[design·M1] Owner-live: answers exceed the reading budget — verbosity is the next usefulness bottleneck after #827

[atomchung]

Where this came from

The #827 blind A/B (receipt on #828). The candidate won the owner's blind preference 3/4 with one tie — but the owner's cross-cutting feedback hit both arms, so this is not a #828 regression; it is the next bottleneck.

Owner-live evidence (sanitized quotes)

• Cash-deployment scene: the opening paragraph of one arm 「第一大段話完全沒有意義」; overall 「廢話太多」 said of both arms.
• Comparison scene (judged a tie despite correct engine math): 「內容已經太多了，我沒有辦法看完這些東西了」 — the answer's usefulness was capped by its length, not its content.
• Discovery scene (the candidate's win): preferred because it answered at all, not because its length was right.

Reading

#823/#827 removed refusal shapes and effort ceilings; nothing now expresses length-to-decision-value. Leading paragraphs restate context instead of deciding; tables and enumerations run past the owner's reading budget even when every number is right.

Pattern siblings, cross-referenced rather than duplicated: #552 (disclosure narration drifts into hedge/caveat filler) and #579 (decision-first answer, not a disclosure dump). This issue adds the owner-live, post-#827 evidence that the remaining gap is expression economy, not permission.

Suggested direction (owner to rule)

An expression-contract-level reading-budget principle — lead with a decision core the owner can absorb in ~30 seconds; every additional block must change the decision — not a mechanical character cap (the #543-era ceiling #827 just deleted; do not reintroduce it as a number).

Refs #827, #828, #552, #579, #823.

[atomchung]

Owner ruling + root-cause diagnosis (2026-08-20)

Owner ruling: confirmed as the next problem to solve. Owner's words: 「這產品太囉嗦，一定是我們下一個要解決的問題」「故事性非常差，可閱讀性非常的爛」.

Root cause — three layers, each individually correct

1. The engine hands every answer an obligation list. One consider call emits ~15 must_state facts, plus rule_effects.must_convey, quote_verbatim, and disclosures_display sentences that SKILL.md instructs to paste as written. The model's cheapest discharge of an obligation list: one sentence per item, arranged into tables. The wall of tables the owner cannot finish reading is obligation discharge, not judgment.
2. The expression layer disclaims quantity by design. expression-contract.md §2: “Expression governs where an owed fact is said, never whether it is owed.” V1–V9 govern the lead, tension count, question timing, and manufactured content — no rule anywhere governs total volume or its distribution. So every owed fact lands in body prose.
3. History. These contracts grew in the anti-hallucination / anti-overreach / anti-refusal era ([design] Agent narration of engine disclosures drifts into hedge/caveat filler beyond the documented concise fact #552, [design·M1] Decision-first TradeEvaluation answer — one supported judgment, not a disclosure dump #579, [design·P1] When consider cannot compute, no contract owns the answer — so the agent narrates the gap instead of framing the decision #674). Every rule is a not-wrong rule, and a not-wrong rule is always discharged by writing one more sentence. [implementation·skill] Move limits to claim/write/action boundaries — free reasoning, research and recommendation #827 deleted the refusal shapes and ceilings but kept the obligation-to-paragraph habit — and amplified it: in the [implementation·skill] Move limits to claim/write/action boundaries — free reasoning, research and recommendation #827 A/B, the candidate's first answers were longer than baseline in all four scenes (discovery scene >3×).

Owner's three complaints map one-to-one: “first big paragraph meaningless” = the lead discharges obligations instead of answering the question; “the all-in-one-name simulation is meaningless” = the model manufactured a comparison scene to give obligations a carrier; “too much to finish reading” despite correct math = correctness taxed away by volume (scene 4 judged a tie with every number right).

Fix design — obligations unchanged, discharge changes

1. V10 — reading budget (voice registry + oracle): the first screen (~200 zh chars) must be self-sufficient — stance + the deciding reason + one concrete action/condition. Every later block must pass “would deleting this change the decision?”. This is a budget (core first), not a length cap (the [design] Freeform conversational answers have no effort/latency bound — an ad hoc portfolio question triggered a 34-turn chart build #543-era ceiling [implementation·skill] Move limits to claim/write/action boundaries — free reasoning, research and recommendation #827 just deleted; do not reintroduce as a number). Oracle: deterministic first-N-chars probe (recommendation anchor + the deciding must_state present) + owner-live 30-second-core witness.
2. D7 — disclosure aggregation: sources / as-of / unchecked dimensions collapse into one compact end-block, one line per item, non-narrative; only truth-critical qualifiers stay beside their number (existing D-series already draws that line). disclosures_display retargets from body sentences to end-block lines.
3. Redefine “state” for must_state (trade-consequence.md + SKILL.md): stated = the fact appears with its anchor — inline number, table cell, or end-block line all qualify. 15 obligations ≠ 15 sentences.
4. Narrative axis flip (SKILL answer shape): organize by the reader's question chain (what you asked → answer → why → what would overturn it → what you'd do), not by payload field order. Extend V4's “one lead tension” to one storyline per answer.
5. Bloat-fails-closed evals: add episode/mutation cases where an all-obligations-discharged but core-less answer FAILS — otherwise a green suite reads as a governed output again (the One expression contract, decoupled from every output surface (closes #823) #824 lesson).

Expected surfaces: docs/output-voice.md (+V10 + witness), docs/expression-contract.md (§2 wording: volume distribution becomes expression's business), skills/fomo-kernel/SKILL.md, references/trade-consequence.md, references/freeform-answers.md, tests/agent/check_voice.py + witnesses, episode bank.

Per the maintainer route, implementation starts in a fresh session against latest main, not in this issue's session.

[atomchung]

Owner redirect (2026-08-20): deletion-first, not budget-first

Owner ruling on the fix philosophy, superseding the V10-first design above: 「什麼東西是一定要的，什麼東西是不必須的……沒有辦法說出一定要理由就全部刪掉」. Audit the obligation inventory itself; anything without a must-have justification is deleted. The reading-budget rule (V10) is demoted to a backstop — consider it only if deletion alone leaves answers too long.

The obligation inventory, audited (from a real consider payload)

One call emits 15 must_state items + unchecked list + case_required + disclosure sentences. Exhibit A for the audit: basis.state_version (a content-hash) sits on the answer-owes list — machine plumbing conflated with human-owed content.

Keeps (a stated must-have reason each)

1. Stance + the deciding reason (case_required rec+support) — the product's point.
2. rule_effects when non-empty — silence = helping the user break their own rule (owner-live incident 2026-08-02).
3. The one or two consequence numbers that would flip THIS decision — that is the answer, not a disclosure.
4. A falsifier attached to any directional recommendation — the only element the owner singled out as praise in the [implementation·skill] Move limits to claim/write/action boundaries — free reasoning, research and recommendation #827 blind A/B.
5. "recorded a consideration ≠ executed" one-liner, only when a canonical write happened.
6. Source + as-of on public facts (compact, inline).
7. One warning line when pricing is degraded (cost-basis case can invert position sizes — real fixture evidence).

Deletes (no defensible must-have; per-item blast-radius noted)

• basis.state_version hash — never human-relevant; stays in the payload for machines.
• Basis four-piece recital every answer (source / completeness / stale_days / as_of) — say staleness only when it could change the decision; otherwise one footnote line.
• Concentration family every answer (top3 / ai_pct / max_sector_pct / concentration_triggered) — the direct source of the owner's 「第一大段完全沒有意義」 verdict; appears as a sermon in all four scenes of both arms ("那條線是系統預設，不是你的" — a sentence both arms wrote is itself obligation-discharge). Surfaces only when it is the deciding fact or touches the user's own cap. Machine layer unchanged: user-cap collisions stay mandatory (keep Remove Vincent Yu name from demo (keep distillation, prep multi-KOL) #2); the system-default trigger becomes may-state.
• Cash balance/weight every answer — only when cash is the question or a floor is being crossed.
• unchecked enumeration (liquidity/valuation/tax/position_fit) — payload stops emitting it, or SKILL forbids enumerating; a material gap earns one footnote line (「我沒查估值」 appeared twice per answer in the A/B).
• Standalone counter-case section — the counter-side exists only as the falsifier line (keep [P1] α 誠實化：樣本不足時不准用「真本事」語氣出 alpha（雙閘門） #4). counter_case: when_material in practice produced a section every time.
• Basis/caliber narrated in body AND footer — body copy deleted; one footnote block, one line per item (disclosures_display retargets there).

The structural flip

From whitelist obligations ("state these 15") to blacklist prohibitions + decision relevance ("never misstate or invent — machine-gated, zero words; what gets said is chosen by what this decision needs"). All integrity gates (engine-owned numbers, no source laundering, explicit-intent canonical writes incl. #829, no execution claims, privacy) are untouched — they cost no reading.

Named risk of deletion

Selection of "the deciding fact" moves from a list to model judgment — the failure mode is omitting a consequence that mattered. Backstops: user-rule collisions remain hard-mandated; eval side gains fail cases for "deciding fact omitted" alongside the existing bloat-fails-closed cases; owner-live witness remains the verdict.

Estimated effect on the #827 A/B answers: ~1,500–2,700 zh chars → ~300–600.

Implementation surfaces update: evaluation_challenge.py (must_state generation — split machine anchors from human-owed content), references/trade-consequence.md, skills/fomo-kernel/SKILL.md, docs/expression-contract.md D-series, references/freeform-answers.md, tests + episode bank. Implementation in a fresh session against latest main, per the maintainer route.

[atomchung]

Target answer shapes (acceptance templates) + claim

Claiming for implementation (dispatched this session, isolated worktree off latest main). These three templates are the acceptance shape — implementation turns them into witness fixtures. All issuers fictional per the tests/agent convention (Widgetron WDGT / Gridcore GRDC / Fabrion FABR); the real-data versions that the owner approved live private-local.

Template 1 — three-way comparison (was ~2,300 zh chars in the A/B; target ~280)

**三個裡我會選 GRDC 加 15 股。**決定性理由：三案對組合的影響都在一個百分點內——誰都不改變你的集中度——真正有差的只有事件風險：WDGT 六天後出財報、預期已拉滿（來源, 日期），這時把最大倉再加大，是三案裡波動最大的；FABR 18 股只佔 1.2%，公司再好這個大小也改變不了結果。GRDC 下次財報在十月底，中間乾淨，上季主業 +82%（公司財報, 日期）撐著。

反面就一條：前三大會從 51.3% 升到 51.9%（GRDC 本來就是第二大）——嫌集中的話這是三案共同的問題，答案是減碼不是選誰。
會讓我改口：你本來就想賭財報超預期——那 WDGT 反而是最直接的表達，排序整個反過來。

（帳本 M/DD、價格 M/DD 收盤；三案動用 $4.4K／$5.1K／$4.9K；估值未評）

Template 2 — recording a selection (was ~550; target ~3 lines)

**記好了：GRDC 買 15 股（@M/DD 收盤 xxx.xx）已存成一筆考慮紀錄——記的是「你考慮過並選了它」，不是「已成交」；真的成交後把交易紀錄丟給我接上。**比較用的另外兩案沒有留下任何正式紀錄。

Template 3 — no-book single-name decision (target ~300)

**WDGT 這家公司的證據支持買，但「現在進場」我不背書——缺的是估值，不是基本面。**上季營收 +24%、EPS 超預期、同日上修全年（公司新聞稿, 日期），這是硬的；但這些已公開三週，而我查不到現在的估值倍數，「好消息是否已在價格裡」這一半我答不了。

$5,000 試水溫，真正的洞在出場：「故事變了就賣」跟你的進場理由是同一個變數，等於沒有獨立的認錯線。可檢查的替代就用它自己簽的支票：全年營收財測或毛利率轉向，擇一寫下來，就可以進場。

（價 M/DD 收盤 xxx.xx；你未提供持倉，部位佔比與重疊無法評）

The one grammar rule behind all three

A fact lives on exactly one floor: deciding facts → opening body; truth-critical qualifiers → parenthetical beside their number; sources/caliber/material gaps → one compact end-block, one line each, non-narrative; machine anchors (hashes, state versions, validator detail) → payload only, never rendered. A fact appearing on two floors is a bug. Completeness lives in the data layer — the engine still computes everything; the user can ask to expand (「其他數字？」) at any time.