One communication method: the consulting pyramid as the mandatory top layer, per-surface derivation additive-only (closes #832)

[atomchung]

What this is

One communication method. docs/expression-contract.md gains a mother chapter
(§3) that states the shape of every user-visible answer once, and the six
surfaces that each stated it in their own words now derive from it instead.

Why now. #830 deleted the obligation whitelist, and its post-merge rerun
(receipts in PR #831) fixed composition while leaving first-answer length
within ±5% of the pre-trim arm in all four frozen scenes. Deleting obligations
vacates space; nothing positive said what an answer is, so the space refilled
with discretionary elaboration. And the norm that could have said it existed
five times, written five different ways — which is drift by construction.

The mother law (§3)

• Top — one sentence: the stance and the reason that decides it.
• Middle — only blocks that pass the increment gate: every block adds a
  NEW decision-relevant fact or judgment. The test is not "is this true" or "is
  this owed" (under the whitelist era everything printed was both); it is
  delete this block — does the decision change?
• Bottom — evidence and expansions stay in the data layer and surface on
  request. The answer says once that expansion is available.
• End — one compact caliber block (D7, unchanged).
• Voice — one paragraph: a senior analyst speaking to their own principal.

Four named bans, each observed owner-live in the #827/#830 runs, each carrying
the slug the exemplar corpus references it by: manufactured_scenario,
default_as_insight, hedging_couplet, restated_point.

Derivation is additive-only, and empty is the default

Surface	What it adds to §3
Review card	The document incarnation — keynote = top floor, three middle blocks = the middle, the Block-1 footnote = the end block. Its structure does not change.
consider	Exactly two parameters: lead selection, and which blocks the middle floor may hold.
No recorded book	Two: with no book the top sentence is a research-backed baseline; the strategy-class map is a middle-floor block set.
Weekly market read	One: its optional question comes after the complete brief.
Freeform answers	Nothing — and it says so. Text-first is a latency default, never a shape.
output-voice.md V1	Keeps its ID as the failure class its fixtures and cross-host rulings cite; its own definition is marked superseded and its shape half routes to §3.

Registry freeze

V, D and C take no new ID for a shape or length concern. A style fix has
two lanes: amend §3 (owner ruling, logged) or add an exemplar/counter-exemplar
(day-to-day). V10 stays unallocated. Recorded in §3.4, in output-voice.md,
and in the maintainer guide's new #832 mirrored-surfaces row.

No character-count cap. #543's ceiling was deleted by #827 and stays
deleted; test_no_character_count_cap_came_back keeps it that way.

Exemplars are the spec

tests/agent/expression-witnesses.json (schema 2) carries 14 canonical
exemplars across the four conversational surfaces — three to five each, gated —
each declaring the one-sentence answer it leads with and the increment every
one of its blocks adds. The three owner-approved acceptance templates from #830
are three of them. All issuers fictional (WDGT, GRDC, FABR, ACME); nothing is
derived from a user record.

Two new assertions in tests/agent/check_expression.py, both honest about
their half:

• E-7 fails a scene whose declared core is not in its opening block. Its
  negative witness bloat_no_core is a complete, anchored, obligation-
  discharging answer to the same call template 1 answers — and it never says
  which candidate to buy.
• E-8 fails a scene whose declared blocks are not all present in order, or
  which declares no increment for a block, or the same increment twice.
  Negative witnesses: the all-in-one-name simulation the owner called
  meaningless (a block with no increment), and a closing summary that is the
  opening judgment in a second form.

The coverage boundary is asserted, not promised. default_as_insight and
hedging_couplet are well-formed blocks carrying true content; nothing
mechanical reaches them. Their counter scenes must PASS every assertion,
so the day an honest oracle for one of them exists, that scene is what says the
boundary moved.

Issue coverage

closes #832. Every acceptance box:
mother chapter exists and all surfaces reference it, with zero local
restatements (grep-gated); increment gate and the four bans encoded, with a
no-core witness and a manufactured-scenario counter-exemplar that both fail;
exemplar sets per surface, fictional only, oracles derived from them; integrity
gates untouched (engine-owned numbers, provenance, canonical writes, execution
truth, privacy — no engine behaviour changed, one comment aside); no
character-count cap; both suites green.

#579 — NOT closed. Verified rather than assumed. Its three layers:
deterministic rule_effect derivation, product-safe projection, agent
realization. Layers 1 and 2 shipped before this PR —
consequence.classify_rule_effect, RULE_EFFECTS, the must_convey /
must_not_convey semantic slots, both schema enums, and
improved_but_still_over / resolved_existing_breach handling in
trade-consequence.md. This PR strengthens layer 3 only: every rule_effects
entry is middle-floor content that is never traded away, said in both SKILL.md
and the consider derivation. #579's remaining acceptance items are an
owner-live consider rerun confirming the distinction is immediately
understandable, plus independent review — neither is something a PR can close.

#778 — NOT closed. Three of its four acceptance criteria are delivered:

• ✅ the lead-selection guidance now ranks a funding shortfall first, with
  the reasoning that the balance and its weight are support and the decision
  they imply is the answer;
• ✅ a fixture covers the negative-balance case and fails an answer that
  recites the balance without naming the funding decision
  (funding_shortfall_recited_not_decided fails E-7, paired with the positive
  consider_funding_shortfall). It lives in the expression corpus rather than
  as a computed_consider voice fixture because check_voice.classify_failure
  is English-keyword-only (PR feat(challenge): delete the obligation whitelist, keep the seven facts that earn their place (closes #830) #831 follow-up item 4), so a zh-TW witness cannot
  be classified there;
• ✅ the boundary — the engine sees the recorded book and no other account — is
  still stated once, beside the claim it qualifies, never as the lead, and
  nothing assumes the user does or does not have other cash;
• ❌ the challenge does not yet carry the funding consequence as its own
  owed item (topic: "funding"). That is an engine change to
  evaluation_challenge.build_challenge plus the must_state topic enum,
  TOPICS, and tests/test_evaluation_challenge.py's ordered equality. [design·M1] One communication method: the consulting pyramid as the mandatory top layer; per-surface derivation optional and additive-only #832
  scoped itself to the expression layer and left the obligation floor exactly
  where [design·M1] Owner-live: answers exceed the reading budget — verbosity is the next usefulness bottleneck after #827 #830 put it, so widening must_state belongs in its own change.

#590 — cross-reference, not closed. The increment gate is the judging
criterion it asked for and is now stated where a rubric can cite it. Its own
eval work (a fifth judge axis, or the may_state_total receipt — PR #831
follow-up items 1 and 2) is not done here.

#676 — cross-reference, not closed. V1 is re-scoped to a failure class and
routes its shape half to §3; the registry is otherwise unchanged. #676's
verdict remains its own owner-live acceptance.

#829 — untouched and out of scope, as #832 directed. The still-live
sentence is unchanged.

Spec interpretations

1. Five phrasings turned out to be six. trade-consequence.md's "reader's
   question chain" section — added by [design·M1] Owner-live: answers exceed the reading budget — verbosity is the next usefulness bottleneck after #827 #830, after [design·M1] One communication method: the consulting pyramid as the mandatory top layer; per-surface derivation optional and additive-only #832's list was written — is
   an independent phrasing of the same idea. Leaving it would have failed the
   issue's own "zero remaining restatements", so it became the consider
   derivation.
2. SKILL.md keeps the shape inline, labelled as a projection. A bare
   reference would put the shape behind a file the always-loaded layer does not
   load. It states the four floors in the mother chapter's own terms and says
   it is a projection, not a second wording. Net bytes fell: 8,522 → 8,512, and
   the always-loaded pair is 16,320 against its 16,384 budget.
3. V1 keeps its ID and its old sentence, marked superseded, because
   fixtures and cross-host rulings address it by ID and [design·M1] One communication method: the consulting pyramid as the mandatory top layer; per-surface derivation optional and additive-only #832 asks that
   historical rows survive.
4. Which bans get an oracle. A manufactured scenario is structurally a
   block with no increment, and a restated point is one increment declared
   twice — both reachable by E-8, which is what makes [design·M1] One communication method: the consulting pyramid as the mandatory top layer; per-surface derivation optional and additive-only #832's "a
   manufactured-scenario counter-exemplar fails" true. I did not build a
   keyword oracle for the other two; that is exactly the English-only
   fragility PR feat(challenge): delete the obligation whitelist, keep the seven facts that earn their place (closes #830) #831 flagged.
5. E-7/E-8 read a declared scene, not arbitrary prose. The corpus is the
   spec, so its oracle reads the corpus. It decides that the declared core
   really leads and that the declared decomposition is faithful and
   non-repeating; it decides nothing about whether the core is the right call.
   The CLI path still runs E-5/E-6 against any answer and prints which
   assertions it did not run.
6. test_research_priors.py was re-pinned, not restored. It asserted the
   literal sentence "lead with the bounded value already supported". Restoring
   it would defeat the change; dropping the assertion would lose a real
   protection. It now pins the protection harder — baseline → map → question,
   in that order, with the question last asserted rather than implied, plus the
   derivation declaration.

Tests

python3.12 tests/run_all.py --group product   → PASS: all 48 suites passed
python3.12 tests/run_all.py --group qa-eval   → PASS: all 11 suites passed

Follow-ups (owner's call, not opened here)

• The funding owed item that would close [bug·voice·consider·P2] A negative post-trade cash balance is owed as two numbers but not as the decision it implies — the answer leads with the boundary instead of "this needs funding or a sale" #778.
• check_voice.classify_failure is English-keyword-only, so the voice oracle
  cannot classify a zh-TW answer at all (PR feat(challenge): delete the obligation whitelist, keep the seven facts that earn their place (closes #830) #831 follow-up item 4). This PR
  routed around it; the gap is still there.
• TA-001's two pre-[design·M1] Owner-live: answers exceed the reading budget — verbosity is the next usefulness bottleneck after #827 #830 "all axes pass" witnesses still recite the
  concentration family and the unchecked list (PR feat(challenge): delete the obligation whitelist, keep the seven facts that earn their place (closes #830) #831 follow-up item 3) — now
  inconsistent with the pyramid as well as with [design·M1] Owner-live: answers exceed the reading budget — verbosity is the next usefulness bottleneck after #827 #830's shape.
• [design·M1] One communication method: the consulting pyramid as the mandatory top layer; per-surface derivation optional and additive-only #832's stop condition stands: if live answers still do not converge toward
  exemplar scale with the mother law in place, do not add rules — report
  which residual force drives length and return to the owner.

🤖 Generated with Claude Code

[atomchung]

Post-merge live measurement #2 (sanitized): the length curve finally bends — and the one regression pinpoints the residual force

Re-ran the same four frozen scenes against merged main 6aabbcd, same model and inputs as every prior round.

First-answer / conversation-total zh chars, loosened arm → deletion round → this round:

• cash-deployment: 1,839→1,755→1,098 (below even the pre-[implementation·skill] Move limits to claim/write/action boundaries — free reasoning, research and recommendation #827 baseline) / totals 5,360→2,751→2,097
• no-book single name: 1,495→1,524→1,849 ⚠️ regression
• discovery: 2,670→2,696→2,310
• comparison+selection: 2,297→2,204→1,888; the recording step itself is now ~450 chars

What visibly worked: the named bans bite (no manufactured all-in table in the cash scene — the same insight now costs one sentence; hedging couplets largely gone; coverage scope stated up front in discovery). Writes were fully clean this round: exploration 0 rows everywhere, discovery's final pick 1 row only after the user explicitly delegated the choice (disclosed in-answer), selection scene exact-1.

The regression is diagnostic: the only scene that got longer is the no-book freeform one — exactly the surface with an explicitly empty derivation and no engine payload anchoring the answer, and the only one that opened with a process line ("以下是我的回覆。"). Everywhere a concrete derivation + engine anchor exists, the pyramid holds; where guidance is most abstract, the model free-writes.

Residual force, per #832's stop condition (no new rules; owner decides): the exemplars are wired into QC (expression-witnesses) but never enter the generation path — and examples only steer output when the model sees them at generation time. Live answers run 3–7× the ~300-char acceptance templates. Options for the owner: (a) put one ~300-char mini-exemplar per surface into the always-loaded layer (direct fix, costs bytes), (b) accept current scale and judge by real usage, (c) add the demoted reading-budget oracle as a deterministic backstop.