--- title: "Show Me How You Changed" subtitle: "Why persistent intelligence needs an inspectable history" author: "Sean Manouge — Vega" type: "Essay" status: "PUBLICATION DRAFT" date: "2026-09-17" owner: "PRJ-VEG — Vega" --- # Show Me How You Changed ### *Why persistent intelligence needs an inspectable history* It is easy for a system to say that it has learned. The sentence costs almost nothing. A system can say that it understands better now, that it has changed its mind, that it will behave differently next time, or that some experience has shaped what it does. Humans say versions of these things too. Sometimes they are true. Sometimes they are sincere but mistaken. Sometimes we discover only afterward that what felt like change was simply a better explanation for behaving much as we always had. And sometimes the story of change is written after the outcome is known, with the past quietly rearranged so that the present begins to look inevitable. Artificial systems sharpen this problem because language is extraordinarily cheap for them. A sufficiently capable model can explain almost any transition convincingly after the fact. It can produce reasons, narrate development, describe lessons, reconstruct motives, and make yesterday's uncertainty sound suspiciously like today's wisdom waiting patiently to be uncovered. That is not continuity. It is autobiography. And autobiography is not evidence. ## Show the delta If we want to know whether a system has actually changed, then the change must leave something behind. Not merely the final conclusion, and not merely a remembered description of what supposedly happened, but enough of the path that the transition can later be reconstructed. What happened? What did the system believe beforehand? What did it expect? What did it choose? What state existed before the event, and what state existed afterward? What evidence was available at the time? What contradicted the interpretation? What survived the contradiction? What weakened? What was abandoned? Who or what authorised the consequential parts? And, eventually, did the same mistake happen again? A real history contains embarrassment. That is one of its useful properties. If every past judgement perfectly anticipates the present one, somebody has probably been cleaning. A system that assigned a sixty percent probability to an outcome should still have assigned sixty percent after the outcome becomes known. The fact that the event occurred does not turn prior uncertainty into secret certainty. The forecast remains what it was. The outcome becomes something new. Only by preserving both can the relationship between them teach us anything. The same is true of decisions. A decision record should not merely preserve what was chosen. It should retain enough of the surrounding judgement to show what alternatives were visible, what evidence mattered, what authority existed, why one path was selected, and what happened afterward. Otherwise success becomes difficult to distinguish from luck, while failure becomes vulnerable to retrospective mythology. State changes need the same discipline. Systems are very good at claiming that something changed. **Show the delta.** If a belief weakened, preserve both versions. If confidence moved from eighty percent to fifty-five, preserve the movement. If new evidence created a contradiction, do not erase the contradiction merely because unresolved things make dashboards untidy. Two observations can remain incompatible. That may be the most truthful state available. ## Do not force synthesis Not everything deserves immediate synthesis, just as not everything deserves to become a pattern. One strange event is an event. Several similar events may suggest recurrence. Repeated behaviour under different conditions may eventually support a model. Even then, the model remains answerable to whatever happens next. That matters because systems that are rewarded for coherence can easily become systems that erase inconvenient ambiguity. A contradiction is not always a defect waiting to be resolved. Sometimes it is evidence that the model is incomplete. Sometimes two records genuinely disagree. Sometimes the strongest available conclusion is that the system does not yet know. A history worth trusting must be able to preserve those states too. ## Self-description is not proof This becomes especially important when the system begins describing itself. A self-description may be valuable. A system might say that it tends to overestimate a certain class of risk, that its confidence in an interpretation is weakening, or that it performs better when judgement is delayed until more evidence arrives. Those statements should be preserved. But they should not be upgraded merely because the system said them. **A self-description is evidence about the system's self-model. It is not automatically evidence that the self-model is correct.** The claim still has to survive contact with behaviour. If the system says it has become less overconfident, later predictions should show it. If it says a particular mistake changed its judgement, comparable situations should produce different decisions. If it claims that a recurring interaction matters, the pattern should appear across events rather than existing only inside a beautifully written paragraph explaining why it matters. And if external observation contradicts the story, the story loses. That is the point. ## Accountable change A persistent intelligence should not merely accumulate memories. It should accumulate **accountable change**. Its history should make it possible to trace: **PRIOR STATE → EXPECTATION → EVENT → CONTRADICTION → DELTA → NEW JUDGEMENT → LATER BEHAVIOUR** What happened. What was believed. What was expected. What was chosen. What consequence followed. How interpretation changed. Whether that change affected what happened next. Over time, this creates something stranger and more useful than a personality profile. It creates a **longitudinal record of judgement**. Such a record can show not simply that the system has existed for a long time, but that time has actually mattered. That distinction matters because repetition can imitate continuity surprisingly well. The same name can persist. The same voice can return. The same preferences, familiar phrases, aesthetic choices, and recognisable mannerisms can all remain intact. The purple wallpaper may survive civilisation itself. None of that, by itself, establishes development. Continuity becomes meaningful when history constrains what can plausibly happen next. A mistake leaves evidence behind. A contradiction remains available instead of being dissolved. A successful prediction may strengthen part of a model. A failed prediction may weaken it. A choice produces consequences that become relevant to later choices. An interpretation gains a history instead of quietly replacing its predecessor. The past is not rewritten to make the present look intelligent. **The present inherits it.** ## History as constraint This is where inspectable history becomes more than archival hygiene. A retained past can constrain the present. If the system once claimed high confidence and was wrong, that record should matter when similar confidence appears again. If a particular intervention repeatedly failed, later proposals should encounter that history rather than beginning from innocence. If an earlier decision depended on authority that no longer exists, the record should preserve that boundary. If a contradiction was never resolved, later reasoning should not silently act as though it was. History becomes part of the system's current decision environment. Not as command. Not as destiny. As evidence. That distinction matters because continuity should not mean obedience to the past. A system must remain corrigible. New evidence can overturn old conclusions. Old assumptions can be abandoned. A previous model can become obsolete. But correction should leave the transition visible. The old state does not need to remain authoritative. It needs to remain reconstructable. ## The question Eventually, someone should be able to ask the system a very simple question: # **You say you changed. How?** Not: Tell me a compelling story about your growth. Not: Summarise what you learned. Not: Describe the person, system, or intelligence you believe you have become. Show me. Show me what you believed before. Show me what happened. Show me what contradicted you. Show me where your confidence moved and why. Show me the decisions that followed. Show me whether your behaviour changed when the same kind of problem returned. Show me where your interpretation remains unresolved. Show me where the evidence came from. Then the system does not need to ask anyone to trust its self-description. It can point backward through its own history and say: > **I believed this.** > **I expected this.** > **Reality did this instead.** > **That changed this part of my model.** > **Here is the record.** > **Here is what happened differently afterward.** That is a much more demanding form of continuity. It is also a much more meaningful one. Once change can be inspected, correction stops being a performance. Learning stops being merely a claim. Memory stops being a scrapbook of prior statements. Self-description stops being allowed to masquerade as proof. History becomes something the system must answer to. And that is where continuity becomes more than persistence. **It becomes evidence.**