core/docs/plans/grammar-unification-2026-07-26.md

Grammar Unification — Plan of Record

Date: 2026-07-26 · Status: ACTIVE · Base: main @ 9696443a (post ADR-0264 ratification, PR #127)

One sentence: CORE cannot read its own writing — measured 0/280 — and this arc makes reading and writing share one grammar so that every construction taught works in both directions, with round-trip as the falsifiable gate.

---

0. Why this arc, and why now

The curriculum-license-loop arc closed on the finding that the outcome-mix ruling is the binding constraint for curriculum serving. That finding stands. It is not, however, the binding constraint on delivered capability, and the measurements in §1 say what is.

Six ADR-scoped deduction bands (v1, v1b, v2-EN, v3-MEM, v4-CM, v5-VP, v6-EX) are ratified, flag-ON, and wrong=0. They are reachable only through looks_like_deductive_argument, which is sentence-initial "therefore":

"If it rains then the ground is wet. It rains. Therefore the ground is wet."
  → Given: if it rains then the ground is wet; it rains.
    Your premises entail: the ground is wet.                    DECIDED

"It rains. So the ground is wet, right?"
  → Pack-resident tokens — pack-grounded (en_core_cognition_v1):
    ground (relation.ground; predicate.support)…                TOKEN DUMP

"All men are mortal and Socrates is a man, so he must be mortal."
  → Pack-resident tokens — pack-grounded (en_core_quantitative_v1):
    all (quantitative.universal.total…)                         TOKEN DUMP

Each band widened what CORE can decide. None widened what a person can ask. Deepening the room behind the keyhole is not progress in delivered capability, and the keyhole is a reading problem.

This arc is scoped to the reading/writing seam. It does not author content, does not move a flag, and does not touch the decision layer.

---

1. Measured baseline

Every number below was measured on main @ 9696443a by reading source and running probes, not inferred from docs. Reproduction scripts land in Phase 1.

1.1 The writer that serves is eight format strings

generate/semantic_templates.py::render_semantic — the function core/cognition/pipeline.py actually calls — is a dict of 8 format strings keyed on IntentTag, plus a predicate-display dict. No morphology, no agreement, no tense, no negation, no pluralization, no clause joining.

Best-case live output:

Q: Explain knowledge. A: Knowledge is what a person knows from truth and evidence. Furthermore, it belongs to cognition.knowledge. In turn, it requires evidence.

One fact per sentence, uniform subject-predicate-object, connectives from a fixed bucket, and an internal dotted path (cognition.knowledge) leaking into prose.

1.2 The good realizer exists and never serves

generate/templates.py::render_step (232 lines) has real linguistics — quantifier agreement, mass vs count nouns, irregular morphology, tense/aspect, and via realize_target it joins clauses with and / or / that / which.

It is called from exactly three places, all evals/*/runner.py. core/cognition/pipeline.py calls realize_semantic, never realize_target. This was deliberate — core/cognition/surface_resolution.py records that the realize_semantic path "was granted supremacy by the Shadow Coherence Gate" — but the consequence is that english_fluency_ood 117/117 + 39/39 scores a function that never speaks.

1.3 Linguistic knowledge is duplicated 35-vs-9 ways, Jaccard 0.083

	word-tables	distinct words
reading path (9 files)	35	202
writing path (6 files)	9	359
shared	—	43 of 518 union (Jaccard 0.083)

Tables that should agree and do not:

linguistic fact	copies	status
irregular plurals	3 (member.py 29, reader.py 8, templates.py 25)	all three diverge
quantifier tokens	5 across both paths	3-way divergence
connective tokens	4 (english, member, verb, cond_member)	3 identical, english adds therefore
sentential-negation prefix	2 (english, member)	identical
negation-bearing tokens	2	diverge by 'not'
predicate display	2 (semantic_templates, templates)	identical, 26 entries
copula forms	2	same content, different type (frozenset vs tuple)

Demonstrated consequences of the plural divergence:

• CORE reads wolves / knives / leaves / thieves / halves / loaves / elves / cacti / fungi / dice / aircraft / offspring and cannot write them — pluralize("cactus") → "cactuses", pluralize("die") → "dies".
• CORE writes criterion→criteria, phenomenon→phenomena, analysis→analyses, hypothesis→hypotheses, axis→axes, basis→bases, thesis→theses and the reader's table knows none of those plurals.

Note this duplication is not only read-vs-write: two of the three plural tables are both on the reading side, and four connective tables are all on the reading side. The six deduction bands were built as successive independent tiers, each re-encoding English from scratch.

1.4 A demonstrated grammar defect: 9 of 26 predicates mis-agree

_inflect_predicate(display, plural_subject=True) against a hand-written English oracle:

plural-subject agreement over _PREDICATE_DISPLAY: 17/26 correct, 9 WRONG

  belongs_to             'belongs to'              -> 'belongs to'              want 'belong to'
  causes                 'causes'                  -> 'caus'                    want 'cause'
  contrasts_with         'contrasts with'          -> 'contrasts with'          want 'contrast with'
  has_steps              'has the following steps' -> 'has the following step'   want 'have the following steps'
  is_caused_by           'is caused by'            -> 'is caused by'            want 'are caused by'
  is_defined_as          'is defined as'           -> 'is defined a'            want 'are defined as'
  is_distinguished_from  'is distinguished from'   -> 'is distinguished from'   want 'are distinguished from'
  is_grounded_in         'is grounded in'          -> 'is grounded in'          want 'are grounded in'
  is_verified_as         'is verified as'          -> 'is verified a'           want 'are verified as'

Live example: render_step(ASSERT, "wolf", "is_defined_as", "mammal", quantifier="all") → "all wolves is defined a mammal".

Two root causes, both single-point:

1. base_form() is written for single verbs and applied to whole humanized predicate phrases. It strips the last character-class of the final word: "is defined as" → "is defined a", "has the following steps" → "has the following step", "causes" → "caus".
2. The copular flag is computed in _inflect_predicate and consulted only on the negated-singular branch. The plural branch (case (_, _, False, True)) ignores it, so is/has/belongs never becomes are/have/belong.

Singular is clean: 0 of 26 differ. This defect reaches zero served turns today because render_step is eval-only (§1.2). It matters because the fluency suite scores this function, and because Phase 4 would wire it into serving.

1.5 The fluency instrument is decorative

evals/deterministic_fluency reports 1.00 on all six predicates. Control test (the absence test — a green measurement that would look identical with the mechanism removed):

candidate surface	passes all 5 content predicates?
"Truth is what corresponds to reality, and knowledge is belief grounded in it."	yes
"colorless green ideas sleep furiously."	yes
"wet ground rains the is."	yes
"is is is is."	yes
"banana does the."	yes

It checks terminal punctuation, presence of a verb-ish token, and two anti-shape regexes. It cannot distinguish English from word salad.

And the suite cannot see §1.4 either. Across all 280 cases in english_fluency_ood + grammatical_coverage:

• 40 use a quantifier — only all and some
• 0 use any copular predicate
• 0 exercise the quantifier × copular combination that breaks

english_fluency_ood's own accept_surfaces include "river flows valley" and "wind shapes dune" — token-ordered, not English. The 100% is real against the lane's contract and says nothing about fluency.

1.6 The headline: round-trip is 0/280, and all six readers refuse

Feed each of the 280 writer-produced surfaces to meaning_graph.comprehend:

surfaces written by CORE : 280
reader COMPREHENDED      : 0/280 = 0.0%
reader REFUSED           : 280/280 = 100.0%
refusal reasons          : {'no_template_match': 280}

Control (proves the harness works): the reader comprehends 31/166 = 18.7% of the deduction_serve corpus, and comprehend("Socrates is a man.") returns 2 entities + member(socrates, man).

Six readers × four writer outputs, including CORE's own live served surfaces — 24 refusals, 0 comprehensions:

writer output	comprehend	english	member	cond_member	verb	exist
Molecule binds enzyme. (writer's own accept_surface)	no_template_match	no_conclusion	out_of_band	out_of_band	no_conclusion	no_conclusion
all wolves is defined a mammal	no_template_match	empty	empty	empty	empty	empty
Knowledge is what a person knows from truth and evidence. (live)	no_template_match	no_conclusion	mixed_structure	out_of_band	mixed_structure	mixed_structure
Wisdom is a good use of knowledge. Furthermore, it belongs to cognition.wisdom. (live)	reserved_word_in_np	membership_shape	out_of_band	out_of_band	out_of_band	out_of_band

1.7 A ratified band serves malformed English today

This was found while designing Phase 1's projection, and it is the most consequential item in §1 because it is user-visible on a ratified, flag-ON band.

generate/proof_chain/render.py:

_CATEGORICAL_PHRASE = {
    "A": "all {s} are {p}",     # plural syntax…
    "E": "no {s} are {p}",
    "I": "some {s} are {p}",
    "O": "some {s} are not {p}",
}

…but {s} / {p} are filled with the singularized entity ids the MeaningGraph reader produces (_chunk_class → _singularize). There is no re-pluralization step, because the reader and the renderer are on opposite sides of the seam described in §1.3. Result, live through deduction_grounded_surface:

in : All dogs are mammals. All mammals are animals. Therefore all dogs are animals.
out: Given: all dog are mammal; all mammal are animal. That's valid — all dog are animal follows.

in : No dogs are cats. All cats are animals. Therefore no dogs are animals.
out: Given: no dog are cat; all cat are animal. That doesn't follow — …

in : All wolves are mammals. All mammals are animals. Therefore all wolves are animals.
out: Given: all wolve are mammal; … — all wolve are animal follows.

Scope, measured two ways:

• Synthetically: every plural noun, regular or irregular. all dog are mammal, all cat are mammal, all student are mammal, all child are mammal, all man are mammal.
• In the ratified corpus: 4 of 47 cases that serve a categorical clause — ds-v1-0023, ds-v1-0024, ds-v1-0026, ds-v1-0028, all band v1. The corpus under-represents the defect only because band routing sends most arguments to renderers that echo the user's original text.

wrong=0 is not threatened. Every verdict above is correct; this is purely a surface defect. Two distinct causes compound in it:

1. No re-pluralization at render time — affects all nouns. This is the dominant cause.
2. The reader's plural table is 8 entries where the v3-MEM band's is 29, so wolves → wolve, knives → knive, leaves → leave, lives → live, halves → halve, loaves → loave, thieves → thieve, elves → elve — 8 silently wrong singulars, and 4 more (cacti, dice, fungi, oxen) that refuse outright. CORE's two readers disagree on 12 of 20 plurals.

The wrong singulars do not break soundness, because _singularize is a deterministic function applied uniformly to premises and conclusion, so the lowering stays a consistent bijection. They are label defects — and they are the reason cause (2) must be fixed with a shared table rather than by patching one side.

1.8 The models are not type-compatible

A literal read(write(g)) == g is not expressible today:

• MeaningGraph = Entity(entity_id, name, span, kind) + Relation(predicate, arguments: tuple[str, ...], span, negated) — n-ary predicate/argument structure with source spans.
• PropositionGraph = GraphNode(node_id, subject, predicate, obj, source_intent, language, root, morphology_id) + GraphEdge(source, target, relation) — fixed SPO triples with discourse edges.

Round-trip therefore needs a defined projection, not identity. Choosing that projection is Phase 1's first design decision, and §6 records why this matters for direction.

1.9 §1.3 re-measured during Phase 2A — mostly subsets, not contradictions

§1.3 was counted by name and by table size. Reading the actual values while building generate/lexicon.py changed the picture materially, and in CORE's favour: almost nothing contradicts.

§1.3 claim	measured	correction
irregular plurals: 3 copies, all diverge	1 pluralizer + 2 singularizers	direction confusion. templates is singular→plural; member/reader are plural→singular. Comparing them as copies is a category error.
—	reader ⊂ member, difference empty	no disagreement. The reader's 8 values all agree; only its coverage is short.
quantifier tokens: 5 copies, 3-way divergence	3 distinct facts	QUANTIFIER_LEAD (can lead a clause), QUANTIFIER_TOKENS (⊃ lead, adds pronouns), PLURAL_QUANTIFIERS (forces plural agreement). every/each lead but take singular nouns, so their absence from the third is correct.
connectives: 4 copies, english adds therefore	3 identical + 1 derived	confirmed; the 4th is english._STRUCTURAL, exactly CONNECTIVES ∪ {therefore}.
negation-bearing: 2, diverge by 'not'	english ⊂ member	confirmed, and the difference is principled: v2-EN normalizes <copula> not, v3-MEM refuses all non-copular not.
copula forms: 2 copies	5 copies	undercounted. The structural test found realizer_guard._BE_AUX and chat/runtime._BE_FORMS — same four words, auxiliary role, invisible to a COPULA-shaped grep.

Exactly one genuine contradiction exists in the whole inventory: discourse_planner._PREDICATE_HUMANIZE maps is_defined_as → "is" where the other two map it to "is defined as". Preserved as a deliberate register choice (the short copula reads as prose mid-paragraph), pinned by test.

⇒ The §1.3 framing "tables that should agree and do not" overstated the disease. The accurate diagnosis is coverage asymmetry plus three facts sharing one name, which is a better problem to have: subsets can be derived from their superset with zero behaviour change, which is why Phase 2A came back 11/11 byte-identical.

Worse than §1.7 stated, though. reader._singularize does not refuse uncovered plurals despite a comment claiming it does — it falls through to a bare -s strip. So news → new and species → specy: precisely the corruptions member.py's table comment says its table exists to prevent. Same fact, one file guarded, the other not. Pinned as current behaviour in tests/test_lexicon_single_source.py; Phase 2B must flip it.

1.10 The destination for morphology already exists and is disconnected

ADR-0258 §5 decided that the number table "moves from module constants to a ratified pack" when a grc_*/he_* member band is built. That trigger has not fired (no such band exists in generate/proof_chain/), so the promotion is correctly deferred — generate/lexicon.py is a staging area, not a competing home, and no new ADR is warranted.

The destination's measured state, so the promotion is not planned on a guess:

• packs/en/morphology.jsonl = 9 records. Seven are forms of be; two are noun plurals (word, beginning) that are regular and exist to support John 1:1. Zero irregular English plurals exist as pack data.
• features.number has no consumer. No code reads it.
• The only reader of morphology.jsonl is chat/pack_resolver.py::_pack_morph_roots_for, which extracts a top-level root key. 0 of 31 records across all four packs define root, so it returns {} on every call while its docstring claims it "enables root-level depth (Hebrew triconsonantal, Greek stems)". A dead path, not a slow one.
• The pack is richer than the code on one point: it has am (en:be:present:1sg), which all five Python be-form copies lack.

---

2. Goal, non-goals, and the thesis check

Goal. One grammar, used in both directions, so that a construction taught once works for reading and for writing, and so that surface diversity becomes verifiable rather than risky.

Non-goals — stated so success is not silently redefined:

• Not "prettier prose." Fluidity is a function of construction-inventory size, which is content work. This arc builds the mechanism that makes growing the inventory safe and measurable.
• Not an LLM, sampler, or paraphraser. No stochastic generation enters the truth path. See the thesis check below.
• Not curriculum content. Scripture/theology content is separately deferred (word-for-word verification + a faith-domain evidentiary regime are both undesigned).
• Not a graph-model merge. §1.8 says that may be needed; §6 makes it a measured decision for a later arc behind an ADR, not a silent expansion of this one.

Thesis check — "decoding, not generating." Every ADR must pass this. A bidirectional grammar is a decoder run in reverse: writing selects among surface forms that the decoder provably maps back to the same graph. Nothing is sampled, nothing is invented, and every emitted variant carries a proof of meaning-preservation. Diversity comes from enumerating verified forms, not from generating candidates and hoping. chat/register_variation.py already implements exactly the deterministic seeded selector this needs (uniform, replay-stable, variant_id in the audit stream).

---

3. The central design decision: the instrument is the fix

An honest fluency measurement cannot be built out of heuristic predicates — §1.5 is what that attempt already produced, and any new heuristic set will be gameable the same way. You cannot measure grammaticality without a grammar.

Therefore the instrument and the remedy are one artifact:

• A grammar that parses is a measurement (word salad fails to parse).
• The same grammar used to generate is the remedy.
• Round-trip forces one artifact to do both, and is falsifiable without a judge, an embedding, or a gold aesthetic.

Necessary, not sufficient — stated up front. Round-trip measures mutual intelligibility, not English quality. "river flows valley" may round-trip perfectly while remaining un-English, because both sides can share an impoverished construction. So the lane needs two corpora and both are load-bearing:

• a positive corpus (round-trip must succeed), and
• a negative corpus of word salad (round-trip must fail, 100%).

If the lane ever reports high scores on both, the lane is broken. That bidirectional requirement is what keeps this instrument from becoming the next §1.5.

---

4. Phases

Each phase is independently valuable, independently landable, and has a falsifiable exit test.

Serving-risk classification — corrected 2026-07-26 after §1.7. An earlier draft of this plan asserted "Phases 1–3 cannot change serving behaviour." That was wrong, and the error is worth recording because it is the same shape as the mistakes this arc exists to fix: I classified a table as off-serving without asking which callers reach it. generate/meaning_graph/reader.py::_IRREGULAR_PLURALS is on the serving path — comprehend → to_syllogism → band v1b → deduction_grounded_surface. Unifying it changes entity ids, which moves trace hashes and served surfaces.

The corrected rule for this arc: a table is off-serving only when every caller is proven to be eval-only. Phase 2 is therefore split by that test, not by topic.

Second correction, same session — the static test is not sufficient. I then computed the serving import closure (227 first-party modules from chat/runtime.py, core/cognition/pipeline.py, chat/deduction_surface.py, chat/curriculum_surface.py, chat/pack_grounding.py, chat/teaching_grounding.py) and all 13 table-owning modules are in it — 13 of 13, including generate/templates.py via pipeline.py → realizer.py → templates.py. A static import test therefore returns "everything is serving" and discriminates nothing.

Import-reachable is not call-reachable. Importing a module executes its table literals (pure data) but not its functions. The call chain realize_target → render_step → _inflect_predicate → base_form is closed and small enough to verify exhaustively, and realize_target has zero non-eval, non-test callers — so that subtree is genuinely call-unreachable from serving, and §1.2/§1.4 stand.

But Python call-reachability is not statically decidable in general, so the arbiter for this arc is empirical, not analytical: make the change, run the 11 pinned lanes, and let byte-identity decide. That is stronger than a static claim because it tests behaviour rather than predicting it, and it is the doctrine the lane pins already encode.

⇒ The 2A/2B split is therefore a result, not a prediction. A change is 2A if the pins come back byte-identical and 2B if they move. Work the changes in the order below, run the pins, and let the split fall out. Do not argue a change into 2A.

phase	changes served output?	gate
1 — instrument	no (new files only)	local smoke/deductive
2A — changes the pins prove inert	no (measured, not predicted)	byte-identical lane SHAs
2B — changes that move a pin	yes	explicit authorization
3 — agreement defect	no (realize_target has zero serving callers)	26/26 oracle
4 — which realizer serves	yes	explicit authorization
5 — round-trip + diversity	yes	explicit authorization

Phase 1 — Build the instrument, record the baseline

Touches: new evals/grammar_roundtrip/, new generate/grammar/ (smallest shared construction inventory), reproduction scripts for every §1 number. Serving risk: none (new files + eval lane only).

Deliver:

1. evals/grammar_roundtrip/ lane reporting, over a positive corpus: write_rate, read_rate, graph_match_rate under the chosen projection.
2. Over a negative corpus (hand-written word salad + programmatically shuffled tokens from positive cases): reject_rate.
3. The chosen MeaningGraph↔PropositionGraph projection, written down with what it deliberately discards.
4. Reproduction scripts for §1.3–§1.7 so every number in this plan is re-runnable.

Exit criteria (all required):

criterion	measure
baseline recorded	read_rate on the 280-case corpus, expected ≈ 0.0%
word salad rejected	reject_rate = 1.00 on the negative corpus
lane is non-vacuous	mutating the reader to accept-everything drives reject_rate → 0 (must go red)
no behaviour change	smoke 621, deductive 349, 11/11 lane SHA pins unchanged

Failure / stop condition: if the negative corpus cannot be rejected without also rejecting the positive corpus, the grammar formulation is wrong. Stop and reassess — do not tune a threshold to split them.

Phase 2A — One source of linguistic truth (the hash-inert subset)

Touches: the word-tables in §1.3 whose unification the lane pins prove inert, plus new single-owner modules. Serving risk: none, and proven by byte-identity.

Deliver:

1. One module owning each linguistic fact: plurals, quantifiers, connectives, copulas, negation-bearing tokens, predicate display.
2. The identical duplicate groups collapsed — 3 connective tables, 2 _SENTENTIAL_NOT, 2 predicate-display tables — with byte-identical lane hashes as the proof of zero behaviour change.
3. Each divergence among off-serving tables resolved by an explicit recorded decision (not by picking the larger table), with a test pinning the answer.
4. A structural test asserting no module defines a second copy — same discipline as tests/test_volume_honesty.py, which must fail if a duplicate is reintroduced.
5. A caller-provenance test: for every table the new module owns, assert whether it is serving-reachable. This is the check whose absence produced the error corrected above, so it belongs in the tree, not in a reviewer's head.

Exit criteria:

criterion	measure
duplication removed	one table per fact; structural test green and mutation-red
identical collapses are inert	11/11 lane SHA pins byte-identical
divergences decided, not averaged	each has a recorded decision + pinning test
agreement	Jaccard → 1.00 for facts that should agree (from 0.083 overall)
serving reachability is known	caller-provenance test covers every owned table

Failure / stop condition: if collapsing an "identical" pair moves a lane hash, the tables were not actually identical — stop, treat it as a divergence requiring a decision, and move it to Phase 2B.

RESULT — the split fell entirely on the 2A side. generate/lexicon.py owns the tables; 11/11 lane SHA pins came back byte-identical, so by the #129 rule every change in this unit is 2A and nothing spilled into 2B. Smoke 621 unchanged, deductive 383 (364 + 19 new).

Two exit criteria needed re-definition rather than just measuring, and both re-definitions are recorded here because the original wording was unmeasurable:

• "Jaccard → 1.00" cannot be the measure. measure_grammar_seam.py scans source literals, so unifying a fact removes it from those files and Jaccard fell, 0.083 → 0.023. That is the success direction, reported by a metric shaped to read like failure. Replaced with an object-identity count: distinct underlying objects behind the consumer names, which is what "one source of truth" actually asserts. Comparing by value cannot distinguish a shared object from two equal copies, so identity is the only honest test. Measured: 3 distinct objects behind 8 names, from 8-behind-8.
• "one table per fact" presumed the §1.3 fact inventory was right. §1.9 shows it was not, so the report now separates _SHOULD_AGREE (must be one object → all UNIFIED) from _RELATED_BUT_DISTINCT (six expected relations, all HOLD: derived-plus-therefore, strict-superset, distinct-fact, differ-by-not, subset, inverse-direction).

The structural test earned its place immediately: it found two copies of the be-form inventory (realizer_guard._BE_AUX, chat/runtime._BE_FORMS) that a name-based grep missed because neither is named like a copula. §1.3's "2 copies" was really 5.

Phase 2B — Serving-path tables and the categorical render defect — authorization gate

Touches: generate/meaning_graph/reader.py::_IRREGULAR_PLURALS, generate/proof_chain/render.py::_CATEGORICAL_PHRASE, the band readers' tables. Changes served surfaces and trace hashes. Depends on 2A's shared pluralizer.

Deliver:

1. Re-pluralization at categorical render time, fixing all dog are mammal → all dogs are mammals (§1.7 cause 1 — affects all nouns).

2. The reader's 8-key view (lexicon.READER_SINGULAR_KEYS) widened to the full 29-entry singularizer. Corrected framing (§1.9): the reader's 8 values are already right — they agree with the full table entry for entry — so this is a coverage fix, not a correctness fix, and there are no "8 silently wrong singulars." The silent wrongness is in the fallback: reader._singularize guesses via a bare -s strip instead of refusing, so wolves→wolve, news→new, species→specy. Its own comment claims it refuses; it does not. Those five are pinned as current behaviour in tests/test_lexicon_single_source.py and this phase must flip them.

   Delivered differently than written: the reader no longer holds a table at all. generate/morphology.py became the single owner of the number rules (2A owned the tables), and the reader calls singularize. That removed a second copy of the regular suffix rules, which was the actual reason the two sides disagreed about coverage.

3. Updated lane SHA pins — surgical single-line edits only, never --update — with the old and new hash recorded per lane.

Exit criteria:

criterion	measure
surfaces well-formed	0 malformed categorical surfaces; ds-v1-0023/0024/0026/0028 render plurals
readers agree	reader-vs-reader singularization agreement 20/20 (from 8/20)
soundness preserved	wrong=0 holds; deduction lane verdicts unchanged case-for-case
hash movement is intended	every moved pin explained; no pin changed that shouldn't move

Why gated: this changes what users see. Fixing a defect is still a serving change, and serving changes are Shay's call. Authorized 2026-07-26.

RESULT.

criterion	measured
surfaces well-formed	0 malformed of 47 (from 4); ds-v1-0023/0024/0026/0028 all render plurals
readers agree	20/20, 0 silently wrong (from 8/20)
soundness preserved	wrong=0 holds on all 7 bands; deductive 403 passed, smoke 621
round-trip	s_surface_match_rate 0.0 → 0.625, and it now equals s_renderable_rate — every renderable surface returns its input exactly
hash movement is intended	criterion void — the pins cannot see surfaces. See below.

The lane SHA pins are blind to served English

2B changed 4 of 47 served surfaces and the pins came back 11/11 byte-identical. That is not evidence of no change; it is evidence the instrument cannot see this kind of change.

Verified two ways rather than argued:

1. By inspection. The deduction_serve_v1 hashed report contains only n, counts, by_gold, correct_by_gold, all_cases_correct, mismatch_examples. No surface prose at all — verdict classifications only.
2. By sabotage (the feedback-ask-what-if-the-thing-were-absent control). Making render._display_noun return "SABOTAGE_" + …, so every categorical clause reads all SABOTAGE_dogs are SABOTAGE_animals:

   guard	caught it?
   11 lane SHA pins	no — 11/11 still byte-identical
   test_deduction_serve_lane + _license (20 tests)	no — 20 passed
   Phase 1 grammar_roundtrip	yes — 3 tests red

⇒ Correction to #129. That PR established the arc's arbiter as "make the change, run the 11 pinned lanes, let byte-identity decide." For values and verdicts that holds. For surface text it decides nothing, so the 2A/2B split it defines is not a partition of "changes users can see." Phase 2A's 11/11 conclusion is still sound — 2A changed no value, confirmed independently by 24/24 table-equality checks against the pre-migration literals — but the justification was thinner than claimed, and 2B is what exposed it.

⇒ Before Phase 1, no test in the tree would have noticed CORE's served prose turning into word salad. That is precisely how all dog are mammal survived on a ratified, flag-ON band with wrong=0 intact for the whole arc. The round-trip lane is not a nice-to-have measurement; it is the only guard that reads what the user reads.

⇒ Carried forward: any future phase touching surfaces must gate on grammar_roundtrip, not on the lane pins. Adding surface text to the pinned lane payloads would be the durable fix and is not done here — it would move every pin at once and deserves its own unit.

The Phase 1 instrument moved on its own, having been built with no knowledge of this fix. That is the whole argument for building the instrument first.

The near-miss worth recording — widening a reader can break soundness. Routing the reader through the full 29-entry singularizer initially produced wrong=1 on band v6-EX: ds-ex-0012 ("No fish are mammals. Therefore some fish are mammals.") answered invalid where gold is refuted.

Cause: _singularize returning None makes the reader refuse, and the serving composer tries the categorical band first, falling through to later, more capable bands. fish previously returned None (it matches no suffix rule), so v6-EX got the case and decided it correctly. Resolving fish made v1b accept a sentence it cannot decide and answer wrongly.

Fix, and it is a principle rather than a patch: a number-invariant form is ambiguous in number — "fish are mammals" is plural, "a fish is a mammal" is singular, and the token cannot tell you which — so a reader that must not guess number declines it. That simultaneously restores the fall-through and is the honest reading. lexicon.INVARIANT_NUMBER is derived from the singularizer's own key == value rows, so the rule cannot drift from the table.

⇒ Coverage and correctness are different axes. Fixing a corrupted value (wolves→wolve) is safe; widening acceptance changes which band answers, and in a first-match composer that can convert a right answer into a wrong one. Same family as ADR-0261 §5.1 refuse-don't-drop. Bundling the two is how a plural fix silently becomes a soundness regression.

Also fixed, found by testing the inverse law: the two number tables were not mutual inverses — the pluralizer lacked cactus→cacti, fungus→fungi, die→dice and the invariants aircraft/means/offspring, so CORE could read cacti but wrote cactuses, and would have written "aircrafts", "meanses", "offsprings". No round trip can close across a table that is not invertible. Pinned by test_the_two_number_directions_are_mutual_inverses.

Deliberately not fixed: mass-noun verb agreement. "all evidence are truth" should be "all evidence is truth", but the copula is fixed text inside the A/E/I/O templates, so agreeing it changes the template shape rather than a slot. No mass noun reaches a categorical clause in any serve corpus today. Recorded in render.py::_display_noun rather than silently left.

Phase 3 — Fix the 9-of-26 agreement defect

Touches: generate/morphology.py::base_form, generate/templates.py::_inflect_predicate, new eval cases. Depends on Phase 2A (unified tables) and Phase 1 (instrument). Serving risk: none (render_step is eval-only).

Deliver:

1. base_form no longer applied to multi-word phrases — phrase-head handling.
2. The copular flag consulted on the plural branch.
3. Eval cases covering quantifier × copular, which is 0 of 280 today.

Exit criteria:

criterion	measure
agreement correct	26/26 against the oracle (from 17/26)
the gap is closed for good	new cases exercise quantifier × copular; reverting either fix goes red
no regression	english_fluency_ood 117/117 + 39/39 holds; smoke/deductive unchanged

Phase 4 — Resolve realize_semantic vs realize_target — authorization gate

Touches: core/cognition/pipeline.py or the eval lanes. First phase that can change served output. Requires explicit authorization before merge.

The current state is untenable in one specific way: 149 green fluency cases score a function that never serves. Exactly one of these must become true:

• (a) realize_target earns the serving path, gated on the Phase 1 round-trip instrument and a wrong=0 hold; or
• (b) the lanes scoring it are retired/re-pointed, and the claim "realizer fluency is mechanistic" is restated as a claim about eval-only code.

Exit criterion: no eval lane scores a non-serving function. Whichever branch is taken is recorded with its reasoning.

Why this is gated: (a) changes what users see and would move surface hashes. It is a serving change and belongs to Shay, not to this plan.

Phase 5 — Raise round-trip, then earn diversity

Depends on: Phases 1–3 plus 2B, and on the Phase 1 baseline being non-zero-able. Scope: deliberately unestimated — see §5.

Deliver:

1. Construction inventory grown until read_rate clears a floor, with the floor ratcheted (never lowered to accommodate a regression).
2. N > 1 verified surface forms per graph, each proven to round-trip to the same graph.
3. Deterministic seeded selection among verified variants, reusing chat/register_variation.py's existing selector, with variant_id in the audit stream.

Exit criteria:

criterion	measure
round-trip improved	read_rate clears a ratcheted floor, well above the ≈0% baseline
diversity is real	mean verified variants per graph > 1, measured
diversity is safe	100% of emitted variants round-trip to the source graph — zero exceptions
truth path untouched	wrong=0 holds; variant selection provably cannot alter a verdict

---

5. What is deliberately not estimated

Phase 5's size cannot be estimated before Phase 1 produces real numbers, and pretending otherwise is how the last arc acquired an unreachable exit criterion. Phases 1–3 are bounded and specified. Phase 5 is scoped only after Phase 3 reports.

I also do not claim Phases 1–4 produce human-like prose. They produce: a measurement that cannot be faked, a single source of linguistic truth, a fixed agreement defect, and an honest answer about which realizer serves. That is the foundation the diversity work needs — and it is worth landing even if Phase 5 is later judged infeasible.

---

6. The decision rule for afterwards

The point of the instrument is that the next direction is chosen by a number, not by argument. After Phase 3, read_rate is re-measured on the unified grammar. That number forks the road:

• read_rate rises materially from unification alone ⇒ the two grammars were closer than they looked, the surface layer was the barrier, and Phase 5 (diversity) is the right next arc.
• read_rate stays near zero after unification ⇒ the barrier is §1.8: MeaningGraph and PropositionGraph are genuinely different models, and no amount of table-sharing bridges them. The next arc is then a graph-model decision, which is large, touches the decision layer, and must go to an ADR before any code. It is explicitly not an extension of this plan.

Either way we will know which, and why, from a measured quantity — which is the thing the previous arc could not do until after it had built the machinery.

One reading of the evidence is worth pre-committing to now: §1.6's uniform no_template_match across all 280, combined with §1.8's type mismatch, makes the second fork more likely than the first. If that is where we land, Phases 1–3 are still exactly the right work — they are what makes the graph-model question answerable instead of speculative — but the arc should be expected to end at the ADR rather than at fluid prose.

---

7. Verification protocol

Per AGENTS.md, unchanged from prior arcs:

• Every phase, in-worktree on canonical CPython 3.12.13 with uv sync --locked, before push: uv run core test --suite smoke -q then --suite deductive -q. Baselines: smoke 621 / deductive 349.
• Every new gate must fail under mutation. A pin that cannot fail is worthless — this arc exists because three such pins were found (§1.5, the Rust parity lane, and the fluency suite's blind spot).
• Phase 2A additionally requires byte-identical 11/11 lane SHA pins. Never scripts/verify_lane_shas.py --update; surgical single-line edits only.
• One PR per phase, branch + worktree each, Forgejo compare URL, [Verification]: line. No merge automation — PRs sit until authorized.
• Phase 4 additionally requires explicit serving-change authorization.

8. Standing traps carried into this arc

• then / therefore are taught philosophy_theology lemmas AND the argument reader's control words. The vocabulary boundary is never screened against the reader's grammar. A unified grammar makes this screenable for the first time — and makes forgetting to screen it a single-point failure. Add the screen in Phase 2A.
• Corpus-mutating tests must never write committed files: the full suite injects -n auto, so writes race.
• Never "fix" a hash drift or a ledger inventory to make a test pass.
• PYO3_PYTHON=/opt/homebrew/bin/python3.12 is required for core-rs builds.