-descriptions   what the data looks like
1b-themes         what patterns/topics emerge
1c-claims         what generalizes (the ledger)
1d-advice         what the evidence advises (the deliverable)   <- THIS RUNG

Read first: ../../../PHILOSOPHY.md, ../../../haipipe-application/SKILL.md (Stage Gate Protocol).

What's special: advice is the ladder's deliverable

1. It delivers W, and every entry is derived from a claim. Paper stops at K (defended claims); the application climbs one rung further to W — advice the artifact work can execute. Derivation is mandatory: an advice entry without a claim behind it is vibes, the exact fabrication the ladder exists to prevent. Every entry must pass the W-actionability test — "could the artifact stage write the exact message move from this line?" — or it is a claim restated, and pushes back to 1c.

2. Advice is counsel, not mandate. This rung writes ALL the guidance the evidence supports; the venue-ALIGNED stages (pitch/narrative/display/artifact) then SELECT, each recording which A<n> it adopts or declines with a one-line why. Declining is a design choice, not a failure — a declined entry waits for the next venue or round. Adoption records live DOWNSTREAM, keeping this doc venue-FREE; claim-audit traces artifact -> adopted A -> C -> anchor, and no rule says every entry must be used.

3. Content-WHAT, not the venue's channel-HOW. 2-venue.md Artifact Principles are channel-HOW (length, cadence, format for THIS modality; venue-ALIGNED, rewritten on retarget). This doc is content-WHAT (what the content should do to work; venue-FREE, survives retarget). Downstream stages read both; they never merge.

4. Every entry carries an exploit|explore role. Exploit entries rest on settled evidence and take the settlement bars below; negative advice ("avoid X", derived from a refuted claim) is exploit-role — the refutation IS settled evidence. An explore entry is a deliberate test-to-learn bet: it MAY derive from weak/GAP claims provided it (a) carries the explore tag visibly, (b) names the C<n> its deployed arm will settle (Settles: C<n> via iterate), and (c) states its compliance rails — miss the contract and it fails CHECK regardless of venue. The A/B result flows back (iterate -> 1a backfill -> C flips) and the entry graduates to exploit or moves to Rejected; deploy itself becomes an evidence probe. Explore is a strategy, not a loophole: the visible tag, the named settling C, and the rails are what separate a bet from vibes.

The four phases, in advice

DRAFT   re-mine last round's Rejected reservoir (did a refuting C flip?), read 1c-claims.md (statuses +
        campaign) and 1b-themes.md, elicit taste on guidance priorities; CONSUME every supported/weak C —
        an A entry, a Rejected entry, or a No-action line with a why; consider explore bets on promising weak/GAP
PROBE   rarely fires: derivation is in-stage work. An entry exposing a NEW evidence gap routes BACK to 1c-claims
        (raised there as a question SECTION in 1-probes/), never gathers here. Mechanics:
        ../../../2-phase/1-probe/haipipe-application-probe/SKILL.md
REVISE  actionability pass (every A survives the test), scope tightening, caveat wording
CHECK   the LADDER GATE for light/medium venues (batched per the Stage Gate Protocol): every exploit A meets the venue's
        settlement bar, every explore A is tagged + names its settling C + states rails, actionability passed,
        every supported/weak C consumed, no unresolved STALE tags; user confirms -> Gate Ledger row(s) -> advance

Advice RARELY probes: a new evidence gap is not gathered here but routed back to 1c-claims, whose PROBE phase binds it. Each phase runs its worker (haipipe-application-draft / -probe / -revise / -check); announce every phase boundary and never let CHECK go implicit. Rounds + back-routing (loop-until-dry for medium+ venues, [ROUND n] / [ROUTE -> claims] in _LOG) follow ../../../haipipe-application/SKILL.md (Stage Gate Protocol). The ladder doc (1d-advice.md) is the record; adopted/declined A-ids with a why live there.

The artifact

0-lifecycle/1d-advice/1d-advice.md — full skeleton in ref/advice-template.md:

Advice      one A<n>: guidance in one sentence + role (exploit|explore) + derivation (>=1 C id) + scope/boundary
            + status; close with a No-action line (every supported/weak C consumed by no entry, each with a why)
Rejected    entries considered and dropped, one sub-item each with the refuting C id or reason (the reservoir; may be empty)
Q-consumer  the questions this rung raises, one `## Q-Advice-<n>` block each (Ask / Why / Answer; rare — derivation is in-stage work; may be empty);
            APPROVE adds each -> 1-probes/ pointer + state

One entry, one sub-item: **A1 - majority framing where compliance is high - exploit - active** / Use descriptive-norm framing only where cohort compliance > 50%. / Derivation: C1 (supported), C2 (weak - boundary caveat). / Scope: adult cohort; revisit if C2 settles. An entry citing only weak claims carries the caveat inline; whether that passes the gate is venue-scaled (below). Sidecar: _LOG_1d-advice.md (phase journal). Ids A<n> are ladder-local; artifact and claim-audit trace artifact -> adopted A -> C -> anchor. Formatting: ===== title, ----- sections, **bold** sub-items, one sentence per line; content sections use no #, Q-consumer questions use ## Q-Advice-<n>.

Settlement coupling (venue-scaled, read at CHECK)

The venue's claims_settlement bar (STATUS.md) applies through the derivation chain to EXPLOIT entries. Explore entries are exempt from the bar BY CONTRACT, not by leniency: tag visible + settling C named + rails stated, or the entry fails CHECK regardless of venue.

light    an exploit A may cite weak claims with an inline caveat; GAP-derived exploit forbidden
medium   load-bearing exploit A cite supported-or-weak-with-caveat; others may caveat
full     every exploit A cites supported (judged) claims only

Absent claims_settlement (venue unpinned) → apply light provisionally.

Questions this stage typically raises

DRAFT's RAISE+PLAN step raises what the draft cannot answer. These are the kinds this stage is prone to — read this list, then walk the draft against it.

🎬 actionability     Can the recipient actually DO this? Name the barrier if not —
                     advice nobody can act on is not advice.
⚠️ harm              Who is made worse off if this advice is followed and the claim
                     turns out to be wrong? Ask it before deployment, not after.
📐 dosage            How much / how often / for how long? Advice without a dose is
                     a sentiment.
🏛️ standard of care  Does this contradict published guidance? If so, on what evidence?

Exits

promote -> /haipipe-application venue    pin the modality (or -> pitch if venue already pinned)

This CHECK is usually the ladder gate; end every reply with the closing block (stage line via ../../../haipipe-application/stage-strip.sh).