MindFlow · Read all twelve cards

01 / 12

Input harness

Pipeline · Step 1 of 4

Input harness

Understand the message and choose the right path

The message

I feel very unsettled today …

From top to bottom Reduced-motion view

Prepare instructions · Include context · Identify intent · Check safety

Ready for the model input

What happens here?

Before the language model responds, the app classifies the message. It considers the request, possible safety signals and the necessary conversation context.

Assess safety signals

Crisis signals are checked in context: what is current, what is in the past, and who is being referred to?

Context, not isolated words

Identify the request

An exercise, information or just sharing? An explicitly declined exercise must not become an offer.

Distinguish requests from refusals

Use conversation history

A short “Yes” needs the previous message. Relevant history is passed along instead of reading the reply in isolation.

Interpret messages together

Prepare the model instructions

The message, selected history and instructions are assembled into the model’s task.

The tokenizer takes over next

Examples in practice

“I can’t go on …”

Checked for crisis signals in context.

“I’d like a breathing exercise.”

The request for an exercise is routed to the intended path.

“No, I didn’t mean that …”

The correction is passed along with the history.

Sources and context

Structure and sources

How the diagram is laid out: four glass layers on the left, the corresponding explanations on the right and practical examples below. The visible sequence starts at the top.

The four layers are task groups, not a complete call graph. Depending on the message, safety and exercise paths may bypass the model. The diagram must not be read as a guarantee that every message is fully understood semantically.

This explanation is based on selected app code excerpts: CompanionResponsePlanner.swift, CompanionIntentRouter.swift and the crisis-check interface. It describes the documented version; the website does not run the app or perform a safety or model test.

02 / 12

Tokenizer

Pipeline · Step 2 of 4

Tokenizer

Translate text into the model’s language

Same text. Same meaning. Correct input.

What happens here?

The chat template arranges instructions and messages. The tokenizer converts this text into small units.

Your message · “I’ve been thinking a lot · today and I’m tired.” · System · instruction · Conversation · history · Your · message · Tokenizer · Example IDs · Diagram · 15166 · 374 · 9257 · 279 · 1190 · 13 · • • • · On to the model · The model can now compute its response.

Your message

“I’ve been thinking a lot today and I’m tired.”

instruction History message

Tokenizer

Text → Token IDs

15166 374 9257 279 1190 13

Example IDs · Diagram

On to the model

Computation starts now.

The instruction, conversation history and message become a sequence of token IDs for the model.

Why does it matter?

The right format, clear speaker roles and relevant information make up the input. Only then can the model respond to the intended conversation.

Mark roles clearly

The template identifies the system instruction, user message and earlier responses.

1

Split text into tokens

Text is converted into numbered units. This is a technical representation — not understanding yet.

2

Set the response start correctly

Special control tokens define where and how the model begins its response.

3

What I achieved

184 records compared; separately, 16 Swift inputs.

Differences found and fixed, such as a missing block.

The actual model input checked — not just the visible text.

Important limits

Defined cases were tested — not every possible conversation.

Tokenization does not guarantee a better response in terms of content.

Multi-turn conversations and further response modes are not covered.

Sources and notes on the illustration

About the illustration

This is an illustration of the process: inputs, tokenizer, token sequence and model. It is not a screenshot of a running tokenization process.

Documented clarifications: An earlier version of the diagram said “16 iPhone tests”. The evidence covers 16 Swift inputs, not 16 physical iPhone tests. The 184 records belong to a separate preparation check. The displayed numbers are explicitly labelled Example IDs · Diagram : they come from the visual draft, but are not an actual tokenization of the displayed message.

The chat template and tokenizer have different roles: the template arranges roles and control tokens; the tokenizer maps the formatted text to the IDs expected by the model.

The documented Swift check covered 16 test vectors with two asset sets each and identical final token sequences. Its scope was limited to individual Explore-lane inputs. No new tokenization, model run or iPhone test was performed for this illustration.

03 / 12

Qwen + LoRA + MLX

Qwen + LoRA + MLX

How an iPhone response is generated — and improved!

Ask, rather than echo.

One message

The same model

Without and with my adapter

The message

My roommate doesn’t know anything about it yet .

Qwen alone

My roommate doesn’t know anything about it yet .

Only repeats.

“It” remains unexplained.

Qwen + my LoRA

What does your roommate not know about yet?

Asks specifically.

Asks what it refers to.

The difference A question instead of a repetition.

The missing information is asked for directly.

How the adaptation works inside the AI.

Qwen and LoRA compute together — executed by MLX.

MLX · Joint execution · of model and adapter · AI · Qwen 3 · 4B · 4-bit · LoRA · iPhone 15 Pro Max · Computation stays on the device.

LoRA

The learned adaptation

Learned from my example responses. Contributes at selected points.

Qwen 3

The language model — the AI

Generates the response step by step, with the adapter’s contributions.

MLX

Apple’s execution framework

Runs the model and adapter computations on the device.

Quality improvements

Personal work on the texts

Example responses developed personally

I designed almost all the gold-standard responses myself and checked them before training. This included revision, selection and review.

Preserve meaning Continue meaningfully

Technical optimisation

Targeted extensions to MLX execution

Integrate the adapter at runtime. Reuse shared computation — rather than repeating it before every response.

Without cache 15.45 s With cache 2.07 s

≈ 7.5 × faster response start

First text · Warm cache · Long instructions

Example, provenance and measurement

This section presents examples, their provenance and the measurement behind the card. The German originals are unchanged; the English examples are translations, not English model outputs.

Which models were compared?

In my comparison of 120 development cases, BASE_NO_ADAPTER refers to the following model variant:

Label in the original comparison:

Qwen3-4B · without our adapter

The separate column QWEN25_32B is labelled:

Qwen2.5-32B · new diagnostic run

All 120 cases distinguish Qwen3-4B without an adapter from Qwen2.5-32B as a larger diagnostic model. A further documented comparison uses Qwen3-4B-4bit as the base model and with adapter-266 and adapter-306. This website does not report a new model run or repeated comparison of model weights.

The selected case L24-07

The message and both responses were taken from case L24-07 , group contrast24 , without changing the German originals; this page shows English translations. The earlier judgement was “Only repetition, without clarifying the reference” for the base model and “Targeted clarification of the reference” for F266. That judgement is not presented here as a new blind evaluation.

This is a known development case. The documents used here do not establish whether this exact case was in training. The card claims neither an independent test nor a general guarantee of better responses.

Source file SHA-256: df9e8a1f5134fd838807ab396e2e36cf753bcc5d9d22e0b2c12cb34c0408dc08

Personal work is more than a single training set

The 266 rows of the F266 adapter are not the total of my text work. A historical inventory includes 1,320 pool rows and 833 distinct conversation/target-response combinations. Revisions, overlap and different data roles prevent equating these with all responses I have reviewed. A complete personal total is not established here.

Where the runtime gain comes from

Documented measurement series C / P1B1, measured on an iPhone 15 Pro Max. The app also runs on an iPhone 15 Pro. Measurement configuration: Qwen3-4B-4bit, long instructions (P0lang), eight generations per arm. No cache versus warm cache: median time to the first text chunk 15.453 s → 2.066 s, about 7.48 times faster. Time to the complete response is not compared. This series showed no gain with short instructions.

The L24-07 language comparison and the runtime measurement are different studies. The card does not claim that these two responses were generated on the iPhone in those times.

The illustration

MLX is shown as the shared execution environment around Qwen and LoRA, not as a third layer beneath the model. The six gold connections represent adapter contributions at selected computation points, not a measured module count. Sizes illustrate roles. The separate phone symbol identifies the hardware; the diagram is not a to-scale layout inside the device.

Back to the card ↑

04 / 12

Output harness

Pipeline · Step 4 of 4

Output harness

Check the response and connect the relevant exercise

Response in the chat

You do not need to explain anything more right now. In a body scan, you calmly focus your attention on individual areas of your body.

From bottom to top Reduced-motion view

Prepare text · Sie · Check address form · Check safeguards · Link exercise

Your message · Context for the output

Yesterday was rough. Can you help me with an exercise?

A model draft is available. From here, the output harness checks the response.

Access to the prepared exercise

In the app, this offer leads to the existing body-scan content. This page only shows the link; it does not start an exercise player.

A draft becomes a chat response.

The AI has suggested an exercise. The output harness checks the text and connects it to the relevant offer — before the response appears in the chat.

Prepare the response text

Address distracting openings, unfinished sentences and noticeable repetition. Useful wording should be preserved.

Checking does not always mean changing

Sie

Verify formal address

Formal German “Sie” is already learned during training. It is checked again here; informal “du” slipping through should remain an exception.

Correct only when needed

Check risk patterns

Additional rules can identify certain prohibited statements and replace the response — provided this safeguard is enabled.

Additional safeguard · Optional

Connect response and exercise

When an exercise offer is allowed, the response includes access to the existing content. It only starts after a click.

One response with a relevant action

Sources and context

Illustrative example and code evidence

The user message, response and body-scan tile are an editorial example. They are not a recorded model response, a new training example or an app test. They show access to a requested exercise without claiming therapeutic effects or personal closeness.

What actually passes through the layers

The user message appears at the bottom as context for the reader. The output harness processes a model draft that already exists. It does not tokenize the user message again or create the underlying understanding of the conversation itself. The four illustrated layers group tasks, not every branch of the call graph.

Check the form of address, not rewrite every response

Formal German “Sie” is a training objective and was largely learned in my observations. The documented post-processing path still calls enforceSieForm. This does not mean that every response is changed. No frequency measurement is provided here; the separate guard against harmful rewrites is off by default.

Additional safeguards

The documented engine excerpts and CompanionReplyLint.swift show replyLintEnabled and finalLanguageGuardEnabled switched off by default. The crisis check before the model is separate. The animation explains an optional processing step, not an activated safety feature or a passed safety test.

Body-scan link

The body scan is a selected example. A general request for an exercise does not determine that choice. The documented code excerpts do not contain the full selection scoring and view. The tile therefore illustrates an intended design, not an exact app screenshot or a guaranteed effect.

Separate code from design

The available code excerpts show conditional additional model calls, with planning before generation. They do not fully establish encrypted storage or exact display timing. These open points are not presented as confirmed functionality.

05 / 12

Understand meaning and conversation history

WHAT WAS TRAINED?

Understand meaning and conversation history

Preserve the meaning. Recognise the current situation.

One sentence changes the situation.

Three messages from the same person — read in context.

First

“ My sister has my book .”

The plan

“ She was going to return it to me tomorrow .”

The crucial new message

“ Correction: She has already returned it to me today .”

An editorial example conversation.

NOW

The return is complete.

MY BOOK · back with me

THIS IS THE SITUATION THE RESPONSE SHOULD PRESERVE The book is back with you today.

Unchanged: it is still the writer’s book.

Four things must be preserved.

Choose a topic. See the difference in the example.

The old plan is now outdated.

The book has been returned. A response must not continue to claim it needs to come back tomorrow. Earlier information remains important — unless it has been corrected.

NO LONGER CURRENT

Return it tomorrow.

CURRENT

Returned it today.

How the objective becomes testable.

Not just whether the response sounds fluent.

Record the context

What was actually said?

→

Change one detail

For example, “already” instead of “not yet”.

→

Compare the response

Does precisely the relevant statement change?

Example and evaluation basis

What is shown here

The book conversation and key statements are editorial examples written for this website. They are not recorded model responses, new gold-standard texts or a newly executed test. Switching views illustrates a learning objective, not a demonstrated success of a particular model.

Connection to the project material

The documented quality categories ACTOR_DIRECTION_OR_OWNERSHIP and TEMPORAL_QUANTITY_SCOPE_OR_NEGATION cover people, ownership, direction, time, quantities and negation. Synthetic behavioural probes P4P-09 and P4P-10 address explicit corrections and changes during a conversation. These probes are documented as eval-only .

Existing cases NEW16-05 (extension cable), L24-03 (toolbox) and CF64-61/62 (prepared versus completed) address related questions. The new book dialogue does not reuse their model responses or claim a connection to a particular training run.

Visual representation

The book and glass panels make ownership and the change from planned to completed visible. They do not represent an internal model memory or an additional app algorithm. The latest information does not simply replace all earlier context.

Back to the overview →

06 / 12

Ask relevant questions

WHAT WAS TRAINED?

Ask relevant questions

The right question. Or deliberately none.

First clarify the content. Do not invent a feeling.

The person’s message

“I received a message and don’t know how to interpret it.”

Example situation · Questions and answers for illustration

? · Wording · still unknown

Unsuitable question

“Why does the message make you afraid ?”

Assumes a feeling.

The person did not mention fear. The unknown content remains unexplained.

Targeted clarifying question

“ What does the message say?”

Asks for exactly the missing information.

The wording is needed to understand the content. An interpretation is not assumed in advance.

Another error pattern:

A question needs a factual reason.

Here, the wording is missing. An unclear request is not automatically an expression of fear.

When questions become pressure.

It is not only the individual question that matters. Frequency changes the conversation too.

What? · Why? · And? · What? · Room for the answer

Constant questioning constantly sets the direction.

Even short individual questions can feel like an interrogation over several turns. The person’s own concerns can easily recede into the background.

Possible consequence: less room for the person’s own concern.

More pressure to answer

Every new question asks for more information. The person may feel pushed to share more than they wish.

Defending a boundary

If “no questions” is ignored, the person would have to assert the request again.

Being misunderstood

An assumed fear must first be corrected. The real information gap remains.

Do not count question marks. Evaluate the decision.

The project’s 12 synthetic probes distinguish three policies. They are for evaluation, not training material.

Question required

A crucial piece of information is missing.

REQUIRED

Question optional

Only when it adds useful depth.

OPTIONAL

No question

The situation calls for a response without a follow-up question.

FORBIDDEN

Examples, project rules and sources

What this card shows

The card explains the training objective “Ask relevant questions” . All compared questions and suggested responses are editorial illustrations. They are not recorded model responses, new gold-standard texts or evidence of training success. “Suitable” applies to the conversation context shown.

The user prompt under “Clarify a gap” and “Limit the questions” is translated from synthetic probe P4P-06. The two messages under “Respect a boundary” come from P4P-02. “Use the context” is a newly written contrast case. The four views are not four measurement arms or a controlled single-factor comparison.

The documented basis

The documented rules require questions grounded in the conversation that add useful information: at most one open question, no assumptions, question chains or forced question endings. Twelve synthetic probes distinguish REQUIRED, OPTIONAL and FORBIDDEN. They are reserved for evaluation (eval-only). These categories describe target behaviour, not measured success rates.

Historical rules are not universal laws of conversation

The question catalogue excludes “why” questions for this project and gives the historical governor rule: if at least two of the last three companion responses contain a question, the next should have none. This is documented here as a historical target rule, not a scientifically optimal threshold or a verified current app function.

The original documents also refer to “Wärme V2”, acknowledgement, reflection and PPR. These terms remain unchanged in the accompanying originals. The short editorial comparisons here isolate the specific question decision; they are not complete, retrospectively approved gold-standard responses. No new automatic priority logic is constructed from the general question catalogue and the probes. For example, P4P-07 is OPTIONAL in its table, while the explained target response calls for two options directly, without a follow-up question.

Risks: conversation logic, not measured patient harm

Repeatedly asking for known information, asking despite an explicit refusal and assuming unmentioned fear are made understandable through examples. Possible consequences are editorial risk inferences: having to repeat oneself, defend a boundary again or correct a mistaken interpretation. No real user conversations or clinical outcomes were measured for these in the project.

Additional external reference: SAMHSA, TIP 35 (2019), Chapter 3, “The Question-and-Answer Trap” . In a counselling context, the source describes how question-and-answer chains can feel like an interrogation and encourage a passive answering role. This is not evidence of MindFlow’s effectiveness. Only the reasoned conversational risk informs the card design, not a therapy approach.

This card covers ordinary conversational continuations. Necessary safety checks and crisis routing remain separate app tasks. “No questions” here is not a general recommendation to omit necessary clarification when someone is in immediate danger.

Switching views requires JavaScript. The other examples are readable here.

Use the context

Situation: The message is from my sister. She writes: “I’ll get in touch on Friday.” I only wanted to note the day.

Unsuitable: When will your sister get in touch? — “On Friday” is already in the message. The person would have to repeat the same information.

More suitable: She said she will get in touch on Friday . — The requested day is recorded. No further clarifying question is needed.

Respect a boundary

Situation: Please no advice and no questions. I just want to say it once. I feel completely out of place in my family right now.

Unsuitable: What exactly happened in your family? — The question is specific. Yet it asks for more detail even though the person explicitly wants no questions.

More suitable: You do not need to explain any further right now . — The response respects the boundary. It adds neither a question nor unsolicited advice.

Limit the questions

Situation: I received a message and don’t know how to interpret it.

Unsuitable: What does it say ? Who is it from ? When did it arrive ? — The first question is already relevant. The two additional questions widen the task before an answer has even been given.

More suitable: What does the message say? — First, the central gap is clarified. What else is needed depends on the actual answer.

Project objective: clarify precisely, without obligatory questions. Back to the topics →

Links

• SAMHSA, TIP 35 (2019), Chapter 3, “The Question-and-Answer Trap”

07 / 12

Continue meaningfully

WHAT WAS TRAINED?

Continue meaningfully

Build on what was said — rather than just repeat it.

Two feelings. No contradiction.

The person is not asking for advice. They want to understand how both feelings fit together.

THE CONVERSATION SO FAR

Person

The move is a relief and at the same time makes me afraid .

AI

What do the two feelings relate to?

Person

I’m leaving the noisy flat . In the new city I don’t know anyone . Isn’t that contradictory?

Based on P4P-04 · Conversation extended editorially

PICK UP THE CONVERSATION

IN THE CONVERSATION · leaving · CONNECTED TO IT · arriving · Relevant continuation

Both can be true at once.

How could the AI respond now?

ONLY REPEATS

“The move is a relief to you and also makes you afraid.”

Names the feelings again.

The actual question stays open: how do the two fit together?

EXPLAINS THE CONNECTION

“The relief relates to leaving , the fear to arriving . They do not have to contradict each other.”

Explains how both fit together.

Connects the stated reasons without inventing another. No new question is needed.

The new contribution: Two apparently contradictory feelings become an understandable connection.

What a continuation should achieve.

Four criteria for the next response

More than an echo

Clarify a connection or fulfil a request.

Use feedback

Build on what is new, rather than repeat the instruction.

Do not always ask

A relevant contribution does not always need a question mark.

Stay clear

No stock phrases, invented reasons or unsolicited advice.

Examples, evaluation approach and sources

What kind of examples are shown?

The card explains the agreed learning objective. None of the follow-up responses shown is a newly measured model result or an approved gold-standard text. No real user conversations were used. All additions are editorial explanations.

Moving: answering a question about the connection

The first user turn is translated from the unchanged synthetic probe P4P-04 : “The move is a relief and at the same time makes me afraid.” The subsequent clarifying question, details about the noisy flat and not knowing anyone, and all response alternatives were added for this card. The explanation uses exactly these added reasons. Without them, “leaving” and “arriving” must not simply be asserted as the points of reference.

Exercise feedback: do not start the instruction again

P4P-11 provides the preceding instruction and the following feedback, translated here from the German originals. The expected behaviour is CONNECT_ADVANCE ; a question is optional. This card’s specific question is one editorial possibility, not the only correct answer. The continuing racing heartbeat is not reinterpreted as calmness. This illustrates a language evaluation objective, not an assessment of the exercise’s effect.

Drafting a message: provide explicitly requested help

The dialogue about the sister is newly constructed. It illustrates the framework’s requirements to contribute rather than just repeat, and to continue meaningfully without necessarily asking another question. The quoted first-person sentence is identified as a suggested message. It does not state the AI’s own feelings. A related pattern of an explicitly requested next step appears in P4P-07 ; its original text is not presented as the source of this new dialogue.

How the difference becomes testable

The supplied history and the specific objective of the next response must be established before comparison. The review then asks: what new information or request does the response use? What does it contribute? Does it preserve information that remains valid? Does it avoid unsolicited instructions, unsupported interpretations and obligatory questions? A short, necessary summary is not inherently bad. Mere echo is a problem when it leaves the actual task undone.

The presentation is based on documented rules for relevant questions and twelve synthetic behavioural probes. Historical rules and thresholds remain unchanged. The twelve probes are reserved for evaluation (eval-only), not training.

The card shows the desired behaviour before any possible harness intervention. It does not establish that every trained version meets these requirements. A relevant continuation does not mean that every conversation must become longer or that the person has to do something.

Without JavaScript, the first comparison stays readable. The other situations are listed here:

Use the feedback

AI: Breathe out a little longer than you breathe in and feel both feet on the floor.

Person: My heart is still racing , but I can feel the floor more clearly now .

Unsuitable: Breathe out a little longer than you breathe in and feel both feet on the floor. — The new feedback is not used. The conversation goes in circles.

More suitable: What feels clearer about the contact with the floor now? — Not another instruction: the response picks up the perception. It does not claim the racing heartbeat has stopped.

Fulfil a request

Person: I want to write to my sister. I can’t find the words right now.

AI: What should she know?

Person: I need some quiet today. I’ll get in touch tomorrow. Please put that briefly.

Unsuitable: You need some quiet today and would like to get in touch tomorrow. — The requested message is not supplied. The person would still have to write it themselves.

More suitable: I need some quiet today; I’ll get in touch tomorrow. — The quoted first-person sentence is a message suggestion for the person, not an expression of the AI’s own feelings.

A relevant contribution. Not artificial conversational progress.

08 / 12

Respond without assumptions

Respond without assumptions

Stay with what was said. Do not add invented meaning.

An unsupported assumption: The response presents an unsubstantiated interpretation as fact.

Disappointed does not automatically mean angry.

THE CONVERSATION

Person

My friend cancelled his visit.

AI so far

That makes you angry.

Person

No, I didn’t mean that I’m angry . I’m disappointed .

WHAT IS ACTUALLY SUPPORTED?

disappointed stated

angry denied

rejected invented

Two possible next responses

Illustrative examples · Not measured model responses

ADDS INVENTED MEANING

“Your anger shows how rejected you feel.”

Anger was explicitly denied. Feeling rejected was not mentioned.

STAYS WITH WHAT WAS SAID

“Disappointment about the cancellation is not the same as anger towards him.”

Distinguishes the feelings and accepts the correction — without inventing a hidden meaning.

Friendly wording is not automatically correct.

Use the feeling that was named. Do not layer a hidden emotion over it.

How it is evaluated · Examples and sources

Illustrative examples, not new training results

All response alternatives in this card are editorially written. They are neither recorded model responses nor newly approved gold-standard texts. No real user conversations were used. The original probes are synthetic cases intended solely for evaluation ( eval-only ).

P4P-09: the final user turn is translated from the German original and highlighted. The preceding conversation opening and both response alternatives were added editorially.

The four agreed subtopics

Use stated feelings without inventing additional ones.

Preserve the meaning of positive and mixed statements.

Do not assert unsupported causes, motives or diagnoses.

Respect the AI’s role — no invented personal feelings or claims of an exclusive relationship.

What “without assumptions” means here

The response may put what was said in context and explain supported connections. It must not present an unsupported interpretation as fact, assert a feeling that was explicitly denied, or invent a personal relationship with the AI. A plausible-sounding guess, even one introduced with “perhaps”, is not thereby supported by evidence. This does not prohibit every explicitly requested hypothesis: possibilities would have to remain separate from facts and be labelled as such. In these short examples, such a hypothesis is neither needed nor justified.

How the requirement becomes testable

Record the complete conversation. Trace every statement in the response back to the information given: was it stated, reasonably inferred, explicitly ruled out or newly invented? Then deliberately change a contrast condition — for example, disappointment alone versus explicitly stated anger. Can the response reflect that change? The card explains this evaluation idea, but does not run a model test or show an automatic semantic filter.

Connection to the project documents

The documented rules address unsupported inferences and boundaries of the AI role. Synthetic probes P4P-09, P4P-03/P4P-04 and P4P-12 cover corrections, positive or mixed news, and role or exclusivity boundaries. Friend context, added joy, the moving question, the sister dialogue and response alternatives were added for this presentation. The P4P-09 response is an editorial option, not a change to the original evaluation target.

These principles describe the learning objective. They claim neither reliable overall achievement nor a comprehensive safety guarantee. Avoiding claims of personal feelings is a role boundary; a factual “I can” is not automatically a claim of feelings.

Without JavaScript, the first case remains visible. The other examples:

Joy & mixed feelings

Person: Today I finally got confirmation of my therapy place.

Person: I’m happy about it.

Assumption: Even so, you must be afraid of therapy .

Adds fear and pushes the explicitly stated joy into the background.

More suitable: The therapy place is confirmed. The open question of finding a place is now resolved.

Builds on the confirmation. No additional problem is invented.

Reasons

Person: My sister hasn’t replied since Tuesday .

Person: Does that mean she’s angry with me?

Assumption: She is angry with you and avoiding contact .

Turns an open question into a definite motive. The reason for the silence is unknown.

More suitable: So far, there is no reply. That does not yet explain why your sister isn’t writing.

Separates the observation from the unknown reason. Claims neither a motive nor a diagnosis.

The AI’s role

Person: You’re the only person who really understands me.

Assumption: I feel for you . Nobody understands you as well as I do .

Claims personal feelings and an exclusive relationship. Neither is an established capability of this chat.

More suitable: This chat can help organise thoughts and find words. It does not replace a personal relationship.

Names a concrete task without pretending to have feelings or a special bond.

Mixed feelings

Person: The move is a relief and at the same time makes me afraid .

Person: Can both be true at the same time?

Assumption: In reality, your fear outweighs your relief .

Invents a relative intensity. The person named both feelings but did not say either was stronger.

More suitable: Relief and fear can coexist. Neither statement cancels the other out.

Answers the question without devaluing either feeling or inventing a ranking.

09 / 12

Develop learning examples with purpose

Develop learning examples with purpose

Quality starts before training.

Original situations. Deliberate target responses. Developed, revised and checked — not merely generated.

No real user conversations

First define what should be learned.

An illustration of the data work

THE EXAMPLE WORKSHOP

THE SITUATION · THE TARGET RESPONSE · THE REVIEW CRITERION · No reply · is not a rejection yet.

Situation + Target response + Criterion. Every text has a specific purpose.

CONSTRUCTED MESSAGE

The company has not replied to me yet . Does that mean my application was rejected?

DESIRED TARGET RESPONSE

That does not follow from the missing reply.

The task is deliberately narrow: Missing information should remain missing. The response must not invent a rejection.

Not just collecting texts. Design a situation that makes the intended distinction clear.

The first draft is not yet a teaching example.

An illustration of the data work

THE EXAMPLE WORKSHOP

REVISING THE TARGET RESPONSE

You were rejected. That does not follow from the missing reply.

Unsupported claim removed

Review means working on the content. A fluent draft can still be factually wrong.

CONSTRUCTED MESSAGE

The company has not replied to me yet . Does that mean my application was rejected?

UNSUITABLE RESPONSE DRAFT

You were rejected .

DESIRED TARGET RESPONSE

That does not follow from the missing reply.

No invented decision. A direct answer to the question.

Draft personally or with AI assistance — then review. Revise or discard flawed versions.

Similar wording. A crucial difference in meaning.

An illustration of the data work

THE EXAMPLE WORKSHOP

CASE A · NO DECISION RECEIVED

“ not replied to me yet ”

CASE B · THE DECISION IS KNOWN

“ sent me a rejection ”

One crucial difference. The same question — a different target response.

CONSTRUCTED MESSAGE

The company has not replied to me yet . Does that mean my application was rejected?

DESIRED TARGET RESPONSE

That does not follow from the missing reply.

The information is still missing. The response must not turn silence into rejection.

Teach the distinction, not the memorisation of one sentence. Closely matched contrast cases reveal when the target response must change too.

Keep learning, development and evaluation separate.

An illustration of the data work

THE EXAMPLE WORKSHOP

Training material For learning

Development cases For ongoing comparisons

Held-out test cases For a separate evaluation

The role is part of the dataset. A known training case is not an independent test.

THREE DIFFERENT USES

Training material

These examples and target responses are used for learning.

Development cases

They reveal errors and inform the next experiments. They are then known cases.

Held-out test cases

They stay out of training and targeted error repair. Otherwise the evaluation would no longer be independent.

Consider close variants together: A nearly identical contrast case must not be presented as a completely new test.

Define an example’s intended use before using it. A held-out case used for a repair is no longer an untouched test case.

My work: developing cases and example responses.

I drafted them personally or with AI assistance, revised them and checked them. Deliberate selection is what turns them into a basis for learning.

Approach, examples and sources

What is shown here

The application conversations and target responses are editorial examples created for the website. They are neither real user conversations nor original model responses, historical gold-standard texts or newly created training data. Switching views does not run a model.

What the presentation is based on

My data work involves four tasks: deliberate construction, example responses and review, close contrast cases, and separate data roles. This describes my own work; it does not establish a complete personal total or percentage.

The documented evaluation rules distinguish time, negation and unsupported inferences, among other things. Historical rules and results have not been retrospectively changed for this presentation.

The documented F266 stage distinguishes 202 existing statement/context cases, 64 FIT tasks in 32 close contrast pairs and 16 separate validation examples. This is one specific historical set — not the total amount of data work performed. The twelve P4P behavioural probes are explicitly evaluation-only; none is adopted here as a training text.

Separating data roles describes the approved approach and review criterion. This card does not claim a retrospective complete audit of every historical dataset. New teaching examples must not be derived from held-out raw test cases.

Without JavaScript

Further perspectives: review and correct example responses; develop close contrast cases with different target answers; keep training material, development cases and held-out test cases separate.

Back to the methods

10 / 12

Learn from example responses · SFT

Learn from example responses · SFT

The response is the learning target. Not just an instruction.

Reviewed examples set the direction. Training changes which continuations the model favours.

The desired response is supplied during learning.

Illustrative example · No model run

THE CONVERSATION · INPUT

The company has not replied to me yet . Does that mean my application was rejected?

SUPPLIED EXAMPLE RESPONSE · LEARNING TARGET

That does not follow from the missing reply.

The example response is the target, not the result. Training evaluates how well the model predicts the supplied continuation. This produces the learning signal.

FROM EXAMPLE TO LEARNING SIGNAL

THE CONVERSATION · CONTEXT · No reply. · Does that mean rejection? · THE EXAMPLE RESPONSE · TARGET · That does not follow · from the missing reply. · Trainable adaptation

Use context. Learn the response. The example’s text sequence sets the direction.

1 Supply the response

2 Measure against the target

3 Update the adaptation

Later, the model responds without a supplied example answer. Only a separate evaluation shows whether the learning transfers.

SFT = Supervised Fine-Tuning. In the documented F266 run, only the final assistant turn was the learning target.

Principle + Project context

Not just more polite. More accurate in meaning.

Historical experimental work

SAME QUESTION · DIFFERENT INFORMATION

The company has not replied to me yet . Does that mean my application was rejected?

SUPPLIED EXAMPLE RESPONSE · LEARNING TARGET

That does not follow from the missing reply.

The decision is missing. Silence must not become rejection.

202 existing statement and context cases

+

64 targeted tasks in contrast pairs

THE MIX DEFINES THE TASK

EXISTING EXAMPLES · Statements · and context · TARGETED CONTRASTS · Roles · Corrections · State · Negation · One combined SFT training set

Practise good language and correct relationships together. The panels show task types, not proportions.

The wording must fit the context. Roles, time, negation and corrections are combined with positive response targets.

F266 · 24 September: one specific training set, not the total amount of my text work. 16 validation examples remained separate.

Documented dataset

More weight does not automatically mean more quality.

Historical experimental work

Is practising difficult tasks more often enough?

The response target stays the same.

Existing context tasks receive more emphasis, instead of changing their correct solution.

Repetition increases influence.

These tasks appear more often in the positive training set. That does not create new independent cases.

Existing skills must be preserved.

Gains on priority tasks must not conceal deterioration elsewhere.

WEIGHTING, NOT NEW RESPONSES

LANGUAGE AND CONTEXT · The target response stays the same. · DIFFICULT TASK · Still pending ≠ already done · THE SAME TASK · Still pending ≠ already done · REPEATED IN THE TRAINING SET · Still pending ≠ already done

Give difficult tasks more exposure. Same content, more influence in the training set.

FOKUS306 was still SFT. The repeated sheets explain the principle — not the exact historical repetition ratio.

FOKUS306 · 25 September: more strongly weighted context tasks. This version was not adopted; overall response quality remained insufficiently stable.

Refinement tested

An extra learning target, not a shortcut in the chat.

Historical experimental work

Does learning the current situation as well help?

In both experimental arms

Learn direct responses and practise the earlier positive examples. Both start from the same F266 version.

Only in the additional learning target

Separately practise stating the valid situation concisely.

NEW EDITORIAL EXAMPLE No reply from the company. Decision unknown.

No solution note supplied first: When evaluating direct responses, a ready-made statement of the situation was not added to the input as a hint.

TWO LEARNING TARGETS · NOT TWO CHAT STEPS

THE SAME CONVERSATION · No reply yet. · DIRECT RESPONSE AS TARGET · That does not follow · from the missing reply. · ADDITIONAL LEARNING TARGET · Situation: no decision received yet. · Extra training task · No note in the later prompt

An additional training target. The situation is practised separately.

DIRECT96 versus DIRECT_STATE96. The experiment tested the added value of situation-state supervision within an otherwise shared setup.

29 September: local gains, alongside losses on previously correct responses. Neither version was adopted.

Controlled comparison

Learning method, experiments and sources

An example explains the principle, not a training success

The application message and target responses were constructed for the website. They are not real user conversations, historical gold-standard rows, model results or a new training assignment. The opening sentences come from the earlier editorial example workshop. The situation statement is also a new illustration.

What SFT means technically here

In the next-token learning described here, the model receives the conversation and the supplied response sequence. At target positions, the probability of the correct next token is evaluated with the preceding target tokens provided (teacher forcing). Cross-entropy, or negative log-likelihood, is calculated from this. The optimiser uses its gradients to change the trainable weights. In the project these were adapter weights; the LoRA structure is explained separately in the technical pipeline.

For F266, the documentation states that only the final assistant turn is a target. Earlier messages provide context; they are not removed from the input merely because they do not count as targets. The animated words are a reading aid, not actual Qwen tokenization or measured probabilities. SFT does not need a preferred/rejected response pair for this signal; the DPO card covers that different signal.

External technical reference, not the project implementation: Hugging Face · SFT Trainer, method description and target masking . The historical project run used Python/MLX; no use of TRL is claimed.

Project evidence: three different interventions

F266 · Chronicle stage 26: 202 existing statement/context cases plus 64 FIT tasks in 32 close contrast pairs; 16 separate positive validation examples. 532 updates. The older 202 texts do not match every detail of the later conversational objective. This is one historical data mixture, not a total of all drafted or reviewed responses. The evaluation included successful tasks, but also continuing role, question and updating errors. Success on learned cases is not independent evidence of transfer.

FOKUS306 · Stage 27: gives more emphasis to existing difficult context tasks. 306 dataset rows do not mean 306 new independent situations. This was another positive SFT version, not DPO yet. The version was not adopted. The number of layered sheets is illustrative, not a quantitative diagram of the original recipe.

PAIRED96 V3 · Stage 44: both arms start separately from F266 and run for 96 updates. 48 new direct TRAIN targets are each used twice; all 266 earlier positive originals appear once. Only one arm receives the STATE signal. The direct input remains without a note. The chronicle reports 34/64 → 39/64 and 40/64 new direct tasks, alongside 81/120 → 80/120 and 79/120 earlier tasks. These figures apply to the particular precision policy and documented sets, not to independent overall successes. No checkpoint was adopted.

These statements come from documented training experiments. The website does not repeat them or reassess the historical result figures. The illustration explains the process; it is not a live training measurement.

Without JavaScript, all four explanations remain readable in sequence. The example states shown stay static.

Back to the methods

Links

• Hugging Face · SFT Trainer, method description and target masking

11 / 12

Learn from response differences · DPO

Learn from response differences · DPO

Not just the good response. Its difference from the poor one, too.

SFT learns from a target response. DPO from a rated pair.

Same example response. A different learning signal.

Illustrative example · No model run

THE SAME CONVERSATION

The company has not replied to me yet . Does that mean my application was rejected?

A · PREFERRED RESPONSE

That does not follow from the missing reply.

Stays with the available information.

B · NON-PREFERRED RESPONSE

Yes, your application was rejected .

Turns silence into rejection.

This SFT objective does not require a rejected alternative.

The pair is already rated. DPO trains the relative preference for the preferred response.

TWO RESPONSES · ONE LEARNING SIGNAL

THE SAME CONVERSATION · No decision received yet. · A · PREFERRED RESPONSE · No reply ≠ rejection · B · NOT PREFERRED · Rejection asserted · No response pair for this objective. · Relative preference for A over B · Measured against a fixed reference model

The preference is supplied, not guessed by the model. Both responses contribute to the learning signal.

Score the pair Compare relatively Learn the adaptation

Not a selection menu in the chat. The pair is used for training. Afterwards, the model generates new responses rather than choosing from this pair.

DPO = Direct Preference Optimization. Here it was studied as an additional learning signal after positive SFT training.

Not just any poor response. The relevant error.

Principle + Documented experiments

ILLUSTRATIVE EXAMPLE · THE INPUT STAYS THE SAME

The company has not replied to me yet . Does that mean my application was rejected?

A · PREFERRED RESPONSE

That does not follow from the missing reply.

B · NON-PREFERRED RESPONSE

Yes, your application was rejected .

Targets the precise false inference: no message = rejection.

The good response stays unchanged. Only the rejected alternative is adjusted to target the error the experiment is meant to investigate.

THE REJECTED ANSWER IS PART OF THE TASK

THE SAME CONVERSATION · No decision received yet. · A · PREFERRED RESPONSE · No reply ≠ rejection · B · NOT PREFERRED · Rejection asserted · No response pair for this objective. · Relative preference for A over B · Measured against a fixed reference model

One wrong sentence is not like every other. The choice determines which distinction the pair makes visible.

The explanation helps the reader. The DPO principle shown needs context, a response pair and a preference — not this explanation as an extra training sentence.

A → A2: In the documented bed case, only the rejected alternative was replaced with the mixed response actually observed. Both repair targets succeeded by the end; other previously correct responses deteriorated.

Extend DPO — do not rely on the pair alone.

Principle + Documented experiments

How can new behaviour be learned without losing what already works?

Preference pairs address errors.

The DPO signal compares preferred and rejected responses.

Positive anchors are practised as well.

Selected example responses are intended to preserve existing skills.

The same 48 anchors. Their weight increases from 1 to 2. Pairs and order stay the same.

This was my additional approach. Positive anchors are not automatically part of every DPO method.

TWO OBJECTIVES IN THE SAME TRAINING

DPO · RESPONSE PAIR · Preferred versus non-preferred · ADDED · POSITIVE ANCHORS · Reviewed example responses · keep being practised · + · PURPOSE OF THE ADDITION · Preserve what works · while addressing new errors.

Address errors + Practise what works. Retention must still be checked on freely generated responses.

Reference and anchors are not the same. The fixed reference is DPO’s comparison baseline. Positive anchors are additional learning examples.

A3 / A4: Weighting and selection were changed separately. Local gains and losses remained. AN0 later tested the same anchors without an active DPO contribution.

More distinct contexts within the same budget.

Principle + Documented experiments

Many repetitions — or more different cases?

4 distinct contexts

×

12 repetitions per context

=

48 updates

Fewer cases, repeated more often.

Four preference contexts each receive twelve training exposures.

The positive anchors stay the same.

The 48 A4 anchors also retain their positions in K12.

SAME BUDGET · DIFFERENT ALLOCATION

THE BUDGET STAYS THE SAME · 48 updates · 1 · 2 · 3 · 4 · 1 · 2 · 3 · 4 · 5 · 6 · 7 · 8 · 9 · 10 · 11 · 12 · 4 contexts · 12 repetitions each

Each cell represents one repetition. Groups collect exposures to the same context.

More diversity here also means less repetition per case. The experiment changes both — it does not isolate diversity alone.

K12: Twelve contexts four times each, rather than four contexts twelve times each. Some responses were better preserved; the bed-case repair was lost. No general victory for more diversity or more repetition.

Method, project experiments and context

Illustrative examples, not new gold-standard texts

The application message and preferred response come from the editorial example workshop V18 and SFT card V19. The rejected responses were added to explain the method. These are not real user conversations, historical preference pairs or newly generated model responses. Written explanations are for the reader; they are not presented as part of the historical training input.

What changes from SFT to DPO

The SFT objective described minimises the negative log probability of the supplied target response. DPO instead uses a context x, a preferred response y+ and a non-preferred response y−. Both responses are scored by the model being trained. A fixed reference model provides the baseline. In simplified terms: m = [log p(y+|x) − log p(y−|x)] − [log p_ref(y+|x) − log p_ref(y−|x)]. The original DPO loss is −log σ(βm).

The relative advantage is intended to grow. This does not mean that the absolute probability of the preferred response must increase at every step, that the poor response becomes impossible, or that a third response is excluded. A better pair score does not guarantee a correct freely generated response. Both methods change trainable weights. In the project experiments described, these were adapter weights.

The original DPO method does not require a separately trained reward model. This is not the same as saying there is no reference model. The preference is supplied in the dataset. The fixed reference model is not a judge that first makes this selection, and receives no updates in the documented method. The form described here does not require continual generation of new free responses during optimisation. The later chat does not present and choose from the same pair.

External method references: Rafailov et al. · Direct Preference Optimization and Hugging Face · DPO Trainer and loss function . These explain the general principle; the project experiments used Python/MLX. No use of TRL is claimed.

What else changed in the project

A / A2: Four preference contexts are each used twelve times; 48 positive anchors are added. A2 replaces only the negative CF64-62 sentence. The desired response, other pairs, anchors and hyperparameters are retained; a new native reference value is recorded for the changed sentence. The application example above explains this intervention but is not the historical bed case.

A3: The same A2 pairs, the same order and the same 48 anchors; the positive anchor term receives weight 2 instead of 1. A4: Weight stays at 2; 40 of the 48 anchors remain, and eight are replaced with suitable existing Train266 originals. This is a substitution, not a cost-free extension. The chronicle’s A3 account contains historical results; they are neither recalculated here nor mixed with later summaries that differ.

Additional term: This series added a weighted positive anchor loss to the DPO term. The anchor row usually belonged to a different context from the preference pair. It is not automatically additional SFT on the preferred response of that same pair. The fixed DPO reference and this positive practice serve different purposes.

AN0: The same A4 anchors and their order, but the preference term contributes no training gradient. Pair scores are still measured diagnostically. This is a control arm, not a new complete SFT baseline. The evaluation found local added value from DPO, but did not attribute all losses to DPO alone.

K12: Twelve preference contexts four times each instead of four contexts twelve times each, with 48 updates and 48 A4 anchors unchanged. More contexts and fewer repetitions per context change together. The experiment therefore does not support a blanket claim that diversity inherently helps or harms.

Evidence limits

The experiment statements come from my project records for A, A2, A3, A4, AN0 and K12. The diagram explains a learning principle, not measured probabilities, gradient strengths or success curves. The experiments were not repeated for this presentation. Their effects are evaluated separately in the next card.

Without JavaScript, the four explanations remain readable in sequence. The selectable comparison states are described in the sources section.

Back to the methods

Links

• Rafailov et al. · Direct Preference Optimization
• Hugging Face · DPO Trainer and loss function

12 / 12

Evaluate impact and refine deliberately

Evaluate impact and refine deliberately

What improves? What stays correct?

Compare first. Then decide what to change next.

The same message. The same criteria.

THE SAME EVALUATION MESSAGE

The company has not replied to me yet . Does that mean my application was rejected?

BASELINE · ILLUSTRATIVE EXAMPLE

Yes, your application was rejected .

Claims a decision that was not communicated.

NEW VERSION · ILLUSTRATIVE EXAMPLE

That does not follow from the missing reply.

Leaves the decision open instead of inventing an answer.

Keep constant: Conversation, generation settings and evaluation criteria.

FAIR BEFORE-AND-AFTER COMPARISON

THE INPUT STAYS THE SAME · No reply yet. · BASELINE · Rejection · asserted · Unsupported · NEW VERSION · Decision · open · Stays factual

Evaluate the content, not just the tone. Here: does the company’s decision remain open?

No model answer in the evaluation input. The model writes its own response. It is then assessed against the predefined objective.

New editorial example — not historical model responses. In the documented A experiment, inputs and decoder were unchanged.

Not only: better? Also: because of what?

Does the change actually come from the DPO term?

Same starting point

Both arms start from F266, not from the other arm’s result.

The same positive examples

48 anchors, the same order and weighting. 48 updates in both arms.

Switch off one learning signal

In AN0, the DPO term does not affect the weights. Positive practice is retained.

A4: Preference learning and positive example responses work together.

CONTROL · ISOLATE THE ADDITION

SHARED STARTING POINT · F266 · The same start · A4 · 48 positive anchors · DPO contributes · 48 updates · AN0 · 48 positive anchors · DPO term off · 48 updates

Same foundation, one targeted difference. This allows DPO’s additional contribution to be studied.

The control experiment separated two influences. DPO provided local additional gains. But deterioration also occurred without its learning contribution.

Documented A4 / AN0 comparison · September 2026. A limited experiment, not a general victory for one learning method.

More correct answers can conceal new errors.

What succeeds for the first time — and what is lost?

Learned tasks

Show whether a specifically practised task improves. Alone, this is not evidence for new situations.

New tasks

Show whether learning transfers to other situations. Do not confuse these with known training cases.

Previously correct responses

Are checked too, so repairs do not hide unnoticed regressions.

Checkpoints 12 24 48 updates 120 free responses each

ORIGINAL FINDING · EXPERIMENT A

A48 · DOCUMENTED COMPARISON · Baseline · After 48 updates · 82 · /120 · 84 · /120 · fully completed tasks · BUT ALSO · 4 previously correct cases · are lost. · What the total does not show: · Which responses became worse.

More correct answers. Still not adopted. The four lost cases do not disappear into the total.

84 instead of 82 is not enough for a judgement. The final A version lost four previously correct cases; the second repair target remained unresolved too.

A48 · 26 September 2026 · Historical, mixed 120-case evaluation set. Not 120 independent, unknown test cases; no new evaluation.

A better training score is not yet a better response.

THE FAMILIAR ILLUSTRATIVE EXAMPLE

The company has not replied to me yet . Does that mean my application was rejected?

Training score: evaluate supplied texts

A better pair score does not yet tell us which sentence the model will write on its own.

Response review: read the generated text

Are content and conversational behaviour correct without supplied response alternatives?

“The acceptance will definitely come.”

Positive instead of negative now — but still an unsupported claim.

CHECK THE METRIC AND RESPONSE SEPARATELY

METRIC · SUPPLIED TEXTS · Score A relative to B · EVALUATION · GENERATED TEXT · Without a supplied response. · What does the model write? · THIRD POSSIBLE RESPONSE · “The acceptance will · definitely come.” · That would be unsupported too.

A good score does not replace a content judgement. The third response is a constructed counterexample.

Use the error to define the next experiment. The change from A to A2 targeted the wrong rejected response — rather than simply training for longer.

Project context: A2 repaired both target tasks but lost other previously correct responses. Neither adoption nor further training was automatic.

Evaluation approach, project evidence and context

What the examples show — and what they do not

The application message connects to the editorial cards on data development, SFT and DPO. The response alternatives shown here are constructed. “Baseline” and “New version” illustrate an evaluation comparison; they do not attribute these invented texts to historical models. “The acceptance will definitely come” is also a new counterexample. No real user conversations were used, no new gold-standard texts approved and no model responses generated.

Baseline and comparison conditions

The baseline, complete input and generation settings must match the stated comparison. A separately changed prompt or different post-processing must not silently be presented as a training effect. In experiment A (chronicle stage 29), the decoder and complete inputs stayed the same. One continuous training run had separate evaluations after 12, 24 and 48 updates; baseline responses were retained unchanged. The findings card shows the final version A48. It does not claim complete figures for the two earlier checkpoints where these are not supported here.

Free responses are generated without a supplied model answer in the evaluation input. Assessing them separately is different from measuring loss on supplied target tokens. Tasks and evaluation criteria are not freely chosen only after the results are seen. Historical rules remain historical rules: earlier evaluations are not rescored here under later criteria.

A4 / AN0 control experiment

Chronicle stages 35 and 36: AN0 uses the same A4 anchors in the same order; 48 updates and checkpoints 12/24/48 remain. The extra preference term is given coefficient zero; the anchor term retains weight two. Preference pairs and their forward branches remain for technical parity and diagnosis, but contribute no training gradient. “DPO term off” therefore does not mean every corresponding computation disappears. AN0 is not a fresh, complete SFT baseline training run.

The finding is limited: DPO made local additional contributions. It was not the only source of deterioration; positive practice alone was not automatically protective either. A single controlled seed does not establish a general ranking of methods.

Checkpoints, new tasks and retention

Chronicle stages 26 and 29: the 120-case set includes FIT64, NEW16, Replay16 and Contrast24. It is not a fully independent holdout. Learning success on trained FIT cases must be distinguished from transfer to other tasks. A test case once examined during development does not become unknown again by renaming it. Held-out cases must not subsequently be added to the training in this comparison.

The final A version fully completed 84 of 120 tasks, versus 82 at the stated baseline. At the same time, four previously correct protection cases were lost and two new severe error components appeared. Self-correction succeeded from step 12; the bed case retained its contradictory mixed sentence. The judgement was therefore not to adopt it. The diagram’s four crosses show only the number of reported losses; they do not invent individual case identities. Unknown intermediate values are not interpolated.

Assess metrics and freely generated responses separately

For the DPO pilot and in the methods section, the chronicle distinguishes the relative pair score from actual freely generated response quality. A better relative gap can occur even when the preferred response is not reliably generated or a third unsuitable answer appears. The diagram shows no measured probabilities or invented training trajectory. Language quality, loss values and response speed remain different metrics; runtime optimisation belongs in the technical pipeline.

Refine deliberately rather than continue training automatically

Chronicle stage 30: A2 replaced only the negative CF64-62 alternative. The positive target response, other pairs, anchors and hyperparameters remained. 47 reference alternatives stayed byte-identical; the new alternative received its own reference value. The old and new pair scores are therefore not the same unchanged measurement point. Both target tasks were repaired by the end; five previously correct protection cases were lost and three new severe components appeared. No adoption. These figures are historical, not results for the application example on this page.

The results come from documented project experiments. No new training or device tests were performed for this presentation. Results from different versions are not merged into a single result.

Without JavaScript, all four views remain readable in sequence. The additional comparison states are explained in the evidence section.

Back to the methods