Existing F711 Predicate Realization
F898 addressed the predicate-realization gap left after F897. On the frozen 28-case meaning set, the candidate improved complete F711 coverage from 14/28 to 28/28; a separate blind multi-fact set passed 20/20, and the recorded cold reload passed.
Objective and research question
Objective
Extend the existing natural lexicalization path so supported semantic predicates in the frozen F896 meanings could be realized through learned syntactic families without storing complete blind answers.
Research question
Can the existing F711 path close the remaining predicate-surface gap while preserving arguments, proposition order, blind multi-fact coverage, anti-memorization checks, and cold-reload behavior?
Context from F896 and F897
F896 established that the structured planner could work while the full natural-language pipeline remained insufficient.
F897 then repaired atomic fact-contract transport and noun-phrase composition. That phase moved the frozen surface gate from 0/28 to 14/28 but left 14 meanings without complete predicate realization.
F898 targeted that remaining lexicalization boundary. It did not retest unrestricted natural-language understanding or convert the earlier F896 gate into a pass.
Method
- The candidate was compared with a 14/28 baseline on the same 28 frozen structured meaning representations. This gate evaluated surface-contract readiness and did not generate final Decoder answers for those 28 cases.
- The correction was learned from at least two independent structured records per supported pattern and was evaluated without complete blind sentences stored in the candidate.
- A separate frozen 20-case oracle multi-fact set injected structured evidence directly, generated downstream output, and checked F711 readiness, proposition coverage, order, and duplicate proposition identifiers. It did not test free-text text-to-meaning.
- The phase also checked one live database example, anti-memorization conditions, protected-state integrity, and candidate reload.
Implemented correction
The existing lexicalization path was extended with learned associations between semantic predicates and syntactic realization families.
This supplied surface forms for the semantic relations that remained non-lexical after F897 while preserving the already working noun phrases and arguments.
The report records no complete memorized sentences in the candidate and no modification to the downstream decoder.
Evidence and benchmarks
| Measure | Result | Scope | Evidence |
|---|---|---|---|
| Frozen baseline | 14/28 | Structured surface-contract readiness before the F898 candidate; no final answers generated in this gate. | F898 primary audit |
| Frozen candidate | 28/28 | Structured F711 surface readiness on the same 28 frozen meanings; not an end-to-end answer score. | F898 primary audit |
| Blind multi-fact evaluation | 20/20 | Separate frozen oracle multi-fact set with structured evidence; readiness, coverage, and order each 1.000000. | F898 primary audit |
| Argument preservation | 28/28 | Recorded argument-preservation rate 1.000000 on the frozen candidate set. | F898 primary audit |
| Invented predicates | 0 | No invented predicates recorded in the frozen candidate evaluation. | F898 primary audit |
| Cold reload | PASS | Candidate identity and reloaded schema/coordination state matched; not a rerun of every end-to-end gate. | F898 primary audit |
Results
Structured surface-contract readiness on the frozen 28-case meaning set improved from 14/28 to 28/28. Noun-phrase realization and argument preservation remained 1.000000. These 28 cases did not generate final answer text.
The separate oracle multi-fact set passed 20/20 with full F711 readiness, generated proposition coverage, and order; duplicate proposition identifiers remained zero. Structured evidence was supplied directly, so this was not a free-text understanding test.
The report recorded zero invented predicates and zero incompatible coordinations in the frozen candidate evaluation.
The scoped phase gate passed. The lexical candidate itself was not promoted to an operational checkpoint.
Recorded source outcome: F898_EXISTING_F711_PREDICATE_REALIZATION_PASS. The shorter verdict badge above is a conservative public label, not a replacement for the source decision.
What was validated
- The defined predicate-realization gate for all 28 frozen supported meanings.
- Preservation of the recorded arguments on those frozen meanings.
- Coverage and order on the separate 20-case blind multi-fact evaluation.
- The phase-specific anti-memorization check and cold-reload gate.
- One recorded live database control within this evidence package.
What remains unproven
- Unrestricted natural-language understanding or open-domain text-to-meaning.
- Universal predicate, vocabulary, grammar, or discourse coverage.
- Human-level language or reasoning.
- General intelligence, consciousness, or superiority over Transformer systems.
- Production promotion of the candidate or validation of every downstream chat path.
Known limitations
- The 28-case result applies to frozen meaning representations and a defined set of predicate-realization gaps.
- The separate 20-case oracle multi-fact result covers its own structured-evidence dataset, not arbitrary natural-language questions.
- F898 validates a downstream lexicalization correction; upstream text-to-meaning weaknesses documented by F896 remain outside this gate.
- The structural gate was not a universal grammar or discourse-quality test; later evidence continued to address awkward surface and multi-clause composition.
- The candidate was not promoted, so the report must not be described as a universal production upgrade.
Cold-reload result
- The source records candidate reload as PASS: candidate identity and the reloaded schema and coordination state matched.
- The anti-memorization check reports zero complete memorized sentences.
- Operational checkpoints and protected files were reported unchanged during the test.
- Reload success was not a complete rerun of every end-to-end gate and does not establish unrestricted generalization.
Public evidence and source figures
No original source figure is present in the current F898 public evidence package. The metric chart is an editorial HTML rendering of the verified baseline, candidate, and blind-set values; it is not a new benchmark or source figure.
F898 primary audit
Baseline and candidate comparison, blind multi-fact gate, correction scope, live control, anti-memorization, integrity, cold reload, and terminal decision.
F897 bridge report
The 14/28 incoming baseline and the transport and noun-phrase corrections that preceded F898.