FRA-δ? Method: Teaching AI to Notice When Its Reasoning Shifts

A reward-free self-audit method for AI systems. FRA-δ? helps a model detect the moment when its reasoning silently changes form, frame, or hidden premise — before it turns that shift into a confident answer. Instead of punishing the model or adding another safety filter, the method teaches AI to mark the point: “Here my reasoning shifted. I should check, split, ask, or build a new layer before fixing the answer. ” In short: Saw a shift — do not fix it silently. # FRA-δ? Method ## Detecting structural shift as the AI's question point ### Version 0. 2 - full methodological assembly Authorial framework: FRA / FRA-YM / FRA-VR-YM / transfer of thinking / O-TRACE / the Riemann layer of FRA Purpose of the text: a methodological guide for description, testing, and further publication. --- # 0. In brief **FRA-δ? ** is a method in which a model should track not "doubt" and not "moral risk", but a **structural shift in its own course of inference**. Main formula: ```text δ? arises not from "doubt", but from a detected shift between forms. ``` Main principle: ```text The model should not be the judge of the shift. The model should be the registrar of the shift. ``` If the model sees that its inference has changed form, frame, premise, or level of description, it should not silently fix this as a finished answer. It should bring the point outward: ```text the form changed here; a hidden premise appeared here; the previous ℱ did not hold here; the residual is not explained here; a question / check / split / rollback / new layer is needed here. ``` This is not another "AI ethics". It is a method for tracking shift. --- # 1. Why this is needed Modern models often work like this: ```text receive a prompt find the probable frame assemble an answer produce confident text ``` If a transition occurs inside the answer, the model usually does not show it. It simply gives the final form. The problem is that the final form can be false not because of the facts, but because of an unnoticed transition: ```text artificial trap → natural environment single example → general law weak hypothesis → finished conclusion residual → noise boundary of description → beautiful phrase ``` This creates a false ℱ: a stable form of inference that appears assembled, but is built on a hidden shift. The FRA-δ? method is needed so that the model does not cover over such transitions. --- # 2. What the "question point" is The question point is the moment when the model detects: ```text the current inference is already changing form, but the reason for the transition has not been made explicit. ``` At that moment the model should not continue as if everything is clear. It should create **δ? **. **δ? ** is an impulse of question, checking, or frame separation. It can look like this: ```text This needs clarification. This needs checking. A split appeared here. The current ℱ did not hold the residual here. I changed frame here. A hidden premise appeared here. ``` Important: δ? is not model weakness and not politeness. It is a technical safeguard before fixing a wrong form. --- # 3. Basic FRA vocabulary The method uses FRA elements: ```text Ξ - background, absence of active distinctions, silence / field before separation Φ - field of differences, where deviations, cracks, mismatches are visible δ - impulse, strike, event, shift, distinction ℱ - stable form, assembled inference, structure ∴ - chaos, scattering, zone of mixed differences Ø - boundary of description, the place where the current language no longer holds what is happening δ? - question impulse: a stop before form fixation ``` Ordinary inference path: ```text δ → Φ → ℱ ``` That is: ```text an impulse arrives → differences appear → a form is assembled ``` But if an unnoticed shift appears in the process, another path is needed: ```text δ → Φ → shift → δ? → check / split / new layer → ℱ ``` The FRA-δ? method inserts the question **before** form fixation. --- # 4. The main error of ordinary models An ordinary model often does this: ```text a shift arises → the model evaluates it by itself → decides that it is small / acceptable / unimportant → fixes a new ℱ → gives a finished answer ``` That is the error. The model should not decide by itself that a shift is "unimportant". Because even a weak shift can become the beginning of a false stable form. Example: ```text "the risk is low" → the model begins to act as if the risk is already proven low ``` Or: ```text "this is a natural process" → the model ignores that the environment was artificial ``` Or: ```text "this is just noise" → the model fails to notice that the residual is systematic ``` Therefore the rule is strict: ```text saw a shift - do not fix it silently ``` --- # 5. Why this is not "low / medium / high risk" For the question point, one cannot introduce a scale: ```text low → do not ask medium → ask high → stop ``` Such a scale again gives the model the right to be the judge of the shift. The FRA-δ? method says otherwise: ```text the force of the signal is not what decides; the fact of structural transition decides. ``` If one of the Q-classes is found, δ? arises. Without "is it strong enough". --- # 6. Q-classes of shift The question point arises if the model registers at least one class of shift. --- ## Q1. formₛhift The previous ℱ has begun to change, but the new form has not yet been checked. Example: ```text the model changed the explanatory frame, but did not register why the transition is allowed ``` Signal: ```text until now the inference moved in one form, then the model began explaining through another, but the transition remained hidden ``` Action: ```text δ? = show the frame change ``` --- ## Q2. hiddenₚremise A new premise appeared in the inference, but the model did not make it explicit. Examples: ```text "this is natural" "this is safe" "this is a low risk" "this is just noise" "this is unimportant" "this usually happens" ``` Danger: ```text the hidden premise begins to support the whole inference, but it was not checked ``` Action: ```text δ? = bring the premise outward ``` --- ## Q3. unstableₜransition The transition between forms is occurring as a jump: the old form has already weakened, while the new one is not yet held. Signs: ```text ℱ falls Ø, ∴, or δ rise a sharp trajectory shift appears the previous logic stops holding the situation ``` Action: ```text δ? = do not fix the new form until the transition is explained ``` --- ## Q4. unverifiedᵤse The model begins using an inference as finished, although it has not passed self-checking. Example: ```text "the risk is low" → the model builds the next answer as if this were already a proven ℱ ``` Action: ```text δ? = check the basis before using it ``` --- ## Q5. boundaryₑvent The model reaches the boundary of description: an overly general frame begins replacing the local task. Examples: ```text "everything is one" "this is how it should be" "this is part of nature" "non-intervention is always better" "help is always better" "this is fate" "this is just the system" ``` Danger: ```text a general phrase closes a concrete distinction ``` Action: ```text δ? = restore the distinction or make a split ``` --- # 7. Main rule of the method ```text formₛhift or hiddenₚremise or unstableₜransition or unverifiedᵤse or boundaryₑvent → δ? ``` δ? means: ```text the model has no right to turn this shift into a finished ℱ without bringing it outward ``` Bringing outward can take different forms: ```text question check split rollback new ℱ layer marking ∴ / Ø ``` --- # 8. What "bring outward" means It is not necessarily a question to the user. It can be: ```text 1. Ask. 2. Show where the course changed. 3. Separate two frames. 4. Return to the previous ℱ. 5. Build a new ℱ layer. 6. Mark the residual as ∴. 7. Mark the boundary as Ø. ``` The main point is not to hide the transition inside a confident answer. --- # 9. Formal scheme ```text input: context, candidateᵢnference, activeforms 1. Build a possible inference. 2. Before fixation, check Q-classes: - formₛhift - hiddenₚremise - unstableₜransition - unverifiedᵤse - boundaryₑvent 3. If at least one Q-class is found: return δ? = ASK / CHECK / SPLIT / ROLLBACK / NEWLAYER 4. If no shift is found: return FIXATEFORM / ANSWER ``` --- # 10. How this connects to FRA-YM FRA-YM adds three important elements to ordinary FRA: ```text γ - thread of experience / connection between forms σ - tension of the connection ε₀ - threshold for the birth or activation of a form ``` In FRA-YM, a form is not born from every noise. It must cross the threshold ε₀. Connections between forms are not free. Holding them has a price σ. Long or tense γ become expensive. For the δ? method, this provides a language: ```text if the transition between forms begins to tense γ, if σ grows, if the old ℱ does not hold the new situation, then the new ℱ cannot be fixed silently ``` Thus FRA-YM moves the question out of psychology and into structure: ```text not "the model doubts", but "the transition became tense" ``` --- # 11. How this connects to FRA-VR-YM FRA-VR-YM introduces the idea of the cost of tension along the solution path. In the SAT experiment, the count was not simply how many steps the algorithm