Reading Between the Lines

So that foreign professionals in Japan stop getting stuck between honne and tatemae

PRD ／ 2026-08-15 ／ Companion: Measuring pragmatic comprehension ／ ※ Company, product, volume and site names are anonymised

日本語|EN|한국어

Whose	Of 2.3 million foreign workers in Japan, the 475,790 in professional roles (+67.9% in five years). People who already passed the language exams
What problem	They keep waiting without realising "we'll consider it" was a refusal. They speak directly and get opposed for how they said it. And above all, they never notice the divergence happened
How	Make them decide for themselves first, then reveal the answer. Show which type they are weak at, then have them compose one follow-up question they could actually use

01Problem

In Japanese workplaces refusals are not stated directly. "We will consider it" is, in most cases, a refusal — and it does not appear on any language proficiency test. Japanese speakers themselves treat it as something to be sensed rather than taught, so it is never made explicit.

Yun (29) is a Korean national who came to Japan after high school and completed graduate school here. His Japanese is sufficient for work; he is in his third year in the R&D division of a domestic manufacturer. He still gets stuck from two directions. On the receiving side, he takes "we'll consider it" as progress and only realises a week later that nothing moved. On the sending side, he states his conclusion clearly and gets opposed not for the content but for the delivery. The cost shows up as work stopping — in July 2026 a presentation was received badly, his motivation broke immediately after, and productivity dropped. And there is nowhere to practise. Because the right reading shifts with the person, the setting and the timing, no teaching material forms; you learn in the real thing, and the real thing cannot be undone.

It looks like one person's language problem, but the scale is not small. Japan has 2,302,587 foreign workers across 342,087 workplaces (MHLW, October 2024) — and no official statistic on why they leave, because the Employment Trend Survey carries no nationality breakdown. ※ See the companion document for the measurement side.

02Goals / Non-Goals

1. Less time spent waiting on a misread, or doing work that turns out to be unnecessary／metric: days until the misreading is noticed, and the amount of rework done in between
2. Fewer proposals lost because the speaker unintentionally signalled objection／metric: number of times a proposal was sent back over how it was raised
3. Fewer misreadings of phrases whose meaning turns on intonation／metric: error rate on items presenting the same sentence with different intonation
4. When a high-context phrase does cause a mismatch, it gets noticed more often／metric: share of mismatched items where the user flagged "we may have just diverged"

G1–G3 aim to reduce misreadings; G4 aims to notice them. The first three have a ceiling — you do not get to choose how the other person phrases things. But if you notice, you can ask right there. We believe G4 is what actually pays off at work.

Non-Goals

1. Correcting grammar or keigo — existing services cover this. We handle meaning, not form
2. Sales roleplay — likewise. We do not train delivery itself
3. Asserting a single correct reading — meaning is situational; a single answer would be a lie
4. Reproducing real people or real companies — we build no responses modelled on a specific individual
5. Third-party evaluation or scoring — if someone else grades you, you stop writing honestly, and it gets repurposed for HR review
6. Doing the actual workplace negotiation for you — this is a practice ground, not a stand-in for building internal agreement
7. Training Japanese speakers to be lower-context — there is demand, but this document covers the receiving side only

03User Stories

1. Someone who took a vague reply as progress can check on the spot instead of waiting.
   BenefitReading it as "this may be stalled" lets you ask one question before you start waiting／TodayThey take "we will consider it in that direction" as forward motion, wait, and only discover the mismatch at the meeting a week later
2. Someone whose reading diverged on a high-context phrase notices the divergence itself, and can ask the speaker right then.
   BenefitNoticing enables asking; asking stops the misreading from travelling／TodayThey never notice the divergence, so the work proceeds on a misunderstanding

04Proposed Solution

A practice tool built on short items. It shows a setting and a line, makes you commit to a reading first, and only then gives the answer and the explanation. It does not teach words or grammar; it manufactures the experience of having misread, somewhere safe.

1. Item and self-judgement — show the setting and the line, make the user choose a meaning. The explanation appears only after answering 〈US-001 · US-002〉
2. Answer and cues — why it reads that way, and how it could be verified 〈US-001〉
3. Breakdown by conventionality — accuracy and response time per type, showing which type is weak 〈US-002〉
4. A line you were actually told — for real utterances, write your own reading first, then compare 〈US-001〉
5. Practising the follow-up question — straight after writing your reading in 4, compose one usable follow-up question. Not practice in correcting the other person, but practice in asking whether your understanding is right 〈US-002〉
6. Items with intonation — present the same sentence with different intonation and ask for the difference in meaning. Text alone cannot reach G3 〈G3〉

※ The core is the ordering in 1. Show the answer first and it stops being reading practice — it becomes a dictionary. Judging, then being wrong is what makes it stick. When information is insufficient, do not narrow the candidates; ask what would narrow them.

※ 5 is the crux of US-002. Noticing without being able to ask leaves you exactly where not noticing would. So noticing and speaking are practised as one motion.

※ 6 is not in the current prototype. Items 1–5 run on text alone; intonation items require recorded or synthesised audio. Feasibility is left in Open Questions.

※ Items 1–3 and 6 use no generation and work without a key. Only 4 and 5 run on the user's own key (BYOK); that input never passes through our server.

05（Mockup）

A working demo is deployed separately: Reading Between the Lines. The 10 practice items run without an API key.

06Evidence & detail

What is said, and what it actually means (5 examples)

What is said	Literally	Actual meaning ※
ちょっと難しいですね	a little difficult	refusal
検討させていただきます	we will consider it	refusal
前向きに考えます	we'll think positively	undecided; often refusal
もう少し様子を見ましょう	let's watch a little longer	on hold, or quietly dropped
参考にさせていただきます	we'll use it as reference	will not be adopted

※ These are what is generally said to be meant, not assertions. Meaning shifts with person, setting and timing, so this is not to be treated as a fixed lookup table (R5).

Market size — 475,790 people (+67.9% in five years)

Residents under the "Engineer/Specialist in Humanities/International Services" status numbered 475,790 as of December 2025, up 67.9% from 283,380 five years earlier. They were hired after passing language exams. What blocks them is not language ability but this layer.

Source: Statistics on Foreign Residents (Immigration Services Agency / e-Stat)

Why there is nowhere to practise

Because the premise "the real intent cannot be spoken, only sensed" is widely shared, teaching material never forms in the first place. Commentary articles and a chapter in an advanced textbook exist, but there is no tool to practise with.

On top of that the right answer shifts by situation. The same "that's difficult" means different things depending on the person, the setting and the timing. A fixed table is not enough; you have to be given a situation and see the reaction.

Implementation notes (High-Level Tech Consideration)

• The counterpart is treated as an agent parameterised by position. Not a reproduction of a personality, but the reaction of "a person in this relative position".
• Confidence in each reading is never asserted. Because it shifts with person, setting and timing, it is presented as a range.
• Runs on the user's own API key (BYOK). Since nothing passes through our server, writing comfort is preserved without sacrificing quality.

※ There is an accumulated body of research on refusals and consideration expressions in Japanese language education and pragmatics (Journal of Pragmatics Research vol.14 / Kansai University Wakumon 28, among others). It informs the design of reactions, but no particular theoretical framework is made a precondition of the product.

07Open Questions / Risks

Open Questions (6)

• What underwrites the validity of a reading(→ the distribution of native-speaker answers becomes the reference. How to treat items where that distribution splits is undecided)
• The current frequency of rework and send-backs is unmeasured(→ G1 and G2 cannot yet be stated quantitatively; for now they rest on self-report)
• How to build the intonation items(→ recording imports one speaker's habits; synthesis loses naturalness. Whether G3 is achievable hangs on this.)
• How to measure "noticed"(→ the G4 metric. Relying on self-report misses exactly the people who did not notice, since they do not report either. We need noticing to be observable from outside)
• Is BYOK a barrier for users(→ if so, provide a path that uses our key over abstracted input)
• Could the absence of emotional support drive people awayThis tool deliberately makes you experience bad reactions. Returning nothing when someone is discouraged is a reason to quit. Going further, though, enters counselling. Whether this belongs in Non-Goals is still open.

Risks (6, PMBOK form)

#	Risk	Cause → event → impact	Response
R1	Hard to write	Sensitive detail required → hesitation → practice never starts	Mitigate: fewer free-text fields, mostly multiple choice, plus BYOK so nothing passes through our server
R2	Mistaken for a language app	Looks like Japanese teaching → compared with incumbents → loses distinctiveness	Avoid: no correction feature; lead with component 5 (the follow-up question)
R3	Wrong reading	Highly situational → asserted flatly → false confidence	Mitigate: no single answer; present candidates and their spread
R4	Abandonment	Input burden → not used → self-validation fails	Mitigate: two weeks of self-validation on one real case before deciding to expand
R5	Cultural flattening	Generalisation → diverges from reality → reinforces prejudice	Mitigate: assume organisational and individual variation; no fixed lookup table at the centre
R6	G3 not delivered	No audio available → intonation items cannot be built → G3 becomes an empty promise	Mitigate: place G3 in phase 2 and state explicitly that it is out of the phase-1 promise

Weaknesses of this PRD

① The quantification in Problem is insufficient. 475,790 people and +67.9% describe market size, and the presentation incident is a real case, but there is no first-party number for "how many misreadings, how many hours of rework." It cannot be fabricated, so it sits in Open Questions.

② This document contains no means of confirming the validity of reactions. How close the counterpart's reaction is to a real workplace has not been verified. Framing it as "a rehearsal, not a prediction" sidesteps the issue, but that is not verification. The design of effect measurement remains in Open Questions.