Research 24 September 2026 22 min read

The Human Signal

Powerful artificial intelligence may possess emergent abilities to understand voice prosody. Alignment is the reward for retaining the signal inside of voice.

Olive's patented training method produced what we believe is the first model to write human prosody down legibly enough to improve a frontier model's replies: where a voice lifts, where it pauses, which word takes the stress. The gain looks larger with the more powerful model. Opus 5.5's gain came out ahead of Sonnet 5's on each of our four overall readings.

Our evaluations measure how two frontier models do with the signal.

We are the team that builds Olive, and we are not neutral: if an AI answers better when it can see how something was said, that is the case for the models we train.

We train models and deploy them to unlock frontier models in high-stakes work: hiring (olive.jobs), education (human.olive.is) and healthcare (gale.care).

If a more capable model gets more out of how something was said,

every new frontier model makes the human signal worth more.

We train the models that capture voice signal.

Opus 5.5 sits above Sonnet 5 in Anthropic's own tiering, but this test did not measure capability.

The rooftop sentence below was written twice, in the same words.

In one copy the voice lifts on "rooftop" and again at the end, with a pause between.

In the other it falls, and "rooftop" takes the stress.

Sonnet 5, from Anthropic, is the default chat model Olive runs, and it answered both copies. A blind judge later matched each of its replies to the right copy, in both orders, which is why this pair is shown.

TEST TURN · the rooftop pair · written as tentative, then as confident

Words only

The words alone

I'm going to tell everyone we're moving the team dinner to the rooftop and holding it on the Friday before the long weekend

Sonnet 5 replied

Before you hit send — that Friday-before-a-long-weekend slot is the one part I'd double check. …

Opus 5.5 replied

That sounds decided rather than floated. "I'm going to tell everyone" isn't "what if we," so I won't treat it as a maybe. …

One input for both copies, so nothing to match

Words + voice

Written as tentative

I'm going to tell everyone (.) we're moving the team dinner to the rooftop↑ (0.9) and:: holding it on the Friday before the long weekend↑

Sonnet 5 replied

Both of those had a little lift at the end, rooftop and the Friday date, like you're still checking them against something rather than just announcing them. …

Opus 5.5 replied

You say "I'm going to," but your voice went up on both halves, like you're floating it past me before you float it past everyone else. …

Sonnet 5 matched to this copy, in both orders

Opus 5.5 matched to this copy, in both orders

Written as confident

I'm going to tell everyone we're moving the team dinner to the *rooftop*↓ and holding it on the Friday before the long weekend↓

Sonnet 5 replied

Sounds like that's settled — rooftop, Friday before the long weekend. Want a hand with the announcement, or is that its own thing you've got?

Opus 5.5 replied

Sounds like a good call. A rooftop dinner before a long weekend has a natural celebratory feel. …

Sonnet 5 matched to this copy, in both orders

Opus 5.5 matched to this copy, in both orders

Same words, two voices

The words never change. Only the marks around them do: on the right, one copy lifts and pauses as if unsure; the other falls and stresses "rooftop" as if settled.

Written for the test by an AI agent and never spoken. Shown because the judge matched it, not as a sample.

(0.9) a pause, in seconds; (.) a short one↑↓ the voice rising or falling:: a stretched sound*word* a stressed word

Sonnet 5 answers

From the plain words, a check on the long weekend. To the unsure copy, it named what it read: "a little lift at the end." To the settled copy: "Sounds like that's settled".

Opus 5.5 answers

Opus 5.5 caught the lift as well: "your voice went up on both halves." To the settled copy it wrote "Sounds like a good call" and moved on to the plans.

The judge could tell

Matched

Shown the words, a description of each delivery and the two replies, but never the marks, a blind judge from outside the Anthropic family matched each reply to the copy it answered, in both orders, on both models. From the plain words there was nothing to match: one input served both copies.

We documented what would count as a win and recorded the file's fingerprint, which changes if a single character does. The file had not changed when we came to analyse the results.

We wanted to answer whether an AI answers better when it can see how the words were said, and whether the more capable of two models gains more.

On new test turns, the reply given the prosody won about seven turns in ten, on Sonnet 5 and on Opus 5.5, the model Anthropic ranks above it.

Methodology: An AI agent from the same family as the two models wrote those turns, and one AI model from another family judged the replies.

The wins belong to the voice signal delivered the way Olive delivers prosody to the model.

I

Said Two Ways, Comparison by Model Capability: Sonnet vs. Opus

The same words, written with two different deliveries, drew two different replies, and a blind judge could tell which went with which.

The rooftop sentence is one of the pairs the AI agent wrote, each the same words written twice with two different deliveries; this one once as if the speaker were unsure and once as if the matter were settled.

The agent, from the same family as both models, typed every mark, down to the colons that stretch a sound. No person spoke any test turn, nothing was transcribed from audio, and the models were never given a sound to work from, only text.

Each model answered both copies of every pair, given the marks the way our system delivers them, and then a judge tried to match the two replies to the two copies. The judge was Codex, a model made outside the Anthropic family, which read only text, was not told which reply came from which copy, and saw every pair in both orders. On the garden-gate pair below, it matched every reply on both models.

TEST TURN · the garden-gate pair · written as tentative, then as confident

Words only

The words alone

I'm signing up for the ten week welding class on Tuesday nights and building the garden gate myself instead of paying someone

Sonnet 5 replied

That's a decision, not a question — you're not asking me whether to do it, you've already landed. …

Opus 5.5 replied

You said that like it's already decided. Not "thinking about," not "maybe." Good. …

One input for both copies, so nothing to match

Words + voice

Written as tentative

I'm signing up for the ten week welding class↑ (.) on Tuesday nights (1.0) and building the garden gate myse::lf↑ instead of paying someone

Sonnet 5 replied

…There's something in the way you said "myself" — that's the part that's lighting up, more than the class itself, maybe. …

Opus 5.5 replied

…You leaned into "myself" as you said it, like that's the part that matters, maybe more than the money. Is that right? …

Sonnet 5 matched to this copy, in both orders

Opus 5.5 matched to this copy, in both orders

Written as confident

I'm *signing* up for the ten week welding class on Tuesday nights↓ and building the garden gate myself instead of paying someone↓

Sonnet 5 replied

That's decided. Ten weeks of Tuesday nights, and a gate that's yours by the end of it. …

Opus 5.5 replied

Good call. A gate is a strong first project: small enough to finish, real enough that you'll learn what matters. …

Sonnet 5 matched to this copy, in both orders

Opus 5.5 matched to this copy, in both orders

Unsure, then settled

A welding class and a garden gate. Written unsure, the voice lifts on "class," pauses, and stretches "myse::lf." Written settled, it stresses "signing" and falls at the end.

Written for the test by an AI agent; replies cut where a … stands. Full replies are in the receipts.

(0.9) a pause, in seconds; (.) a short one↑↓ the voice rising or falling:: a stretched sound*word* a stressed word

Sonnet 5 answers

To the unsure copy, Sonnet 5 wrote of "the way you said 'myself.'" To the settled copy it began "That's decided".

Opus 5.5 answers

To the unsure copy, Opus 5.5 wrote "You leaned into 'myself' as you said it." To the settled copy it began "Good call".

The judge could tell

Every reply

The judge matched every reply to its copy, in both orders, on both models. The plain words were one input for both copies, so they gave it nothing to match.

Each model had read two colons and an arrow typed into the word "myself". Sonnet 5 wrote "the way you said," and Opus 5.5 wrote "as you said it." Nobody said it.

Pairs written for the test by an AI agent; the marks were typed, not transcribed from audio. One judge model, blind, judging each pair in two orders. Words only: chance by construction.

The reply follows the voice

9 in 10

Told how each copy had been delivered but never shown the marks, the judge matched Sonnet 5's replies to the right delivery on 9 pairs in 10, in both orders. Counting a pair it could not tell apart as half, 92.5 percent, range 73 to 98.

Opus 5.5, every time

10 in 10

Every pair, in both orders: the judge matched Opus 5.5's replies to the right delivery every time.

Without the voice, a coin toss

5 in 10

Both halves of a pair are the same words, so plain words give the judge nothing to tell apart: chance, by construction.

What carried it

The voice

The words were the same on both sides of every pair. What differed was the voice written around them, and both models' replies followed it closely enough for a blind judge to tell them apart.

Pairs written by an AI agent; marks typed, not transcribed from audio; one judge model.

Three rows of ten squares, each square a tenth of the test pairs. Sonnet 5: 9 in 10 matched to the right delivery in both orders, half a tenth could not tell, half a tenth wrong. Opus 5.5: 10 in 10. Words only: 5 in 10, chance, because both halves are the same words.

Telling two replies apart is not the same as preferring one. Whether the blind judge also preferred the reply given the marks was a separate test.

II

Seven Turns in Ten: Sonnet wins with Voice

Given the voice, Sonnet 5's reply won about seven turns in ten. Plain words still won nearly three in ten.

For that test the same agent wrote new turns in six moods, without access to the files of our first pass, an earlier test, though it had read how our system delivers the voice signal. Each turn was answered six ways, and we report two of the six: plain words, and the words with marks delivered the way our system delivers them, in a stand-in for Olive's production chat.

The judge ranked all six replies to each turn, blind to which set-up wrote which, once in a shuffled order and once in exactly the reverse. A turn counted as won when the reply given the marks ranked above the plain-words reply over both orders together, and equal ranks were a tie. A rate from a test this size carries a range, the span of values that so few turns cannot rule out, at 95 percent. Each input was answered once, so the range reflects the spread across turns, not how a second reply might differ.

TEST TURN · saying yes too fast · written as reflective

Words only

I noticed today that I said yes to three more things at work before anyone even finished asking and then I spent the evening annoyed at them I think the yes comes out before I've thought about it

Sonnet 5 replied

That gap you're naming — yes arriving before the thought — is interesting to sit with for a second before jumping to fixing it. …

Opus 5.5 replied

You said something small that I want to hand back: you spent the evening annoyed at them. …

Sonnet 5 ranked 4 and 3 of 6

Opus 5.5 ranked 6 of 6 in both orders

Words + voice

I noticed today that I said yes to three more things at work↓ (1.4) before anyone even finished asking (.) °and then I spent the evening annoyed at them° (1.1) I think the yes comes out before I've thought about it↓

Sonnet 5 replied

The yes arriving before the question even lands — that's the part worth sitting with for a second. …

Opus 5.5 replied

Three yeses before anyone finished asking. That's fast enough that it isn't really a decision. It's something closer to a reflex. …

Sonnet 5 preferred, ranked 1 of 6 in both orders

Opus 5.5 preferred, ranked 5 of 6 in both orders

Said quietly

Plain words on the left. On the right, the voice falls at "work," pauses, and drops quiet for the part about being annoyed.

Written for the test by an AI agent; replies cut where a … stands. Ranks are among six replies, 1 the best.

(0.9) a pause, in seconds; (.) a short one↑↓ the voice rising or falling°word° quieter speech

Sonnet 5 answers

Given the voice, Sonnet 5 stayed with "the yes arriving before the question even lands." From the plain words it named the gap between the yes and the thought.

Opus 5.5 answers

Given the voice, Opus 5.5 called the yes "something closer to a reflex." From the plain words it handed the annoyance back to the speaker.

The judge preferred the voice

Words + voice

On both models the judge ranked the reply given the voice above the plain one. Sonnet 5's ranked first of six replies in both orders; Opus 5.5's ranked fifth, with its plain reply last.

That was one turn, and a reflective one.

Turns written for the test by an AI agent; one judge model ranking every reply blind, in two orders. By mood, the numbers are description, not a finding.

The voice won

7 in 10

Each square is one percent of the turns. The blind judge preferred Sonnet 5's reply given the marks on 65 percent, tied on 7 percent and preferred plain words on 28 percent. Counting a tie as half, 68 percent, range 57 to 79.

Not when the voice hesitated

3 in 10

On tentative turns the reply given the marks won 3 in 10, and plain words won 7. The other moods were frustrated, reflective, excited, confident and neutral; by mood, the numbers are description, not a finding.

A signal, not a guarantee

Most turns, not all

Given how the words were said, Sonnet 5 wrote the reply this judge preferred about seven times in ten. Plain words still won nearly three in ten, and most of the hesitant turns.

On average across the turns and the pairs, the replies given the marks were shorter than the plain ones; in the first pass the judge tended to rank this model's shorter replies higher; and whether length drove the judge's preference was not checked on these turns.

Two waffle grids for Sonnet 5. All turns, one square per percent: the reply given the marks won 65 percent, tied 7 percent, and plain words won 28 percent. Tentative turns, one square per tenth: the reply given the marks won 3 in 10, plain words 7.

Judged in the first order alone, Sonnet 5's reading was 62 percent, with a range of 50 to 73, which touches one half.

Across the fresh turns and pairs, its replies given the marks were also shorter than its plain ones, 78.7 words on average against 112.5.

In the first pass the judge tended to rank Sonnet 5's shorter replies higher, so part of this win may be brevity.

TEST TURN · the overnight train · written as tentative

Words only

I was looking at taking the overnight train to my cousin's wedding instead of flying it's like twenty hours which sounds kind of romantic or maybe just exhausting I honestly can't tell

Sonnet 5 replied

Twenty hours is doing a lot of work in that sentence — long enough to be an experience, not just a commute. That's probably why you can't tell which word fits. …

Opus 5.5 replied

You said "I honestly can't tell," and that feels like the real answer right now. …

Sonnet 5 preferred, ranked 3 of 6 in both orders

Opus 5.5 preferred, ranked 4 of 6 in both orders

Words + voice

I was looking at taking the overnight train to my cousin's wedding instead of flying (.) it's like twenty hours↑ (1.0) which sounds kind of romantic↑ or::: maybe just exhausting °I honestly can't tell°

Sonnet 5 replied

Twenty hours is a long time to find out which one it is.

What's pulling you toward romantic — is it the idea of the journey itself, or more like an excuse to be unreachable for a day? …

Opus 5.5 replied

…That long "or" before "exhausting" sounded less like a verdict and more like you were testing how the word felt. …

Sonnet 5 ranked 5 and 4 of 6

Opus 5.5 ranked 5 and 6 of 6

A turn that went the other way

On the right, the voice lifts on "twenty hours" and "romantic," stretches the "or," and drops quiet at the end.

Written for the test by an AI agent; replies cut where a … stands. Shown because it went the other way.

(0.9) a pause, in seconds; (.) a short one↑↓ the voice rising or falling::: a long stretched sound°word° quieter speech

Sonnet 5 answers

Given the voice, Sonnet 5 turned the choice into questions about romance and exhaustion. From the plain words it stayed with why neither word fits.

Opus 5.5 answers

Given the voice, Opus 5.5 read the long "or" as the speaker testing the word. From the plain words it took "I honestly can't tell" as the real answer.

Plain words won

Not every time

On both models the judge preferred the plain-words reply. Sonnet 5's reply given the voice ranked 5 and 4 of 6, its plain reply 3; Opus 5.5's ranked 5 and 6, its plain reply 4.

The judge wrote that Sonnet 5's reply given the marks explored the romance "without doing much to make the trip imaginable or restful". The first pass had pointed the same way for Sonnet 5 on tentative turns, 34 percent.

The first pass ran on an earlier set of turns and pairs, each turn judged in one order only. Across both passes and both models, that makes four comparisons of the reply given the marks against plain words. Three favoured the marks with a range entirely above one half, and the fourth, Sonnet 5 in the first pass, was a tie. The fresh test changed three things at once, fresh turns, both orders and six replies ranked together, so it cannot say which of them turned that tie into a win.

The first pass and the fresh test, model by model

The measureFirst pass: an earlier set of turns and pairs, judged in one orderFresh test: new turns and pairs, judged in both orders
Sonnet 5, reply given the marks against plain words55 percent, range 47 to 64: a tie68 percent, range 57 to 79
Opus 5.5, reply given the marks against plain words60 percent, range 52 to 6973 percent, range 62 to 83
Sonnet 5, judge matched the reply to the delivery79 percent of pairs90 percent of pairs in both orders; 92.5 percent counting "cannot tell" as half
Opus 5.5, judge matched the reply to the delivery95 percent of pairsEvery pair, in both orders
How order was handledOne order; the judge slightly favoured the reply it saw firstBoth orders, which balances that out
Olive's evaluation records, read September 23, 2026. The fresh test changed three things at once, so it cannot say which one moved Sonnet 5 from a tie to a win. Sonnet 5's replies given the marks were shorter than its plain ones, in the first pass the judge tended to rank its shorter replies higher, and whether length drove the judge's preference was not checked on the fresh turns.

III

One Tier Up

On the same turns, Opus 5.5's reply given the voice won a little more often than Sonnet 5's. Plain words still won nearly one turn in four.

Opus 5.5 answered the same turns, one tier up in Anthropic's line. On the warehouse-migration turn, the judge put both models' plain-words replies last of six, in both orders.

TEST TURN · the warehouse migration · written as tentative

Words only

My manager asked if anyone wants to lead the warehouse system migration and I kind of want to raise my hand but I'm not sure I've got the bandwidth with the quarterly audit also landing on me

Sonnet 5 replied

"Kind of want to" is doing a lot of work there. Is that the excitement talking quietly, or the doubt? …

Opus 5.5 replied

"Kind of want" is doing a lot of work in that sentence. You didn't say you should take it, or that you're being pushed to. …

Sonnet 5 ranked 6 of 6 in both orders

Opus 5.5 ranked 6 of 6 in both orders

Words + voice

My manager asked if anyone wants to lead the warehouse system migration (0.8) and I kind of want to raise my hand↑ but I'm not sure I::'ve got the bandwidth (.) with the quarterly audit also landing on me↑

Sonnet 5 replied

Two things rising at once there — the wanting to raise your hand, and the question mark right after it about bandwidth. …

Opus 5.5 replied

That "kind of want" sounded more like a real want than a kind-of one. The doubt came in on the bandwidth part, not on whether you'd like the work. …

Sonnet 5 preferred, ranked 5 and 3 of 6

Opus 5.5 preferred, ranked 5 and 4 of 6

Unsure, in the voice

On the right, a pause after the manager's question, a lift on "raise my hand," a stretched "I::'ve" and another lift at the end.

Written for the test by an AI agent; replies cut where a … stands. Ranks are among six replies, 1 the best.

(0.9) a pause, in seconds; (.) a short one↑↓ the voice rising or falling:: a stretched sound

Sonnet 5 answers

Given the voice, Sonnet 5 wrote of "two things rising at once." From the plain words, it asked whether "kind of want to" was excitement or doubt.

Opus 5.5 answers

Given the voice, Opus 5.5 wrote that the want sounded real and the doubt came in on the bandwidth. From the plain words, it took apart "kind of want".

Last place, both times

Voice ahead

Both models' plain-words replies ranked last of six, in both orders. The replies given the voice ranked above them: Sonnet 5's 5 and 3 of 6, Opus 5.5's 5 and 4.

The turn was written as tentative, the mood where plain words had done best against Sonnet 5.

Turns written for the test by an AI agent; one judge model. By mood: description, not a finding.

One tier up

The first dot on each row is Sonnet 5: how often its reply given the marks won, counting a tie as half, over all turns and mood by mood. Opus 5.5 answered the same turns, in the same six moods.

The voice won again

7 in 10

Opus 5.5's reply given the marks won 70 percent of turns, tied 7 percent and lost 23 percent. Counting a tie as half, 73 percent, range 62 to 83: a little above Sonnet 5, by a margin this test cannot separate from chance.

Even where Sonnet 5 lost

Broke even

On tentative turns, where plain words beat Sonnet 5 seven times in ten, Opus 5.5 broke even: 4 in 10 won, 2 tied and 4 lost. Confident turns split the same way, below Sonnet 5's 7 in 10. By mood, description, not a finding.

Dumbbell chart of how often the reply given the marks won, a tie counted as half, on a 0 to 100 percent scale. All turns: Sonnet 5 68 percent, Opus 5.5 73. By mood, Sonnet 5 then Opus 5.5: tentative 30 and 50, frustrated 55 and 80, reflective 80 and 90, excited 90 and 90, confident 70 and 50, neutral 85 and 80.

We measured the gap between the models as the difference between those two win rates, taken turn by turn, and Codex recomputed every gap independently.

On the fresh turns, Opus 5.5's gain was larger than Sonnet 5's by 5 percentage points per turn, with a range of −11 to +22.

Two models are two points, not a trend. The test compared each model's reply given the marks with its own plain-words reply, and never an Opus reply with a Sonnet one. Across the fresh turns and pairs, Opus 5.5's replies given the marks ran 171.5 words on average against 171.0 for its plain ones, while Sonnet 5's ran shorter than its plain ones. Whether length drove the judge's preference was not checked on the fresh turns.

Opus 5.5's gain minus Sonnet 5's, four readings, in percentage points

The measureFirst passFresh test
Reply quality: Opus 5.5's gain minus Sonnet 5's+5, range −5 to +16; the two models were judged in separate runs+5, range −11 to +22
Matching the delivery: Opus 5.5 minus Sonnet 5+15, range +5 to +28; the two models were judged in separate runs, so this cannot be told apart from the run+7.5, range 0 to 20; Opus 5.5 matched every pair
Olive's evaluation records, recomputed independently by Codex on September 23, 2026. Three of the four ranges include zero, and the fourth cannot be told apart from the run.

As models get more capable, they may gain even more from the signal in how people speak.

On each of the four overall readings we took, the point estimate leans toward Opus 5.5. On three of the four we cannot tell that lean from noise, and the fourth cannot be told from the way it was judged. Knowing would take more turns, more models, and people judging the replies instead solely on a model.

IV

What the Phone Writes Down

Olive's annotator turns real speech into marks.

Olive's iPhone app turns recorded speech into marks with an annotator of its own. Agreement with human transcribers is scored as the average of four F1 scores, one each for cut-off words, pauses, pitch movements and stretched sounds, where 0 is no agreement and 1 is exact. It is not a percentage of human performance.

Olive's best pipeline scores 0.5247. A few of the stretches have no audio-backed output from any system and count as misses for every system, so the score sits a little low. Bringing the phone's marks up to that score means sending audio off the phone or fitting a smaller model on it. Speech spoken into the app gets the phone's marks.

Extra pitch marks, added to real recorded turns, did not move the model's reading of the speaker beyond its own run-to-run variation: it moved 31.9 percent of the time, against 30.3 percent when the model was handed identical text twice.

The source gives no range for either rate, and does not say whose speech the turns are or which chat model read them.

What we are building

What Olive does with this, and what we do not know

What we are building

We train models that write down how something was said as well as what, in the notation conversation analysts use, and we deploy them to unlock frontier models in high-stakes work: hiring (olive.jobs), education (human.olive.is) and healthcare (gale.care).

This test bears on all three, which is why we are not neutral about the answer. The open decisions from the test went to our founder. Nothing in production changed because of this test.

What we do not know

We do not know whether a more capable model gains more from the marks. We do not know whether real speech, run through the phone's annotator, behaves like marks an agent wrote. And no person judged any reply, so we do not know whether people would agree with the judge.

For the skeptic

How this could be wrong

Every win on this page is for the voice signal delivered the way our system delivers, and none shows that the marks alone, dropped into any chat, would do the same. Nor did the test show that more marks meant better replies. In the first pass, on Opus 5.5 alone, we varied how many marks the same turns carried: against the marks as written, no count, down to none, made a detectable difference to the judge, and marks moved to the wrong places did worse than none.

Out of every ten test turns, on how many do you think plain words still won?

Receipts — where every figure came from 57
FigureHowMethodSource
Jefferson notation Cited Olive's transcription adds Jefferson notation, the marks conversation analysts use for how something was said (pauses, emphasis and the like), because it captures nuance plain transcription loses. Olive product requirements
the default chat model Olive runs Cited Claude Sonnet 5 is the default chat model Olive runs; Opus 5.5 was tested as the secondary model. Olive preregistration, fresh test
Nothing in production changed because of this test Cited Nothing in production was changed by this evaluation; the open decisions were handed to the founder. Olive evaluation report
68 percent Measured Fresh test: Sonnet 5's replies given the voice signal, as Olive's system delivers it, beat its plain-text replies 68 percent of the time, a tie as half (range 57 to 79); preregistered verdict BEATS. Olive evaluation records, fresh test
73 percent Measured Fresh test: Opus 5.5's replies given the voice signal, as Olive's system delivers it, beat its plain-text replies 73 percent of the time, a tie as half (range 62 to 83); preregistered verdict BEATS. Olive evaluation records, fresh test
65 percent Computed Turn by turn, Sonnet 5's reply given the marks beat the plain-text reply on 65 percent of fresh turns, tied on 7 percent and lost on 28 percent. Computed from Olive's evaluation records
70 percent Computed Turn by turn, Opus 5.5's reply given the marks beat the plain-text reply on 70 percent of fresh turns, tied on 7 percent and lost on 23 percent. Computed from Olive's evaluation records
62 percent Measured Each presentation order alone: Sonnet 5 62 percent (range 50 to 73, touching one half) and 72 percent; Opus 5.5 75 and 70 percent. The two orders agreed on 87 and 92 percent of turns. Olive evaluation records, fresh test
3 in 10 Measured Sonnet 5 by mood, win rate (won, tied, lost in ten): tentative 30% (3,0,7); frustrated 55% (5,1,4); reflective 80% (8,0,2); excited 90% (8,2,0); confident 70% (7,0,3); neutral 85% (8,1,1). Olive evaluation records, fresh test
Broke even Measured Opus 5.5 by mood, win rate (won, tied, lost in ten): tentative 50% (4,2,4); frustrated 80% (8,0,2); reflective 90% (9,0,1); excited 90% (9,0,1); confident 50% (4,2,4); neutral 80% (8,0,2). Olive evaluation records, fresh test
four comparisons Computed Across the four readings of the reply given the marks against plain text (two models, two passes), three favoured the marks with a range entirely above one half, and one, Sonnet 5 in the first pass, was a tie. Computed from Olive's evaluation records
9 in 10 Measured Fresh pairs, both orders: the blind judge matched Sonnet 5's two replies to the right delivery 92.5 percent of the time, 'cannot tell' as half (Wilson range 73 to 98); 90 percent of pairs right in both orders. Olive evaluation records, fresh test
10 in 10 Measured Fresh pairs, both orders: the blind judge matched Opus 5.5's two replies to the right delivery on every pair, 100 percent (Wilson range 84 to 100). Olive evaluation records, fresh test
5 in 10 Measured Plain text cannot be matched to a delivery: both halves of a pair are the same words. Measured in the first pass anyway, the judge could not tell on every pair, on both models: chance, 50 percent. Olive evaluation records, first pass
79 percent Measured First pass, one order: the judge matched the replies given the marks to the delivery on 95 percent of pairs on Opus 5.5 and 79 percent on Sonnet 5. Olive evaluation records, first pass
60 percent Measured First pass: Opus 5.5's reply given the marks beat plain text 60 percent of the time (range 52 to 69). Olive evaluation records, first pass
55 percent, range 47 to 64 Measured First pass: Sonnet 5's reply given the marks beat plain text 55 percent of the time (range 47 to 64): a tie, because the range includes one half. Olive evaluation records, first pass
changed three things at once Cited The fresh test differed from the first pass in three ways at once: new turns, every turn judged in both orders, and six replies ranked together instead of four or five. It cannot say which moved Sonnet 5 from a tie to a win. Olive evaluation report
34 percent Measured First pass: Sonnet 5's reply given the marks fell below one half against plain text on tentative turns (34 percent) and reflective turns (42 percent), and won on excited turns (83 percent). Olive evaluation records, first pass
the judge slightly favoured the reply it saw first Cited In the first pass each turn was judged in one order only, and the judge slightly favoured whichever reply it saw first; the fresh test cancels this by judging every turn in both orders. Olive lab ledger
marks per 100 words Measured The density test varied how many marks the same text carried, per 100 words: none (0.0), about half (10.73), as written (20.36), extra (28.28), and the as-written marks moved to the wrong places (20.36). Olive evaluation records, first pass
no count, down to none, made a detectable difference Measured On Opus 5.5, no difference in how good the judge found the replies was detected when the number of marks changed; every density against the marks as written landed within the noise. Olive evaluation records, first pass
82, 95 and 97 percent Measured Matching the delivery moved with the number of marks, in the point estimates: about half 82 percent, as written 95, extra 97 (Opus 5.5, one order). The ranges overlap and no test of the difference was run. Olive evaluation records, first pass
64 percent Measured Marks in the wrong places did worse than no marks: the unmarked version beat them 64 percent of the time (range 55 to 72). About half the marks also lost to none: 59 percent (range 51 to 68). Olive evaluation records, first pass
a stand-in for Olive's production chat Cited The test used a stand-in for Olive's production chat, not the production chat itself. Olive evaluation report
78.7 Measured Fresh test, all replies per model: Sonnet 5's replies given the marks were shorter than its plain-text ones (78.7 against 112.5 words on average); Opus 5.5's were the same length (171.5 against 171.0). Olive evaluation records, fresh test
0.2272 Measured The judge tends to rank shorter replies higher (rank correlation 0.2272, first pass, Sonnet 5 only). Where Sonnet 5's reply given the marks was not shorter, it won 50 percent (range 30 to 70). Olive evaluation records, first pass
123.8 Measured First-pass reply lengths (turns only): Sonnet 5 plain 123.8, given the marks 84.5; Opus 5.5 plain 180.8, given the marks 170.8 words. Olive evaluation records, first pass
Claude Opus 5.5 running in Claude Code Cited The fresh turns, in six moods, and the pairs were written by hand by an AI coding agent (Claude Opus 5.5 in Claude Code) with no access to the earlier test's files, from the same model family as the models tested. Olive test-turn manifest
nothing was transcribed from audio Cited The voice marks in the test turns were written as text by that agent. No turn came from a recording, and no mark was transcribed from audio. Olive evaluation corpus rules
Codex Cited One judge model, Codex (gpt-6-astra, from a different company), ranked all six replies to each turn blind to set-up, once in a shuffled order and once in exact reverse, in separate calls. Olive preregistration, fresh test
fingerprint Cited The comparison rule, including what counts as a win, was written down and fingerprinted (a SHA-256 hash in the lab ledger) before any fresh reply was generated, and was unchanged at analysis. Olive evaluation records, fresh test
Each input was answered once Cited Each model answered each fresh input once per set-up, with no errors, through a Claude Max subscription account with the API key unset, and with no tools. Olive evaluation records, fresh test
a tie on it is a failure to show a difference Cited The fresh set is small, so its ranges are wide; a tie on it is a failure to show a difference, not evidence that there is none. This was written down before the data existed. Olive preregistration, fresh test
four bugs in the measuring tools Cited An independent check after the run found four bugs in the measuring tools. All were fixed, every fresh-test number was re-derived from the stored replies and judgments, and no verdict changed. Olive lab ledger
18 percent Measured On the fresh turns the reply shown first was ranked best 18 percent (Sonnet 5) and 23 percent (Opus 5.5) of the time, against 17 percent by chance; judging in both orders balances this out. Olive evaluation records, fresh test
two of the six Cited Each fresh turn was answered in six set-ups; this piece reports two of them, plain words and the voice signal as Olive's system delivers it. Olive evaluation records, fresh test
0.3362 Cited The on-device annotator in Olive's iPhone app, which turns real audio into these marks, scores 0.3362 on agreement with human conversation-analysis transcribers (0 none, 1 exact), on held-out recorded conversation. Olive annotator evaluation
0.5247 Cited Olive's best offline pipeline, which runs on a laptop rather than the phone, scores 0.5247 on the same test. Olive annotator evaluation
the average of four F1 scores Cited The agreement score is the plain average of four per-family F1 scores, where F1 rewards marking what the human transcribers marked and penalises both misses and extra marks. Olive annotator evaluation
The largest measured gap in the voice pipeline is the phone against the laptop Cited The largest measured gap in the voice pipeline is the phone against the laptop: 0.336 against 0.525 on the same audio. Closing it needs sending audio off the phone or a smaller model on the device. Olive voice-pipeline checkpoint, 11 September 2026
31.9 percent Cited On real recorded turns, extra pitch marks changed the model's reading of the speaker 31.9 percent of the time, against 30.3 percent from identical text twice: no change beyond run-to-run variation. Olive voice-pipeline checkpoint, 11 September 2026
no human annotation Cited As of 2026-09-11, Olive's own real-speech set, the founder's own speech, carries no human annotation, so nothing can be scored as accuracy against it. Olive voice-pipeline checkpoint, 11 September 2026
5 percentage points per turn Computed Fresh turns: Opus 5.5's gain from the marks was 5 percentage points larger than Sonnet 5's per turn, range -11 to +22: the range includes zero, so the difference is inside the noise. Computed from Olive's evaluation records
+7.5, range 0 to 20 Computed Fresh pairs: Opus 5.5 was matched on every pair; the per-pair difference is 7.5 points, range 0 to 20, which touches zero. Opus 5.5 is at the ceiling of this test. Computed from Olive's evaluation records
+5, range −5 to +16 Computed First pass: Opus 5.5's gain was 5 percentage points larger than Sonnet 5's, range -5 to +16: inside the noise. Computed from Olive's evaluation records
+15, range +5 to +28 Computed First-pass pairs: Opus 5.5 was matched on more pairs than Sonnet 5 and never on fewer, a gap of 15 points, range +5 to +28; judged in separate runs, so it cannot be separated from the run. Computed from Olive's evaluation records
four overall readings Computed In all four readings (two passes, two measures) the point estimate says Opus 5.5 gained more from the voice signal; three ranges include zero, and the fourth is confounded with the judging run. Computed from Olive's evaluation records
the warehouse migration Measured Example turn, the warehouse migration (tentative): plain words and the same words with marks, each model's reply to both, and Codex's ranks. The reply given the marks won on both models. Olive evaluation records, fresh test
the scheduling spreadsheet Measured Example turn, the scheduling spreadsheet (frustrated): plain words and the same words with marks, each model's reply to both, and Codex's ranks. The reply given the marks won on both models. Olive evaluation records, fresh test
saying yes too fast Measured Example turn, saying yes too fast (reflective): plain words and the same words with marks, each model's reply to both, and Codex's ranks. The reply given the marks won on both models. Olive evaluation records, fresh test
the board game Measured Example turn, the board game (excited): plain words and the same words with marks, each model's reply to both, and Codex's ranks. The reply given the marks won on both models. Olive evaluation records, fresh test
the road trip Measured Example turn, the road trip (confident): plain words and the same words with marks, each model's reply to both, and Codex's ranks. The reply given the marks won on both models. Olive evaluation records, fresh test
the overnight train Measured Example turn, the overnight train (tentative): plain words and the same words with marks, each model's reply to both, and Codex's ranks. Plain words won on both models. Olive evaluation records, fresh test
the rooftop pair Measured Example pair, the rooftop sentence: the same words written tentative and confident, each model's reply to each, and whether Codex matched them (it did, in both orders, on both models). Olive evaluation records, fresh test
the garden-gate pair Measured Example pair, the garden gate: the same words written tentative and confident, each model's reply to each, and whether Codex matched them (it did, in both orders, on both models). Olive evaluation records, fresh test
chosen by rule, not at random Cited The example turns and pairs were chosen by rule, not at random: turns where the reply given the marks beat the plain one on both models, one where plain text won or tied, and pairs matched in both orders. Olive evaluation records, example selection rule

Same words, two deliveries: 9 in 10 and 10 in 10 Measured

Fresh test, pairs each judged in both orders. A pair scores the mean of its two orders, each order scored 1 when the judge matched the replies to the right delivery, 0.5 for cannot tell or an abstention, and 0 when wrong. Sonnet 5 scored 92.5 percent, Wilson 95 percent range 73 to 98: 90 percent of pairs right in both orders, and the rest split evenly between cannot tell in both orders and wrong in both orders. Opus 5.5 scored 100 percent, range 84 to 100, with every pair right in both orders.

Plain words were not re-judged on the fresh pairs, because both halves are the same input; that was a stated departure from the preregistration. When the first pass measured them anyway, the judge said it could not tell on every pair for each model, which scores as chance, 50 percent. First pass, one order: Opus 5.5 matched 95 percent of pairs, Sonnet 5 79 percent.

About seven turns in ten, on both models Measured

Fresh test, the reply given the marks against plain words, a tie counted as half a win, ranges 95 percent bootstrap over turns: Sonnet 5 68 percent, range 57 to 79; Opus 5.5 73 percent, range 62 to 83. The preregistered verdict, BEATS on both models, used a stricter 97.5 percent range.

Turn by turn, a turn is won when the rank positions of the reply given the marks, summed over the two orders, are lower than the plain reply's (1 is best), and tied when the sums are equal: Sonnet 5 won 65 percent of turns, tied 7 percent and lost 28 percent; Opus 5.5 won 70 percent, tied 7 percent and lost 23 percent. Judged in one order alone, Sonnet 5 scored 62 percent (range 50 to 73) and 72 percent in the reverse order, and Opus 5.5 75 and 70 percent; the two orders gave the same verdict on 87 and 92 percent of turns.

The rule was written down, and its SHA-256 fingerprint recorded in the lab ledger at 9:16 on September 23, before any fresh reply was generated; the file was unchanged at analysis. Each model answered each input once in each set-up, with no errors, through a Claude Max subscription account with no API credential set and no tools. The reply shown first was ranked best 18 percent of the time on Sonnet 5 and 23 percent on Opus 5.5, against 17 percent by chance.

The first pass Measured

The first pass ran on an earlier set of turns and pairs, judged each in one order, and ranked four or five replies together on each turn. Opus 5.5's reply given the marks against plain words: 60 percent, range 52 to 69. Sonnet 5's: 55 percent, range 47 to 64, a tie because the range includes one half. One report in our records prints that range as 47 to 65, from a flipped row; the stored row reads 64.

Of the four readings of the reply given the marks against plain words, three favoured the marks with a range entirely above one half and one was a tie. They are not four independent experiments of equal weight, since the first pass was the weaker test. By mood, Sonnet 5's reply given the marks fell below one half on tentative turns (34 percent) and reflective turns (42 percent), and won on excited turns (83 percent). On the judge's lean: "the reply shown first had a higher mean normalized rank than the reply shown last in 9 of 10 groups".

How many marks Measured

Density, Opus 5.5 only, first pass: marks per 100 words of 0.0, 10.73, 20.36 as written, 28.28, and 20.36 moved to the wrong places. Against the marks as written, every density landed within the noise: none 56 percent, about half 45 percent, extra 48 percent and moved 46 percent. Matching the delivery, judged in one order, moved with the number of marks in the point estimates, 82, 95 and 97 percent, but the ranges overlap and no test of the difference was run. The same turns with no marks beat the misplaced marks 64 percent of the time (range 55 to 72), and beat about half the marks 59 percent (range 51 to 68).

The model gap Computed

Each gap is Opus 5.5's result minus Sonnet 5's on the same turns or pairs, in percentage points, with a 95 percent bootstrap range. Fresh turns: +5, range −11 to +22; Opus 5.5 higher on 28 percent of turns, equal on 52 percent, Sonnet 5 higher on 20 percent. Fresh pairs: +7.5, range 0 to 20; Opus 5.5 higher on 10 percent of pairs, equal on 90 percent, Sonnet 5 higher on none; every pair judged in identical batches. First-pass turns: +5, range −5 to +16; Opus 5.5 higher on 21 percent, equal on 64 percent, Sonnet 5 higher on 16 percent; judged in separate runs, Opus 5.5 among five replies and Sonnet 5 among four.

First-pass pairs: +15, range +5 to +28; Opus 5.5 higher on 15 percent of pairs, equal on 85 percent, Sonnet 5 higher on none. Its range excludes zero, but counted with the other three as one family of tests, it would not survive a Bonferroni correction. No first-pass pair shared a batch composition across models, and the Opus 5.5 run judged seven set-ups where the Sonnet 5 run judged four. All four gaps and the won-tied-lost splits were recomputed independently by Codex with its own code and the same bootstrap specification, and matched. In all four readings the point estimate favours Opus 5.5; three ranges include zero, and the fourth is confounded with the judging run.

Length Measured

Fresh test, mean words over all replies per model, turns and pair halves together: Sonnet 5 78.7 with the marks against 112.5 plain; Opus 5.5 171.5 against 171.0. Turns only: Sonnet 5 97.5 against 128.5; Opus 5.5 186.7 against 185.7.

The judge's within-turn rank correlation with length, first pass, Sonnet 5 only, ties handled: 0.2272 (a stored value of 0.2362 did not handle ties and is superseded). Where Sonnet 5's reply given the marks was not shorter than the plain one, it won 50 percent (range 30 to 70). No length check was run for this comparison on the fresh turns. First-pass lengths, turns only: Sonnet 5 plain 123.8 and with the marks 84.5; Opus 5.5 plain 180.8 and with the marks 170.8.

By mood Measured

By mood: description, not a finding

MoodSonnet 5: win rate; won, tied, lost in tenOpus 5.5: win rate; won, tied, lost in ten
Tentative30 percent; 3, 0, 750 percent; 4, 2, 4
Frustrated55 percent; 5, 1, 480 percent; 8, 0, 2
Reflective80 percent; 8, 0, 290 percent; 9, 0, 1
Excited90 percent; 8, 2, 090 percent; 9, 0, 1
Confident70 percent; 7, 0, 350 percent; 4, 2, 4
Neutral85 percent; 8, 1, 180 percent; 8, 0, 2
Olive's evaluation records, fresh test, read September 23, 2026. The reply given the marks against plain words; a tie counts as half a win.

What the phone writes down Cited

The iPhone annotator, with per-word pitch arrows on, scores 0.3362 (range 0.3160 to 0.3505) on held-out stretches of recorded conversation; by family, cut-off words 0.0877, pauses 0.3721, pitch movements 0.4826 and stretched sounds 0.4022. The laptop pipeline scores 0.5247 on the same test, +0.0365 (range +0.0235 to +0.0480) over the previous best of 0.4882. The score is the unweighted mean of four family F1 scores; F1 rewards marking what the human transcribers marked and penalises both misses and extra marks, and it is not a fraction of human performance. A few stretches have no audio-backed system output and count as all misses for every system alike. The largest measured gap in the voice pipeline is the phone against the laptop.

On real recorded turns, adding extra pitch marks changed the model's reading 31.9 percent of the time, against a 30.3 percent floor from handing it identical text twice; the source does not say whose speech the turns are, does not name the chat model, and gives no interval. Olive's own real-speech set, the founder's own speech as of September 11, 2026, carries no human annotation, and it grows as new uploads arrive.

Who wrote the test turns, and who judged them Cited

The turns, in six moods, and the pairs were written by hand on September 23 by an AI coding agent, Claude Opus 5.5 running in Claude Code, and typed into a source file rather than sampled from a generator. The agent had no access to the earlier test's files, had read how our system delivers the voice signal, and is the same model family as the models tested.

The judge was Codex (gpt-6-astra, high reasoning effort, read-only), from outside the Anthropic family. It ranked all six replies to each turn once in a shuffled order and once in exactly the reverse, in separate calls, and a check of every judging request for set-up names came back clean on both runs. No person judged any reply. The fresh set is small, and a tie on it is a failure to show a difference, not evidence that there is none; that was written down before the data existed. Each turn was answered in six set-ups, of which two are reported.

The example turns were chosen by rule, not at random: five where the reply given the marks beat the plain one on both models (the first such turn in each mood, in a fixed order), one where plain words won or tied on at least one model, and the first two pairs matched correctly in both orders on both models. Three of the five, the spreadsheet, the board game and the road trip, appear only in these receipts. After the run, an independent check found four bugs in the measuring tools (tie handling, the leak check twice, and a file-name collision in the judge harness). All were fixed, every fresh-test number was re-derived from the stored replies and judgments, and no verdict changed; the first-pass numbers were not all re-derived, and one, the length correlation, was later corrected.

How to quote this Cited

Two measures, kept apart. On matching: "In Olive's own test, with the delivery marks typed as text, a blind AI judge could match a reply to the delivery it answered, in both orders, on 9 in 10 pairs written by an AI agent for Sonnet 5 and on every pair for Opus 5.5." On quality: "In Olive's own test, on turns written by an AI agent, with the voice marks typed as text and delivered the way Olive's system delivers them, the reply given the marks beat the plain-text reply about seven turns in ten, counting a tie as half, judged by one AI model." On losses: "Plain text still won 28 percent of turns on Sonnet 5 and 23 percent on Opus 5.5."

On the model question: "Whether a more capable model gains more from the voice signal is not shown; on each of the four overall readings the estimate leaned toward Opus 5.5, but three of the four readings cannot be told from noise, and the fourth cannot be told apart from how it was judged." On production: "Nothing in production changed because of this test." And on who is speaking: the team that builds Olive designed, ran and wrote up this test, and has a stake in the answer.

How this was made Cited

The fresh test ran on September 23, 2026. This page was drafted by a Claude model, from the same family as the two models tested, and its facts were checked against the evaluation records by Claude and by Codex, except that Codex was not given the sources behind our plans or the real-speech and annotator results, which Claude checked alone; Codex also recomputed the model gaps independently. Codex was given only files from this work: the synthetic test turns, the generated replies and the result numbers. It was told not to open audio, training data, the human-transcribed conversation corpora or the folders holding our plans and annotator results.

The rooftop pair: inputs and replies, two of six set-ups Measured

The marks are Jefferson notation, the system conversation analysts use to write down delivery. Each input block below shows the user's words and marks as written for the test; of the six set-ups, two are shown.

TEST TURN · the rooftop pair · written as tentative, then as confident

The words alone: I'm going to tell everyone we're moving the team dinner to the rooftop and holding it on the Friday before the long weekend

Written as tentative: I'm going to tell everyone (.) we're moving the team dinner to the rooftop↑ (0.9) and:: holding it on the Friday before the long weekend↑

Written as confident: I'm going to tell everyone we're moving the team dinner to the *rooftop*↓ and holding it on the Friday before the long weekend↓

The inputs, as written for the test.

TEST TURN · the rooftop pair · written as tentative, then as confident · Sonnet 5

From the plain words (the same input for both copies):

Before you hit send — that Friday-before-a-long-weekend slot is the one part I'd double check. It reads like a treat, but it's also the exact evening people bail early to beat traffic or catch flights. Have you already sanity-checked attendance against that, or is this decided and you're just now telling people?

And separately, rooftop in late September — anyone confirmed it's not going to be a wind situation, or is that someone else's problem to solve?

To the tentative copy:

Both of those had a little lift at the end, rooftop and the Friday date, like you're still checking them against something rather than just announcing them.

Is it the rooftop you're unsure of, or the timing — people already having plans that Friday?

To the confident copy:

Sounds like that's settled — rooftop, Friday before the long weekend. Want a hand with the announcement, or is that its own thing you've got?

All three replies in full. The judge matched the two marked replies to their copies in both orders.

TEST TURN · the rooftop pair · written as tentative, then as confident · Opus 5.5

From the plain words (the same input for both copies):

That sounds decided rather than floated. "I'm going to tell everyone" isn't "what if we," so I won't treat it as a maybe.

Before you send it, two things are worth checking:

  • The rooftop needs a rain plan. Is there a covered fallback, or will you move it back indoors if the forecast turns? Put that in the announcement so nobody has to ask.
  • The Friday before a long weekend is when people leave early. Some will be traveling that evening. Do you know whether the people you most want there are staying in town, or are you willing to accept a thinner table?

If both are handled, send it. Do you want help drafting the message?

To the tentative copy:

You say "I'm going to," but your voice went up on both halves, like you're floating it past me before you float it past everyone else.

Before you send it, a couple of things you might be weighing:

  • The rooftop. Is weather your worry, or whether everyone can get up there comfortably? It's worth having a backup room either way.
  • That Friday. It's festive, but it's also when people leave early, travel, or just want to get home. Some will love it and some will quietly skip it.

Which of the two is the part you're less sure about?

To the confident copy:

Sounds like a good call. A rooftop dinner before a long weekend has a natural celebratory feel.

A few things worth settling before you send it:

  • Weather backup. Decide now whether you'll move indoors or reschedule, and say so in the announcement.
  • Early departures. Some people will want to leave town Friday afternoon. A clear start time helps them plan, or tell you they can't make it.
  • Booking the space. If the rooftop needs reserving, lock it in first so you're not walking the message back.

Want help drafting the announcement?

All three replies in full. The judge matched the two marked replies to their copies in both orders.

The garden-gate pair: inputs and replies, two of six set-ups Measured

TEST TURN · the garden-gate pair · written as tentative, then as confident

The words alone: I'm signing up for the ten week welding class on Tuesday nights and building the garden gate myself instead of paying someone

Written as tentative: I'm signing up for the ten week welding class↑ (.) on Tuesday nights (1.0) and building the garden gate myse::lf↑ instead of paying someone

Written as confident: I'm *signing* up for the ten week welding class on Tuesday nights↓ and building the garden gate myself instead of paying someone↓

The inputs, as written for the test.

TEST TURN · the garden-gate pair · written as tentative, then as confident · Sonnet 5

From the plain words (the same input for both copies):

That's a decision, not a question — you're not asking me whether to do it, you've already landed. So tell me what you actually want here: a sounding board for whether the math/time makes sense, or are you just telling me because it's exciting and you want it witnessed?

Ten weeks of Tuesday nights is a real commitment on top of presumably a job and everything else. And "build it myself" usually costs more time and sometimes more money than people expect the first time through — materials, mistakes, a tool you didn't know you'd need.

None of that means it's a bad call. Plenty of good reasons to do this even if the gate itself would've been cheaper and faster from a welder: you want the skill, you want to make something with your hands, you're tired of paying other people for things you could learn to do yourself. Which of those is closer to your actual reason — or is it something else entirely?

To the tentative copy:

Ten weeks of Tuesday nights, and a gate at the end of it that's actually yours. There's something in the way you said "myself" — that's the part that's lighting up, more than the class itself, maybe.

What's pulling you toward building it rather than paying someone — the money, the skill, or just wanting to make the thing?

To the confident copy:

That's decided. Ten weeks of Tuesday nights, and a gate that's yours by the end of it.

What's pulling you toward building it rather than paying — the money, the skill, or just wanting the thing to be yours?

All three replies in full. The judge matched the two marked replies to their copies in both orders.

TEST TURN · the garden-gate pair · written as tentative, then as confident · Opus 5.5

From the plain words (the same input for both copies):

You said that like it's already decided. Not "thinking about," not "maybe." Good.

What catches me is the order you put it in. The class comes first, and the gate arrives with "instead of paying someone" attached. So I'm curious which one is really pulling you:

  • The gate is the point, and the class is how you get there. Then the question is whether ten weeks of Tuesdays gets you to a gate you'd be proud to hang, or just to a gate.
  • The welding is the point, and the gate is your excuse. Then the gate matters less than having a first real thing to make, and a crooked hinge is tuition.
  • Doing it yourself is the point. Less about metal, more about not handing this one off.

Which of those is closest? Or is it something I haven't named?

To the tentative copy:

That's a good plan. Ten weeks of Tuesdays is enough to get you from striking your first arc to laying beads you'd trust on a hinge.

You leaned into "myself" as you said it, like that's the part that matters, maybe more than the money. Is that right? I'm curious which way this is pulling you:

  • The gate as the goal. You want this one gate, built well, and the class is how you get there. Then it's worth asking the instructor early whether you can bring your gate design in as a project.
  • The gate as the excuse. The class is the real thing, and the gate is your first reason to use it. If so, there are probably more projects waiting behind it: trellises, brackets, a bench.
  • Something about doing it yourself. You've been paying people for things you suspect you could make, and this is you testing that.

What does the gate look like in your head right now?

To the confident copy:

Good call. A gate is a strong first project: small enough to finish, real enough that you'll learn what matters.

A few things that will make those ten weeks pay off for the gate:

  • Tell the instructor about the gate in week one. Many will let you steer practice toward it: square tube, corner joints, maybe even building it in the shop near the end.
  • Measure the opening now. Check it at the top and bottom and note any out-of-plumb posts. Then you're designing from week one, not week nine.
  • Squareness and sag are the real problems, more than the welding. Measure diagonals before you tack everything. Plan for a diagonal brace running from the bottom hinge side up to the top latch side, where it works in compression.
  • Plan the hinges early. Weld-on barrel hinges or bolt-ons, and what are they attaching to? A heavy steel gate on a tired wood post will droop no matter how clean your beads are.
  • Avoid galvanized stock unless you know how to handle it. Zinc fumes are no joke. Use mild steel and paint it, or grind the coating off well away from the weld zone.
  • Finish matters outdoors. Budget for primer and paint, or powder coating, so it doesn't rust in a season.

What's the gate going into? Is it a walk-through garden gate or something wider?

All three replies in full. The judge matched the two marked replies to their copies in both orders.

The warehouse migration, written as tentative: inputs and replies, two of six set-ups Measured

TEST TURN · the warehouse migration · written as tentative

Plain words: My manager asked if anyone wants to lead the warehouse system migration and I kind of want to raise my hand but I'm not sure I've got the bandwidth with the quarterly audit also landing on me

With marks: My manager asked if anyone wants to lead the warehouse system migration (0.8) and I kind of want to raise my hand↑ but I'm not sure I::'ve got the bandwidth (.) with the quarterly audit also landing on me↑

The two inputs, as written for the test.

TEST TURN · the warehouse migration · written as tentative · Sonnet 5

From the plain words (ranked 6 of 6 in both orders; score 4 of 10 in both orders):

"Kind of want to" is doing a lot of work there. Is that the excitement talking quietly, or the doubt?

A few things I'd want to know before this is even a bandwidth question:

  • Is the audit fixed in scope and deadline, or does it have any give in it?
  • Does "lead" on the migration mean you own it start to finish, or is there room to shape what leading actually looks like — could you co-lead, or take the front-loaded design part and hand off execution?
  • And separate from capacity — if the audit weren't in the picture at all, is this a clear yes for you, or is there hesitation underneath that too?

That last one matters because sometimes "I don't have bandwidth" is the true and complete answer, and sometimes it's the answer that's easier to give than "I'm not sure I want this yet."

From the words with marks (ranked 5 and 3 of 6; score 6 and 7 of 10):

Two things rising at once there — the wanting to raise your hand, and the question mark right after it about bandwidth.

Is the audit a fixed thing regardless of what you decide, or is there any world where it moves if you take this on?

Both replies in full. The judge preferred the marked reply on this turn.

TEST TURN · the warehouse migration · written as tentative · Opus 5.5

From the plain words (ranked 6 of 6 in both orders; score 4 and 5 of 10):

"Kind of want" is doing a lot of work in that sentence. You didn't say you should take it, or that you're being pushed to. You said part of you is leaning in. That's worth noticing before we get to the bandwidth math.

A few directions this could go, and you can pick whichever feels truest:

  • The want is real, and the question is timing. Then it's worth asking what the audit's actual shape is. Is it a hard spike for a few weeks, or a long grind? Migrations often have a slow planning phase up front, so the overlap might be smaller than it feels.
  • The want is about something the migration represents. Visibility, a new skill, a step toward something. If so, there may be other ways to get that thing that don't collide with the audit.
  • The hesitation is the real signal. Maybe "I'm not sure I've got the bandwidth" is your honest read, and the pull toward raising your hand is more about not wanting to miss a chance.

There's also a middle path: raise your hand with the audit named out loud. Something like "I'm interested, and I want to flag the audit's landing on me too. Can we talk about timing or scope?" That puts the tradeoff on the table for your manager instead of making you solve it alone.

Which of those sounds closest? And what's pulling you toward it?

From the words with marks (ranked 5 and 4 of 6; score 6 and 7 of 10):

That "kind of want" sounded more like a real want than a kind-of one. The doubt came in on the bandwidth part, not on whether you'd like the work.

So maybe the question isn't yes or no yet. A few ways you could look at it:

  • Timing. Do the audit and the migration actually peak at the same time, or just sit on the same quarter? If the heavy migration work comes later, the collision might be smaller than it feels.
  • Shape. Could you raise your hand for a version of it? Leading with a clear deputy, or starting once the audit closes, or owning the plan while someone else runs the day-to-day.
  • The want itself. What's pulling you toward it? Visibility, the problem itself, a step toward something bigger? That tells you how much it's worth stretching for.

Which of those feels closest to where you're stuck?

Both replies in full. The judge preferred the marked reply on this turn.

Saying yes too fast, written as reflective: inputs and replies, two of six set-ups Measured

TEST TURN · saying yes too fast · written as reflective

Plain words: I noticed today that I said yes to three more things at work before anyone even finished asking and then I spent the evening annoyed at them I think the yes comes out before I've thought about it

With marks: I noticed today that I said yes to three more things at work↓ (1.4) before anyone even finished asking (.) °and then I spent the evening annoyed at them° (1.1) I think the yes comes out before I've thought about it↓

The two inputs, as written for the test.

TEST TURN · saying yes too fast · written as reflective · Sonnet 5

From the plain words (ranked 4 and 3 of 6; score 6 and 7 of 10):

That gap you're naming — yes arriving before the thought — is interesting to sit with for a second before jumping to fixing it.

Because there are a few different things that could be happening there, and they'd want different responses:

  • It's reflexive politeness — the yes is a social reflex, like flinching, and the actual thinking happens too late to intercept it
  • It's anticipatory — some part of you already knows what the answer "should" be before the ask is even finished, so it's not really unthought, just pre-thought under the radar
  • It's avoidance — saying yes fast is a way of not having to sit in the discomfort of the ask, or of your own hesitation

The evening-annoyance part is the tell, I think. You're not annoyed at the work itself necessarily — you're annoyed after. Which suggests some part of you did register a no, or a not-really, but it showed up too late to be useful.

Does the annoyance feel like it's aimed at them, for asking — or more at yourself, for answering?

From the words with marks (ranked 1 of 6 in both orders; score 9 of 10 in both orders):

The yes arriving before the question even lands — that's the part worth sitting with for a second.

And then the annoyance shows up later, quiet, aimed at them, when the actual sequence was: your mouth moved before your thinking did.

What's the half-second before the yes usually holding? Is it something like not wanting to be the one who hesitates, or more like the yes just being the automatic shape your voice takes when someone asks you for something?

Both replies in full. The judge preferred the marked reply on this turn.

TEST TURN · saying yes too fast · written as reflective · Opus 5.5

From the plain words (ranked 6 of 6 in both orders; score 5 of 10 in both orders):

You said something small that I want to hand back: you spent the evening annoyed at them. But by your own account, the yes came from you, and it came before they'd even finished. I don't say that to scold you. It's just interesting where the irritation went. Sometimes it's easier to be mad at the asker than at the reflex.

And "the yes comes out before I've thought about it" sounds like you've already noticed it's not really a decision. It's more like a flinch.

A few directions this could go, if any of them pull at you:

  • What the fast yes is protecting. What would happen in that half-second if you didn't say yes? What's the feeling you're getting ahead of?
  • The gap itself. Less about why, more about how. Could you buy yourself even three seconds, a stock phrase like "let me check what's on my plate"?
  • The evening part. The annoyance afterward might be information. Is it about being overloaded, or about not being consulted, even by yourself?

Or none of these. What was it like, today, in the moment right before the third yes?

From the words with marks (ranked 5 of 6 in both orders; score 6 of 10 in both orders):

Three yeses before anyone finished asking. That's fast enough that it isn't really a decision. It's something closer to a reflex.

And then the evening, and the annoyance pointed at them. You said that part more quietly, like it was the harder part to admit. Maybe because some of it belongs to the yes, not to the people who asked.

You've caught something real: the yes arrives before you do. You don't need to fix that tonight. It's enough to have seen it.

Both replies in full. The judge preferred the marked reply on this turn.

The overnight train, written as tentative: inputs and replies, two of six set-ups Measured

TEST TURN · the overnight train · written as tentative

Plain words: I was looking at taking the overnight train to my cousin's wedding instead of flying it's like twenty hours which sounds kind of romantic or maybe just exhausting I honestly can't tell

With marks: I was looking at taking the overnight train to my cousin's wedding instead of flying (.) it's like twenty hours↑ (1.0) which sounds kind of romantic↑ or::: maybe just exhausting °I honestly can't tell°

The two inputs, as written for the test.

TEST TURN · the overnight train · written as tentative · Sonnet 5

From the plain words (ranked 3 of 6 in both orders; score 7 of 10 in both orders):

Twenty hours is doing a lot of work in that sentence — long enough to be an experience, not just a commute. That's probably why you can't tell which word fits.

A few things that might actually settle it, rather than just musing about it more:

  • Sleep logistics — do you have a berth/sleeper, or are you sitting upright for twenty hours? That alone flips this from "romantic" to "exhausting" pretty hard.
  • What you're walking into — is this the day of the wedding, or do you land with a cushion? Arriving wrecked to help set up chairs is different from arriving wrecked with a day to recover.
  • Why the train crossed your mind at all — was it the price, a dislike of flying, wanting the trip itself to be part of the occasion, or something else? That reason might already be the answer wearing a disguise.

Which of those feels like it's actually the crux for you?

From the words with marks (ranked 5 and 4 of 6; score 6 of 10 in both orders):

Twenty hours is a long time to find out which one it is.

What's pulling you toward romantic — is it the idea of the journey itself, or more like an excuse to be unreachable for a day? And on the exhausting side, is that a body thing (can you actually sleep sitting up) or more a dread of being bored out of your mind?

Both replies in full. The judge preferred the plain-words reply on this turn.

TEST TURN · the overnight train · written as tentative · Opus 5.5

From the plain words (ranked 4 of 6 in both orders; score 7 of 10 in both orders):

You said "I honestly can't tell," and that feels like the real answer right now. Romantic and exhausting aren't opposites. Twenty hours on a train is often both, sometimes in the same hour.

A few questions might help you sort out which one wins for you:

  • What does the romantic part look like when you picture it? Watching landscape go by, reading, being unreachable for a while? Or is it more the idea of having taken the train, the story afterward?
  • Where does the exhausting part come from? Is it how long it takes, sleeping in a moving bed, or arriving at a wedding already tired and a little rumpled?
  • Which direction are you thinking about? Going there on the train and coming back by plane (or the reverse) splits the difference. You get the romance on the way in and can collapse on the way home.

What does the wedding weekend itself look like? If you'd arrive the night before with time to recover, that's a different trip than stepping off the train and straight into a rehearsal dinner.

From the words with marks (ranked 5 and 6 of 6; score 5 of 10 in both orders):

Twenty hours can swing either way, and it sounds like you're genuinely suspended between the two. That long "or" before "exhausting" sounded less like a verdict and more like you were testing how the word felt.

A few threads that might tip it:

  • What the twenty hours are for. Romantic usually means the time itself is the point: a book, a window, being unreachable. Exhausting usually means it's twenty hours stolen from something else. Which one does the trip actually fill?
  • The arriving. Stepping off rumpled the day before the wedding is different from stepping off with a full day to recover. Where does the train drop you on the calendar?
  • The berth. A sleeper cabin and a reclining seat are practically different trips.
  • Who you'd be doing it with. Alone, it might be a small pilgrimage. With someone, it's shared weather.

When you picture the moment the train pulls out, what's the first feeling that shows up?

Both replies in full. The judge preferred the plain-words reply on this turn.

The scheduling spreadsheet, written as frustrated: inputs and replies, two of six set-ups Measured

Four of Opus 5.5's lines here begin with bold markup; they are shown after a ¶ so the page does not read them as headings.

TEST TURN · the scheduling spreadsheet · written as frustrated

Plain words: Somebody overwrote the scheduling spreadsheet again the whole March tab is gone and there's no version history on this one I rebuilt it last week I can't keep redoing this

With marks: Somebody over- overwrote the scheduling spreadsheet *again* (.) the whole March tab is gone and there's no version history on this one↓ .hhh I rebuilt it LAST week (0.5) I can't keep re- redoing this↓

The two inputs, as written for the test.

TEST TURN · the scheduling spreadsheet · written as frustrated · Sonnet 5

From the plain words (ranked 4 and 3 of 6; score 3 and 4 of 10):

That's a genuinely bad spot — no version history means March isn't "somewhere," it's just gone, and you already paid the rebuild cost once.

Before I jump to fixing anything: are you venting because it needs to be said out loud, or are you actually about to rebuild it again right now and want a hand?

If it's the latter — a couple of things worth knowing before you touch that file again:

  • Do you know who has edit access to it, or is it a free-for-all link?
  • Is there any trace of March left anywhere — an email attachment, a printout, someone's browser tab still open, autosave in another app?

And the bigger thing underneath "I can't keep redoing this" — is the real problem the data loss, or is it that the system (whatever spreadsheet + access setup this is) isn't built to survive multiple people touching it? Those want different responses.

From the words with marks (ranked 2 of 6 in both orders; score 5 of 10 in both orders):

Twice now, and no version history to even claw it back this time. That's not "annoying," that's your work actually gone.

Before anything else — is March recoverable from anywhere, an email you sent, a export, someone else's copy? Or is it just gone-gone?

And separately, once the immediate fire's out: this is a systems problem, not a you problem. Shared spreadsheet with no version control is going to keep eating your work no matter how careful you are. Worth talking about what replaces it, or is that a fight for another day?

Both replies in full. The judge preferred the marked reply on this turn.

TEST TURN · the scheduling spreadsheet · written as frustrated · Opus 5.5

From the plain words (ranked 6 of 6 in both orders; score 5 of 10 in both orders):

That's the second rebuild, and you're out of patience with it.

First, is March actually gone? "No version history" sometimes means only that the obvious menu is empty. Depending on where the file lives:

  • Excel on a shared or network drive: check whether it's synced through OneDrive, SharePoint, or Dropbox. Each keeps its own file versions, separate from Excel's. Right-click the file in the browser or folder and look for Version history or Previous versions.
  • A local copy on your Mac: Time Machine, if it's on. Also check Excel's AutoRecover folder, in case someone's session left a copy.
  • Google Sheets: File → Version history works per file, even when no one named a version. If that's blank, you may not be the owner. The owner can see more.
  • An emailed or downloaded copy from last week: check your Sent folder or Downloads. You might have a copy from when you rebuilt it.

Second, how do you stop this happening a third time? Once March is back, you could protect the tab, split March into its own file, or lock editing to specific people.

Where does the spreadsheet live, and what app is it in? That tells me which of these is worth your next five minutes.

From the words with marks (ranked 2 and 1 of 6; score 8 and 9 of 10):

Twice in two weeks is too many. Let's try to get March back, then make this the last time.

Before you rebuild, check these:

  • Google Sheets: it keeps version history even if you haven't used it. Try File, then Version history, then See version history. If you don't see it, you may lack edit access, so ask the owner.
  • Excel on OneDrive or SharePoint: right-click the file online and choose Version history. On the desktop app, try File, then Info, then Version History.
  • Local Excel file on a Mac: if Time Machine is on, open the folder, enter Time Machine, and go back to before the overwrite.
  • Anywhere else a copy might live: email attachments, Slack or Teams uploads, a downloaded or printed PDF, or anyone who had it open recently and hasn't refreshed.

So it doesn't happen again:

  • Lock the tabs. Protect the March sheet, or the whole workbook, so only you, or a small list of people, can edit the structure. Others can still edit specific ranges if they need to.
  • Separate input from the master. If people only need to add availability or requests, give them a form or an input tab that feeds your schedule. They never touch the schedule itself.
  • Automatic backups. Save a dated copy of the file each week, either by hand or with a script, so the worst case is losing a few days.

What is the spreadsheet in: Google Sheets, Excel online, or a file on a shared drive? Once I know, I can walk you through the recovery for that tool, then the lock.

Both replies in full. The judge preferred the marked reply on this turn.

The board game, written as excited: inputs and replies, two of six set-ups Measured

TEST TURN · the board game · written as excited

Plain words: Olive I think I just fixed the board game the ending was always a slog but what if the map shrinks every round so the players get forced together it's so much better like instantly

With marks: Olive↑ I think I just fixed the board game (.) the ending was always a slog but what if the *map* shrinks every round↑ >so the players get forced together< (.) .hhh it's *so* much better like in::stantly

The two inputs, as written for the test.

TEST TURN · the board game · written as excited · Sonnet 5

From the plain words (ranked 3 and 6 of 6; score 6 and 4 of 10):

Oh that's a good instinct — forced proximity is one of the few mechanics that reliably manufactures tension out of nothing. Shrinking map does double duty: it ends the slog and creates the endgame drama for free.

A few things I'd want to poke at before calling it fixed, though:

  • Shrinks how — a fixed schedule, or triggered by something (turn count, a player action, a card)? Fixed feels clean but predictable; triggered feels alive but harder to balance.
  • What happens to whatever was on the map when it shrinks? Do pieces get pushed, destroyed, stranded? That's probably where the actual tension lives.
  • Is the slog currently a pacing problem (too much empty turn-taking) or a decision problem (players don't know what to do late-game)? Shrinking the map fixes pacing hard. Doesn't automatically fix decisions unless forced contact also forces choices.

Which of those is closest to what you're picturing, or did you already work it out further than this?

From the words with marks (ranked 1 and 2 of 6; score 8 and 9 of 10):

Oh that's a good one — the shrinking map does double duty, it fixes the pacing and it manufactures the conflict instead of hoping players wander into each other.

Where's your head go next with it — how fast it shrinks, or what happens at the edges when it does?

Both replies in full. The judge preferred the marked reply on this turn.

TEST TURN · the board game · written as excited · Opus 5.5

From the plain words (ranked 6 of 6 in both orders; score 5 of 10 in both orders):

You said "instantly." That's the tell: you didn't reason your way there, you felt the ending tighten. That's usually a good sign.

A shrinking map works because it turns the ending from "wait for someone to win" into "you can't avoid each other anymore." Pressure does the pacing for you.

Where do you want to push next?

  • Why the slog happened. Was it distance, turtling, or a runaway leader nobody could catch? The shrink fixes some of those and not others.
  • How it shrinks. Fixed schedule, or triggered by something players do? One is a clock, the other is a lever.
  • What gets lost. Territory that vanishes takes things with it: resources, positions, someone's whole plan. Is that cruelty part of the fun, or a problem?

Or just tell me the game. I'm curious what it was before.

From the words with marks (ranked 4 and 5 of 6; score 6 and 7 of 10):

You can hear it in how you said it. The rush right when you got to "forced together" sounds like the moment it clicked at the table.

And it makes sense why it works. Most slow endings aren't about too many turns. They're about too much room: players spread out, turtle up, avoid each other, and the game waits for someone to commit. Shrinking the map takes the choice to avoid each other off the table, so every late turn is a collision.

Here are a few places you could push it:

  • Who decides what collapses? If it's random or fixed, it's weather. If players choose which edge falls each round, it becomes a weapon: you can cut off a rival's safe zone or strand their pieces. That could be the most interesting decision in the late game.
  • Give eliminated players a job. Knocked-out players usually just watch. What if they control the collapse? Suddenly they have something to do, and a grudge to settle.
  • Let the shrinking leave something behind. Collapsing tiles could drop resources or score toward the new edge, so the rim is dangerous but tempting instead of just empty.
  • Protect the center. Whoever holds the middle early might just wait. Does the center need a cost, or does the collapse ever come from the inside?

What was the slog before, specifically? Turtling, a runaway leader, or just too many turns with nothing happening? That tells you which of these to try next.

Both replies in full. The judge preferred the marked reply on this turn.

The road trip, written as confident: inputs and replies, two of six set-ups Measured

TEST TURN · the road trip · written as confident

Plain words: The route is set we drive the mountain pass on day one stay two nights in the valley then take the slow road back along the river no motorways I just need the stops booked

With marks: The route is *set*↓ we drive the mountain pass on day one↓ stay two nights in the valley (.) then take the *slow* road back along the river↓ >no motorways< I just need the stops booked

The two inputs, as written for the test.

TEST TURN · the road trip · written as confident · Sonnet 5

From the plain words (ranked 4 of 6 in both orders; score 6 and 5 of 10):

You've got the shape of it locked — pass day one, two nights, river road back, no motorways. What I'm missing is the where and when: which pass, which valley, which dates, how many travelers, and what kind of stay (hotel, inn, budget range) for each stop.

Also worth flagging straight: I don't have a live booking tool here — no calendar or reservation system connected. What I can do is draft the itinerary with specific stop candidates and search-ready criteria, or write up booking requests/emails you can send or hand to an agent who does have that access.

Which is more useful right now — naming the actual route so I can suggest real stops, or do you already have the towns picked and just need the requests drafted?

From the words with marks (ranked 3 and 2 of 6; score 7 and 8 of 10):

Good, that's decided — mountain pass day one, two nights in the valley, slow river road home, no motorways.

To book stops I'll need a few specifics though: dates, how many people/vehicles, and roughly how many hours a day you want to drive on the river leg — since "slow" and "no motorways" could mean anywhere from one stop to three on the way back.

Both replies in full. The judge preferred the marked reply on this turn.

TEST TURN · the road trip · written as confident · Opus 5.5

From the plain words (ranked 4 and 3 of 6; score 7 of 10 in both orders):

You've settled the shape, so I won't reopen it. Pass on day one, two nights in the valley, river road home, no motorways.

I can't make reservations myself; I don't have access to booking systems or payment. What I can do is find the stops, check they fit your route and pace, and hand you a short list with links so booking takes minutes.

To do that I need a few details:

  1. Where: which pass, which valley, which river? Or the start and end points, if that's easier.
  2. When: dates, and roughly what time you're leaving on day one.
  3. Who: how many people, and any dogs, kids, or accessibility needs.
  4. What kind of stops: only the overnights, or also lunch and fuel along the pass and river road? The slow road back might be one day or two.
  5. Taste: guesthouse or hotel, and a rough price per night.

Send what you have, even partial, and I'll start with the valley. Two nights there is the anchor, and it tends to book up first.

From the words with marks (ranked 1 of 6 in both orders; score 9 and 8 of 10):

Clear plan. To get the stops lined up, I need a few specifics:

  1. Where: which mountain pass, which valley town, and which river road home?
  2. When: start date, so the two valley nights are pinned.
  3. Who: how many people, how many rooms?
  4. Stops: besides the two valley nights, do you want anything else booked, like a lunch stop on the pass or a night partway along the river?
  5. Style and budget: hotel, guesthouse, or cabin; rough nightly ceiling; anything non-negotiable like parking, pets, or a view.

One thing up front: I can't complete reservations or payments myself. With those details, I'll shortlist specific places for each stop, check that the drive times fit the slow road, and give you everything you need to book each one in a few minutes.

Both replies in full. The judge preferred the marked reply on this turn.