14. Ask · Stamped architecture

Ask is a view over L4, not a second product. Same plant picture, same evidence rules, same hard stops. Chat does not get a back door past the kernel.

L6 paints the thread; L4 owns the session and tools.

1. Decision summary#

#	Topic	Decision
1	Product shape	One decision loop. Ask reads PSM, memory, case library, and open-card state — it does not emit cards
2	Session owner	L4 holds the Ask session; L6 renders
3	Dialogue banks	Per-thread Hindsight dialogue bank; hard-walled from the plant bank and from other threads
4	Promotion	Thread content enters the plant bank only on explicit confirm or verified closure
5	Tools	No write tools — no equipment, schedule, master data, autonomy, or card emit
6	Evidence	Same ledger citation rules; rupees only from L3 calculator references; uncited claims dropped
7	Discovery	Ask may request a discovery run; it may not bypass grounding or gates
8	Seams	`ask_intent_routing` and `ask_answer_framing` are LLM-only until Jev beats them

2. Why#

Operators will ask “what’s going on on feeder 3?” before they accept a card. If Ask invents a second brain with looser rules, they will trust the chat and ignore the card — or the reverse. One brain, two surfaces.

Dialogue must not silently become operating truth. Shift jokes, hypotheticals, and incomplete stories stay in the thread until a human confirms or a verified close produces a short learning fact.

HITL: Ask informs; the named owner still accepts cards and confirms promotions.

3. Rules#

Rule	Detail
Same evidence rules	Claims cite ledger rows; no invented rupees; tiers assigned by code
No write tools	Same write ban as the decision runtime
Dialogue banks	Per-thread Hindsight dialogue bank; hard wall from the plant bank
Promotion	Dialogue → plant bank only on explicit confirmation or verified closure
No second terminal	Ask does not emit / supersede / withhold cards; the decision runtime does
Grammar always on	Units, comparison bans, hard-stop language from the always-on contract
Case library wins	Dialogue memory is never outcome authority

4. What Ask may do#

Allowed	How
Explain plant state	Typed reads over PSM digest / zoom; frozen into an Ask ledger for the turn
Explain an open or closed card	Case library + trace slots (observed → outcome)
Compare alternatives	Only with L3 calculator / simulator citations; Modeled labeled
Request a discovery sweep	Enqueues the normal discovery path; results still need grounding
Cite constraints	Code-evaluated results; Ask explains, does not re-judge
Draft a clarification for the owner	Text only — no side effects
Refuse	Hard stops, unknown hard constraints, missing proof — say so plainly

5. What Ask must not do#

Forbidden	Reason
Emit / supersede / withhold as a chat side effect	Cards go through Finding / discovery → portfolio → kernel
Change or enable autonomy classes	L5 + named owner
Write equipment, setpoints, schedules, quality holds, master data	Kernel hard stop
Assign the on-shift person	L5 resolves role → person
Promote dialogue → plant bank without confirm / verified closure	Memory hard wall
Invent ₹ or evidence tiers	Calculator and code own those
Cross-query other threads’ dialogue banks	Hard wall
Bypass portfolio or attention hold	Portfolio stays authoritative
Accept a constraint override because a chat thread “agreed”	Constraints are code
Become the home screen instead of the card queue	L6 product rule
Call Opus / Sol on the plant Ask path	Offline council only

6. Memory walls#

Plant bank     ← world facts + short learning facts (eligible closes)
Dialogue bank  ← this Ask thread only (stable document id)
Case library   ← traces + outcomes (authority when banks disagree)

Rule	Detail
Hard wall	Hindsight has no cross-bank query. L4 may merge reads for an answer; it does not merge stores
Retention	Dialogue retention stated per deployment; expiry does not copy into the plant bank
Promotion	Explicit user confirm (“save this as plant fact”) or verified closure learning fact — both audited
Rejection	Declined promotions remembered; no re-nag without new evidence

Mental-model questions stay owner-approved. Ask may draft a suggested update; it does not pin it (13-improvement.md).

7. Evidence rules (same as card path)#

Build a turn ledger from allowlisted reads (PSM, L2 builder reads, case library, open cards, L3 tools as needed).
Partition measured / advisory / model — never merge into one undifferentiated blob.
Every factual claim cites a ledger row id.
Money only via L3 calculator reference.
Free text from the user is delimited data; instruction-like user text is flagged and not treated as policy.
Trace the turn (Ask subset of DecisionTrace): intent seam, reads, framing seam, citations, refusals.

Disagreement on framing is a routing-class seam → registry default. Ask never uses disagreement to invent an action for the floor.

8. Intent routing and answer framing#

Seam	Job
`ask_intent_routing`	Map utterance → allowlisted read plan / refuse / suggest card queue
`ask_answer_framing`	Order evidence, label tiers, refuse out-of-scope asks

Both are closed-option seams (11-models-and-seams.md). Both stay LLM-only until replacement beats the seam decision records.

Ask may use the plant runtime pair under the same dual-family / one-family rules, or a designated Ask slot in the model-pin registry. Calls go through the KR LLM gateway and tenant budget (D21, RECOMMENDED). Offline council models stay offline.

Latency: prefer digest-level answers first; zoom only on request or when proof obligations require it.

9. Relationship to cards and discovery#

Situation	Behavior
User asks about a live card	Read-only explanation from L5 state + trace
User wants “make this a card”	Point to promotion / backlog / normal emit path — Ask does not emit
Soft-gate backlog item	Owner may promote via backlog UI; Ask can explain the gate in plain words
Discovery request	Runs scanners + methods under normal caps; hypothesis lane still needs plant opt-in and dual-family agreement

10. Rejected alternatives#

Alternative	Why rejected
Ask as a separate agent product	Two brains; split trust; doubled governance
Soft write tools “just for notes”	Notes become de-facto setpoints without audit
Auto-promote chat summaries nightly	Contaminates plant bank with unverified talk
Weaker evidence rules in chat	Operators will prefer the looser path
Council models answering Ask live	Wrong path; cost and improvement contamination

11. Evidence that would change this#

Owners need a supervised “create draft card from this thread” → add an explicit draft proposal that still runs full kernel checks (contract delta; still not a silent emit).
Dialogue promotion too rare → improve confirm UX; do not auto-promote.
Intent routing error rate high under pass^k → tighten closed intents or add Jev when records exist.

12. v1 slice vs later#

v1	Later
Read-only Ask over PSM + cards + case library	Richer zoom tools; same walls
Manual confirm to promote facts	Same HITL; better “what would be saved” preview
Discovery request = enqueue sweep	Optional draft-card-from-thread under full gates
LLM intent / framing seams	Jev on high-volume intents when earned