Status: Living document. New registrations are added as they are made.
Timestamp of record: The Japanese original (Foundation appendix draft in this repository) is the master document. The authoritative timestamps are its git commit history. This English table is a mirror prepared after the fact; where the two diverge, the Japanese original prevails.
The author (a Japanese novelist with no formal training in AI, linguistics, or mathematics) can directly introspect her own semantic space — naming the axes on which near-synonyms differ, reporting the perceptual attributes attached to words, and observing her own category formation in real time. Since 2026-07-06 she has been learning Spanish (and tracking English, Chinese) while logging predictions before the corresponding acquisition events occur.
Each entry below was written down, with date, before the predicted event. Hits and misses are both recorded; a miss is data about the limits of the underlying theory (difference-driven discrete splitting of decision trees; see 7-1-2 in the Foundation appendix), not a failure of the log.
Legend: ✅ confirmed · 🔶 partially confirmed / on track · ⏳ open (awaiting natural encounter — the author deliberately avoids looking up the answer so the encounter stays natural)
| # | Date | Prediction (registered before the event) | Status |
|---|---|---|---|
| 1 | 2026-07-06 (morning) | The memorized Spanish chunk "yo soy + [name]" — currently stored as one undecomposed unit — will split into frame + slot. | ✅ Confirmed the same day: seeing "yo soy Kokko" (her own name — a large visual difference from the memorized Pablo/Luis) triggered the split. Trigger channel: visual difference, consistent with the author's form-first matching architecture. |
| 2 | 2026-07-06 | Once the frame+slot structure exists, further unknown subjects will enter slot-wise, skipping the whole-chunk phase. | 🔶 Partially confirmed: él es / ella es decomposed hours later with no external trigger ("chained split" — the splitting procedure itself transferred, not the example). Exception correctly carved out in #4. |
| 3 | 2026-07-06 | Spanish gender agreement (-o/-a) will be acquired far faster than English 3rd-person -s ever was, because it is a visible one-letter difference grounded in a real-world distinction, whereas -s carries no meaning. | 🔶 On track — but the entry channel was sound, not vision as predicted: "ella es amiga" was adopted because "it sounds right," before any rule was known. (Judgment precedes explanation; the deviation from the predicted channel is itself logged.) |
| 4 | 2026-07-06 | es will detach first (the minimal pair él/ella + es is already in hand); soy will stay fused to yo (no contrast partner exists); tú will require a new whole-chunk phase, because eres is a third verb form not derivable from existing material. | 🔶 Rear half confirmed same day: tú eres entered as an undecomposed chunk, exactly as predicted. es-detachment: pending. |
| 5 | 2026-07-06 | With the frame quieres + [noun] now productive, new nouns will be usable in full sentences after a single exposure (acquisition cost collapses to near zero — the proposed mechanism of the childhood "vocabulary explosion"). Probe nouns issued: vino, cerveza, pan, manzana. | ⏳ Under observation. Side finding: manzana showed three independent layers in one word — sound (in, via la manzana), meaning-grounding (not yet linked to the image of an apple), grammatical gender axis un/una (lagging). |
| 6 | 2026-07-06 | Overgeneralization seed ①: from the current input range (singular + article chunks only), the false rule "singular → article, plural → no article" will form. It must later collide with Spanish plural articles (los/las, unos/unas) and be re-split. | ⏳ Awaiting natural encounter. |
| 7 | 2026-07-06 | Frame mis-remodeling, path specified in advance: on first encountering a plural, the system will conclude "un/una absent → the head isn't needed → delete the article slot." Collision with reality is guaranteed. Test: does the author's provisional-rule tag (rules formed from thin data stay marked "provisional") extend to the frame level, making the re-split painless — or only to the rule level? | ⏳ Awaiting natural encounter. The author has also reported, in real time, the pull toward the wrong axis ("I keep wanting to say una pan" — a first-person report of shortcut learning while resisting it). |
| 8 | 2026-07-06 | English listening (goal: functional comprehension by March 2027): FF14-native English videos (maximal shared situation — her home ground) will outperform generic audio material; listening comprehension will improve substantially while production grammar (3rd-person -s etc.) remains absent — the absence is a prediction of the architecture (order-free graph; morphological markers carry no meaning to ground), not a failure. | ⏳ Evaluation: March 2027. |
- Synesthetic-report retest: word-attribute reports (color, luminosity, motion, distance, temperature, texture, weight) recorded 2026-07; identical prompts to be re-run after 1–2 months without showing prior answers. Genuine perceptual reports are expected to replicate; constructed associations are not.
- Two-person test (heritability granularity): same word lists given independently to the author and her adult daughter. First result (2026-07-05): the differential axis of きれい/美しい ("sparkle-grain") matched — a shared-side attribute, alongside "weight"; texture remains non-shared. Catch-trials use forced choice between the true axis and an AI-generated decoy axis (the author does not have to voice any false perceptual report herself).
An LLM cannot introspect its own embedding space; it can only regenerate descriptions from training data. The author reads her semantic space directly and reports the axes. Pre-registration turns those reports from anecdotes into testable records: written before the event, timestamped by git, falsifiable, with misses kept on the books.
If you are a researcher and want the full context, start with the Foundation appendix draft (Japanese, master copy) in this repository — sections 7-1-2 (the life cycle of the splitting operation: judgment precedes explanation / discrete split vs. continuous automation / meta-learned splitting / retroactive batch decomposition / expertise as resolution / overfitting as an almost-right proxy axis / provisional-rule confidence tags) and 11-13/11-14 (semantic-space isomorphism, communication-cost asymmetry, dual time representation).
Contact: rufeir@gmail.com