Grow outward from the confident island, at whatever level and direction looks most promising.
Why the other options fail:
- “Strict left to right” reintroduces the pipeline lesson 4 just ruled out. Position on the tape has nothing to do with where the evidence is strongest; ignoring a confident mid-sentence hypothesis to grind through weak data at the start throws away the one thing blackboards buy you.
- “Discard the segment level” confuses “less confident right now” with “not needed.” The segment level still has to be resolved eventually — the point of the island is that the strong word hypothesis now helps resolve it, by telling weak segment-level KSs which few phones to favor instead of searching blind. Nothing is thrown away; a search is narrowed.
- “Wait for every segment” is the pipeline’s assumption in a different costume — that lower levels must fully resolve before higher ones can begin. Bidirectional inference exists precisely so that never has to happen.
The mechanics, concretely. Picture the utterance “the seven items” with segment-level noise scrambling the start but a strong syllable/word match for “seven” landing cleanly in the middle (vowel + fricative patterns for “seven” are acoustically distinctive; “the” said quickly is mush). A lexical KS posts a high-confidence WORD hypothesis for “seven,” linked down to the syllable hypotheses that support it. That hypothesis is now an island. From here, two directions of work both become newly worth scheduling:
- Downward, into the noise: a segment-level KS re-examines the mushy start, now armed with the knowledge that whatever’s there, it’s followed by “seven” — a grammar KS can suggest “probably a determiner or number word here,” collapsing the segment KS’s search space.
- Upward, into structure: a syntax KS proposes phrase hypotheses built around “seven,” and a semantic KS asks what kind of phrase makes sense in the application domain (a database query system expects “the seven items” to be a quantity-and-noun phrase, which further predicts what comes next).
Each level’s hypothesis is a claim about the same signal at a different grain, and the support links between levels are what let confidence at one level migrate to another — a confident word makes its component segments confident by association, the way lesson 4’s “island” metaphor implies: solid ground you build outward from, not a single resolved answer arrived at once.
Pitfall worth naming: islands of certainty is a search heuristic, not a guarantee. A confident-but-wrong hypothesis (the acoustics sound like “seven” but the speaker actually said “heaven”) can seed a wrong island that then wrongly constrains everything grown from it — bad confidence compounds exactly as fast as good confidence does. Nothing in the architecture prevents this; it’s why real systems keep multiple competing hypotheses alive per level rather than committing early, and why the next lesson — control — is the part that decides how much work to spend hedging against exactly this risk.
Where this goes: hypothesis levels tell you what the board holds. The next lesson opens control: the agenda that scores every pending piece of work — extend this island, doubt that hypothesis, try a competing one — and picks what runs next.