Image Sense Gate — design (PHON-205)

Date: 2026-07-23 Ticket: PHON-205 — Root-inherited images misrepresent lexicalized inflections (e.g. rights → right arrow) Status: design approved, ready for implementation planning

Problem

The picture card for rights shows a right-arrow. rights inherits its lemma's image through the root_inherited tier in packages/data/src/phonolex_data/runtime/images.py, and the arrow teaches the wrong word→image association in a child-facing SLP tool.

The ticket framed this as lexicalized inflection — an inflected form whose dominant sense has drifted from its lemma's. That framing was tested against all 10,638 root_inherited pairs and does not isolate the defect.

What the probe measured

Signal	Coverage	Flags	Behaviour
Inflection is its own WordNet lemma, synsets disjoint from root	100%	1,889 (17.8%)	catches arms, glasses, spirits, papers; swamped by abandoned, accepted, absorbing — participial adjectives WordNet lists separately, where the inherited image is correct
SemCor dominant sense of inflection ≠ dominant sense of root	40.6% both-covered	2,314 (21.8%)	dominated by made, told, asked, came — a different verb sense of the same verb, where the image is still correct

Neither flags rights. SemCor gives the dominant sense of surface rights as right.n.01 (12 of 13 tokens) and the dominant sense of surface right as right.n.01 as well. There is no inflection-vs-root divergence to detect.

The actual defect

right.n.01 is "an abstract idea of that which is due to a person or governmental body by law or tradition or nature" — the entitlement sense, which WordNet ranks first for the bare word right.

The shipped card's own provenance in data/mappings/word_to_image.tsv:

right   right   mulberry    CC-BY-SA-4.0    exact   Noun    Descriptive Direction   0       1

Mulberry declares the symbol depicts direction — right.n.02, rank 2 of 8 noun senses. So right's direct card already depicts a non-dominant sense. Root inheritance did not introduce the error; it faithfully propagated one that was already present.

The comparison that isolates the defect is not inflection-vs-root. It is the sense the card depicts vs. the dominant sense of the word it is shipped for.

The third term — what the picture shows — is not in WordNet. It lives in Mulberry's symbol_category, OpenMoji's annotation, and our generated cards' cardspec gloss/depiction. Supplying it is the only genuinely new work.

This comparison is strictly better for the reproducibility requirement on two counts. WordNet orders synsets by frequency, so wn.synsets(w)[0] yields a dominant sense at 100% coverage with no SemCor sparsity problem. And the check applies to direct cards too, so future tranches are validated by construction rather than audited after the fact.

Goal

No word ships a picture card depicting a sense it does not carry. Enforced by a deterministic rule over committed artifacts that re-runs on every future image tranche without a fresh audit.

Design

1. The gate

A verdict per (word, image) pair from three terms:

• card_synset — the WordNet synset the shipped picture depicts
• word_senses — WordNet synsets for the surface form, frequency-ordered, refined by SemCor surface-form counts where there are ≥3 tokens (SemCor is surface-indexed, so it distinguishes rights from right; WordNet ordering is lemma-indexed and cannot)
• verdict — compatible | incompatible

Compatibility is graded, not synset equality. A pair is compatible when the card synset is the dominant synset, is within a bounded number of hypernym/hyponym hops of it, or shares a close common hypernym. Equality would be far too strict: papers (dominant document.n.01) vs a paper.n.01 sheet-of-material card sits in a different subtree and flags correctly, while cats vs cat.n.01 does not.

Asymmetric strictness

Strictness differs by tier. This is the load-bearing decision.

• Inherited cards (10,638 pairs) — strict. Flag when card_synset is not compatible with the inflection's dominant sense.
• Direct cards (~11.4K pairs) — permissive. Flag only when card_synset falls outside the word's top-K senses entirely.

The asymmetry keeps right's arrow (direction is rank 2, inside top-K) while still flagging rights (card right.n.02 vs dominant right.n.01). Without it, WordNet's adult-register frequency ordering would strip child-dominant senses off cards that already passed PHON-185 rubric review.

K is a tunable constant in the gate module with a single documented default; the implementation plan sets it from the alignment-pass results.

Mandatory guards

Prototyping on the 3,150 shipped Mulberry pairs showed the bare card-sense-vs-dominant-sense comparison flags 976 pairs (31%), overwhelmingly false positives. Three guards are required, not optional; together they bring it to 120 (3.8%).

• G1 — corpus support. Trust a dominant sense only when it has SemCor backing (≥2 surface tokens) or is WordNet rank 1 with a nonzero tagged lemma count. WordNet's frequency order is polluted by rare and technical senses: bikini.n.01 is Bikini Atoll, bends is decompression sickness, a is angstrom, b is bacillus. Without G1 these all flag.
• G2 — derivational bridge. A lexname shift reachable through WordNet derivationally-related forms from the card's synset is a morphological shift, not sense drift: bicycling (noun.act) inheriting the bicycle artifact card is correct, as is arrests (noun.act) from a Verb card. The bridge must originate at the card synset, never at all senses of the card word — the looser form is what let rights pass in the prototype.
• G3 — abstain. No confident dominant sense yields no verdict. 45% of Mulberry pairs (1,407) abstain, and the gate must leave them untouched. Absence of evidence is never a flag. abstain is a first-class verdict in the committed artifact, distinct from compatible.

Supporting precision signals

Each is gated to the surface class where it is not noise, and raises confidence on a flag rather than producing one alone:

• WordNet plural marker — synset glosses carrying (frequently plural) / (usually plural) etc. 100 lemmas repo-wide, 43 of our root-inherited pairs. Very high precision for the lexicalized-plural class (authorities, depths, conditions, interests, exercises). Gated to plural surface forms — it otherwise fires on channeled, interesting.
• Own-lemma disjointness — gated to nominal plurals, which removes the participial-adjective noise that made it unusable standalone.
• SemCor surface divergence — gated to nouns, which removes the verb-tense noise.

2. Term 1 — establishing what each card depicts

Term 1 must resolve to a synset, not a sense domain. A prototype that approximated it as a set of WordNet lexnames failed on exactly the case this ticket exists for: right has 8 noun senses whose derivational forms span nearly every lexname, so a domain-level match let the direction card through for rights. Domain granularity is adequate for monosemous words and unsafe for polysemous ones — which are precisely the ones that can be sense-wrong.

Provider	Cards	Signal	Method
mulberry	~1,400	symbol_category (110-value taxonomy) + symbol_grammar	deterministic — category maps to candidate lexnames, which disambiguate the word's synsets to one card synset (Descriptive Direction + right → right.n.02)
generated	9,724	cardspec gloss + depiction	LLM alignment batch
openmoji	~280	annotation + category	LLM alignment batch

Mulberry — the stream where dropping a card actually removes third-party coverage — needs no LLM pass at all. Its category taxonomy is a hand-assigned sense declaration already in data/mappings/word_to_image.tsv. Only the generated and OpenMoji streams need alignment, and generated is ours going forward.

Committed as data/mappings/card_senses.tsv with columns word, provider, image_file, synset, confidence, source_signal.

Forward-looking: research/2026-07-12-image-gen-style-probe/bakeoff/cardspec.py gains a synset field so every future tranche declares its sense at generation time and term 1 is free for new cards. This is an emit, not a hard blocking gate on generation — a rule bug must not be able to block a tranche.

3. Pipeline integration

Follows the existing committed-artifact paradigm; images.py stays a reader.

• New dev-local script packages/data/scripts/build_image_sense_gate.py. Requires nltk (WordNet + SemCor). CI never runs it, consistent with build_image_mapping.py and build_generated_image_mapping.py.
• New committed artifact data/mappings/image_sense_verdicts.tsv — word, provider, image_file, card_synset, dominant_synset, relation, verdict, reason. Human-reviewed before it ships, mirroring the PHON-197 denylist pattern. The rule proposes; it does not auto-apply.
• One change in images.py — filter incompatible pairs out of the shipped pool before preferred-selection and before root inheritance, so a dropped root card cannot be inherited by anything.
• Fallthrough is then automatic: the word drops to the next-best provider, and to imageless if none survives.

4. Remediation batch

Flagged words split by the cardspec's existing depictability call:

• depictable → generate via the existing gen_tranche.py lane → add to data/mappings/generated_word_images.tsv. The direct match then beats inheritance automatically with no further logic.
• UNDEPICTABLE → add to data/mappings/image_policy_blocklist.tsv with reason=sense-divergent. No card ships. A word that cannot be depicted correctly is left imageless rather than wrongly imaged.
• Pixels upload via the existing packages/web/workers/scripts/upload-generated-images-r2.py; the CI coverage verify gate (verify-generated-images-r2.py) already covers the new cards.

5. Validation

• Vision spot-check over the flagged set before verdicts are committed — a model is shown each card with the word it ships for and asked whether it fits. An accuracy check on the rule, not the mechanism.
• Unit tests in packages/data/tests/ with hermetic fixtures for the gate logic: compatibility relation, asymmetric strictness, each supporting signal's surface-class gating.
• Regression assertions: rights no longer inherits right.svg; cats still inherits cat's card; a blocklisted word still ships no card from any provider.

Scope

In: the gate rule, the card-sense alignment pass, the images.py filter, the committed verdicts artifact, the remediation batch for flagged words, the cardspec synset emit.

Out: image quality review (PHON-185's rubric, orthogonal — the gate checks sense, not whether the picture is good); new coverage tranches (PHON-188); changes to the canonical rule or root morphology.

Blast radius

Third-party (Mulberry + OpenMoji) coverage is 3,674 shipped pairs — 1,561 exact, 1,992 root-inherited, 121 other tiers. Mulberry and our generated set overlap on 3 words, so a dropped third-party card almost never has a fallback: 88% of third-party pairs (3,239) would go imageless rather than fall through to another provider.

Bounds on how many are actually dropped:

Bound	Pairs	Basis
Ceiling	3,674	every third-party pair
At-risk	2,648 (72%)	polysemous and senses span ≥2 lexnames — i.e. a wrong sense exists to pick
Measured, guarded gate (Mulberry)	120 (3.8%)	53 inherited + 62 direct, from the prototype on all 3,150 Mulberry pairs
Supporting-signal lower bound	222	inherited pairs tripping plural-marker / own-lemma / SemCor divergence; 209 without fallback

The realistic expectation is low hundreds, all regenerable through the existing gen_tranche.py lane. The at-risk figure is the ceiling only if the gate were strict everywhere — which is why the tier asymmetry matters. The 1,193 at-risk direct Mulberry cards include break (75 senses), cut (70), run (57), play (52), take (44); a uniformly strict rule would put all of them in question, while the top-K permissive tier leaves them alone.

Accepted posture: dropping third-party cards is desirable in itself, since the long-term direction is away from third-party assets. Words that lose a card and cannot be correctly regenerated stay imageless.

Risks

• WordNet frequency order reflects adult written register. No sense-tagged child corpus exists to correct it — CYP-LEX and the FineWeb-Edu grade bands are word-level, not sense-level. Mitigated by asymmetric strictness and by the verdicts file being human-reviewed rather than auto-applied. Accepted, not eliminated.
• The gate checks sense, not depiction quality. A card can pass the gate and still be a poor picture.
• The gate is only as good as term 1. The Mulberry prototype is deterministic and measurable today; the generated stream's alignment quality is unverified until the batch runs. The plan must checkpoint after alignment and sample-verify before any verdict ships.
• G3 abstains on ~45% of pairs. The gate is silent on a large minority — it reduces wrong cards, it does not certify the rest as correct.
• Mulberry cards will be dropped with no fallback. 88% of third-party pairs have no other provider, so a drop means imageless until regenerated. Accepted (see Blast radius), but it means the remediation batch should land in the same release as the gate, not after it.

Acceptance

rights, and similarly sense-divergent forms, either carry a correct image or no image — never the lemma's misleading one. The rule is a committed script plus a committed verdicts artifact, re-runnable on each future tranche without a fresh audit.

References

• Ticket: PHON-205. Related: PHON-162 (mapping), PHON-184 (root population — stale, root is populated), PHON-185/187 (generated pipeline), PHON-193 (WordNet sense signal in the canonical rule), PHON-197 (curated-denylist pattern).
• Ship policy and inheritance: packages/data/src/phonolex_data/runtime/images.py
• Generation lane: research/2026-07-12-image-gen-style-probe/bakeoff/gen_tranche.py