Second Guess Strategy: Probe Again or Start Filtering?

The first guess has a whole experiment behind it, and the filtering that follows feedback has its own solver. The decision between them — the second guess — is usually played by feel. This page replaces the feel with counts: SATIRE was simulated against all 682 answers in this site’s pool, its real feedback patterns were collected, and for the frequent ones the two options were compared directly. Option one is a probe: a word chosen to test letters, even if it cannot be the answer. Option two is a filter move: guess a word that still fits the board.

The result is not “always probe” or “always filter.” It is a threshold, and the threshold is lower than most players assume.

What one strong opener leaves behind

Running SATIRE against every answer with the site’s two-pass feedback rule produces 211 distinct color patterns. Their sizes are lopsided in a useful way:

• 92 patterns are singletons. For 13.5 percent of answers, the board after guess one already identifies the word — the only correct second guess is the answer itself.
• 78.3 percent of answers land in a bucket of 12 or fewer candidates.
• Only 11.3 percent land in a bucket larger than 20, and the single largest bucket holds 34.

So before any tactics, the base rates say the second guess is a filtering move most of the time. The interesting decisions live in a minority of boards, and the three sections below work through real patterns from small to large. The opener itself is held fixed here; how SATIRE earned that role is documented in the entropy ranking of all 1,485 accepted first guesses.

The all-gray start is smaller than it feels

Six gray tiles look like a disaster. Measured, the pattern is the rarest kind of luck: exactly 5 of 682 answers return all gray against SATIRE — COMMON, COMPLY, FOLLOW, HOLLOW, and UPHOLD — a 0.7 percent chance. Striking out on S, A, T, I, R, and E leaves almost nowhere to hide.

Comparing the two options on this five-word set:

• Filter: guessing FOLLOW, HOLLOW, or UPHOLD splits all five candidates into single-word outcomes. Worst case, guess three is forced and certain — and each carries a 1-in-5 chance of ending the game immediately.
• Probe: the best outside words, such as HERALD, achieve the same perfect split but can never be the answer, so the 20 percent instant win is thrown away.

The filter move strictly dominates. The one all-gray trap is choosing carelessly inside the bucket: COMMON and COMPLY leave a two-word tie between FOLLOW and HOLLOW, turning a guaranteed three-guess game into a possible four.

Two yellows: the one board where a probe earns its slot

SATIRE’s largest bucket is yellow R, yellow E, everything else gray: 34 candidates, hit 5.0 percent of the time, ranging from BORDER and FROZEN to WONDER. This is genuine probe territory — and the simulation says which probe.

The best second guess across all 1,485 accepted words is GOLDEN: it cuts the expected remaining list to 1.65 candidates with a worst branch of 3. Notice what GOLDEN is doing. It is not six fresh letters; it moves the confirmed E to a new slot while testing G, O, L, D, and N. GOLDEN cannot be the answer here — it has no R — so it is a pure probe, but a probe that recycles the yellows.

Two comparisons put that number in context:

• The best candidate guess, CORNER, expects 1.76 remaining with the same worst branch of 3 — nearly identical, plus a 1-in-34 chance of being right outright.
• A blind “all new letters” probe, UPHOLD, expects 2.71 remaining with a worst branch of 6. Ignoring the two yellows costs more than the probe-versus-filter choice itself.

The sharpest finding is inside the bucket: the median candidate expects 2.88 remaining, and the weakest expects 6.24. The gap between probing and filtering well is 0.11 words; the gap between filtering well and filtering on autopilot is over a full word of expected list size.

A green plus two yellows: recycle, don’t reset

A colored board tilts things further toward candidates. Take green A in slot 2 with yellow R and yellow E: 21 candidates (3.1 percent of answers), including DANGER, FARMER, and WARDEN.

The measured best probe is RENDER, at 1.67 expected remaining — again a recycler, re-placing R and E twice each while adding N and D. The best candidate guesses, GARDEN and WARDEN, expect 1.95 with a 1-in-21 win chance. The disjoint probes collapse: UPHOLD expects 4.05 remaining and COMPLY 5.67 — worse than a merely median candidate. Three confirmed letters make “test six new ones” an expensive way to ignore evidence.

The pattern across both colored buckets is consistent: once tiles have color, the strongest second guesses — probe or candidate — keep the confirmed letters in play and spend only the leftover slots on new ones.

The coverage pair is a gray-board tool

There is still a place for the letter-coverage instinct, and it can be measured too. SATIRE’s six letters touch at least one letter in 677 of 682 answers — 99.3 percent — and the five untouched words are precisely the all-gray bucket above. Pairing SATIRE with a complement closes the gap: just 15 of the 1,485 accepted guesses share zero letters with it — UPHOLD, COLUMN, and JOYFUL among them — and SATIRE plus UPHOLD reaches all 682 answers while testing nine of the pool’s ten most frequent letters — every one except N. SATIRE plus COLUMN also reaches 682, leaving only D untested. Those frequency claims come from the top-ten letter counts computed from this pool.

But coverage is a plan made before feedback exists, and the bucket measurements show where it belongs: boards that stayed mostly gray. On the two-yellow board, UPHOLD trailed the recycling probe by a full candidate; on the green-plus-yellows board it trailed by more than two. Memorize a pair if it helps — then let the tiles overrule it.

A working rule for guess two

Condensed from the simulation:

1. Count the candidates first. At 12 or fewer — 78.3 percent of answers after SATIRE — guess from the list. In every measured bucket of eight to twelve candidates, the best candidate’s split came within half a candidate of the best probe’s while adding a real chance of winning on the spot.
2. Above roughly 20, consider one deliberate probe — but build it around the confirmed letters, GOLDEN-style, not from scratch. The payoff over a well-chosen candidate is small; take it when a guess is worth spending, skip it when the guess budget is tight.
3. Wherever you land, choose deliberately. The largest measured losses came not from picking the wrong strategy but from picking a lazy word inside the right one — the 2.88-versus-1.76 gap among candidates dwarfs every probe-versus-filter difference on this page.

Counting candidates by hand is the only step that is tedious, and it is exactly what the worked filter sequence that cuts 682 answers to one automates: enter the tiles, read the list size, and apply the rule. To make the count-then-decide habit automatic, a practice board with no daily cap allows as many second-guess reps as it takes.

Data review: August 23, 2026. Pool version: answers-ef976d79dbc2_guesses-27c282147058. Every bucket size, expected-remaining figure, and coverage count on this page was computed by script from that data version, using the same two-pass feedback rule as the site’s tools; the feedback algorithm version, pool hashes, and correction policy are published for reproduction.