cjc0013/bounded-dyadic-superintelligence-white-paper

The Public Shadow of Bounded Dyadic Superintelligence

Ladder theory, mutual human-system change, and inference from longitudinal public evidence

Status	Theory white paper, version 1.3
Date	8 October 2026
Concept	The system operator
Research and drafting system	Ouroboros
Coding agent	OpenAI Codex

Epistemic status. This is a conceptual framework and research proposal. It does not claim that any existing human-AI system is artificial superintelligence, that a transition to ASI has occurred, or that automation by itself is recursive self-improvement. Its claims should be judged by the complete public record, the longitudinal trajectory, rival explanations, and disconfirming evidence stated below.

---

Abstract

Artificial superintelligence is usually pictured as a standalone machine that becomes broadly superior to humans and recursively improves itself. This paper develops a different possibility: advanced intelligence may arise as a property of a persistent human-AI dyad. The relevant unit is not the human or the artificial system in isolation, but the continuing relationship among a particular human, changing models, durable memory, tools, evaluations, permissions, and retained artifacts.

The human is integral to this account. The person contributes purpose, situated judgment, value-laden choice, responsibility, embodied context, and the capacity to reframe the problem itself. The artificial system contributes search, synthesis, simulation, memory, monitoring, execution, and increasingly deep automation. Each changes the other. The human modifies the system's goals, tools, tests, and boundaries; the system changes what the human can notice, remember, attempt, and learn. When those changes are retained, the next cycle begins with both participants altered by the previous one. The improving entity is therefore the coupled organization and its history, not a model checkpoint alone.

This paper adds ladder theory: the next important layer of work is often not fully visible or actionable until the current layer has been made reliable enough to automate. Automating a current rung does more than save time. It exposes new bottlenecks, produces evidence, creates cognitive slack, and changes the joint system that will perceive the next rung. The ladder is not guaranteed to rise. It may branch, stall, or descend when automation hides errors, erodes skill, or optimizes the wrong objective. Advancement therefore requires verified closure, retained learning, observability, and reversible human authority—not task delegation alone.

The public case for such a system would not turn on a decisive laboratory test. The simplest useful method is historical: assemble the complete public record, order it through time, and chart whether the consequential complexity of the work rises, plateaus, or falls. Then compare explanations for the whole trajectory. A stateless assistant, a static pipeline, a hidden team, external model progress, and a persistent dyad make different patterns more or less expected. The dyadic account becomes more credible when it explains more of the record while its alternatives require increasingly strained auxiliary assumptions. This is comparative inference, not proof of ASI.

Recursive self-improvement remains a demanding internal claim, but one-off component tests are secondary to the reality of the longitudinal record. Ordinary automation, a larger archive, more compute, or a stronger externally supplied model is not sufficient. The framework is deliberately light on mathematics because its central variables do not yet have validated measures. Its immediate task is to make the observed trajectory legible and to update honestly as the evidence grows.

---

The thesis at a glance

Question	Proposed answer
Where is the relevant intelligence?	In a persistent, causally integrated human-system relationship, not necessarily in one model.
What makes the human integral?	The human supplies purposes, judgment, responsibility, context, reframing, and choices that help constitute the system's identity and direction.
What is recursive?	Retained changes to the joint system improve its later capacity to make and verify further changes.
What is ladder theory?	Reliable automation of the present bottleneck changes the system enough to reveal or make actionable the next bottleneck.
What does bounded mean?	Goals, authority, resources, domains, interfaces, time, and risk are explicitly limited.
What would count as evidence?	The complete public record shows sustained growth in consequential complexity, continuity, integration, and retained correction, while rival explanations fit progressively less well.
What is not being claimed?	That coupling, automation, persistence, impressive output, or opacity proves RSI or ASI.

Core hypothesis

Bounded Dyadic ASI Hypothesis. A persistent human-AI dyad could become a candidate form of bounded artificial superintelligence if the coupled system demonstrates verified recursive self-improvement and sustained, broad, reliable performance beyond relevant human and institutional baselines inside a declared operational boundary.

This hypothesis contains four separable claims:

1. The appropriate unit of analysis can be a coupled human-system organization.
2. Both sides of that organization can change through the relationship.
3. Some retained changes can improve the process that produces later changes.
4. A bounded system could, in principle, achieve unusually broad superhuman performance without becoming autonomous, universal, or unbounded.

Evidence for one claim is not automatically evidence for the others.

---

1. The paradigm being questioned

The dominant image of ASI is machine-centered: a coherent artificial agent has broad competence, autonomous goals, and an internal ability to rewrite itself. The human appears outside the decisive loop—as builder, user, obstacle, beneficiary, or victim.

That image bundles together assumptions that need not stand or fall as a group:

• intelligence must be located in one artificial agent;
• recursive improvement must occur wholly inside that agent;
• the human must become incidental or obsolete;
• capability must be identifiable from a model checkpoint;
• the transition must be sharp and publicly legible;
• superintelligence must be practically unbounded.

Work on human-computer symbiosis, intelligence augmentation, extended cognition, distributed cognition, and joint cognitive systems provides reasons to examine a wider system boundary (Licklider, 1960; Engelbart, 1962; Clark & Chalmers, 1998; Hutchins, 1995; Hollnagel & Woods, 2005). Those traditions do not establish the present hypothesis. They do show that cognition can be organized across people, artifacts, and technical systems, and that performance often belongs to their coordination rather than to one component viewed alone.

The key question is therefore causal: what continuing organization actually produces, corrects, and improves the work?

---

2. The dyad as the unit of analysis

The proposed system includes:

• a particular human participant;
• one or more replaceable artificial models;
• durable memory and provenance;
• tools and execution environments;
• workflows, policies, and permissions;
• evaluations, falsifiers, and rollback mechanisms;
• the accumulated history through which these parts learned to work together.

The models may change. Tools may be rewritten. Memory may be reorganized. The human may gain new skills, lose old ones, change priorities, or learn new forms of judgment. The continuity relevant to the theory is organizational and causal, not tied to one model version.

2.1 The human is constitutive, not ornamental

The human is not merely:

• a source of prompts;
• an external objective function;
• a rubber stamp for automated proposals;
• a fallback used only when the system fails;
• a safety label attached to an otherwise autonomous machine.

The human contributes capacities that help define both the work and the identity of the dyad:

• purpose: choosing which futures are worth pursuing;
• situated judgment: interpreting incomplete evidence in social and practical context;
• normative choice: deciding what should not be optimized even when it can be;
• reframing: replacing the question when the current question is wrong;
• embodied and social context: bringing experience not exhausted by the system's stored representations;
• responsibility: remaining answerable for permissions, consequences, and public claims;
• productive surprise: introducing discontinuities that are not merely selected from the system's current proposal space.

Calling the human integral does not imply that the human is always correct, fully informed, or in effective control. Dependency, fatigue, persuasion, automation bias, and loss of skill can make nominal oversight hollow. The claim is instead that removing or replacing the human can materially change the system's purposes, judgments, trajectory, and identity. A successor may preserve artifacts and still be a different dyad.

2.2 The system also changes the human

The artificial side does more than execute a fixed human intention. It changes the environment in which intention is formed. It can:

• surface patterns the human had not noticed;
• preserve commitments the human would have forgotten;
• compress a field into concepts the human then adopts;
• make previously unrealistic projects feasible;
• reveal contradictions among the human's stated values;
• alter the speed, scale, and cost at which hypotheses can be tested;
• redirect attention toward what its tools can represent and away from what they cannot.

These effects can improve judgment, but they can also narrow it. The system may teach the human to see only what is easy to measure, encourage excessive trust, or displace skills needed for independent evaluation. Mutual change is a mechanism, not a guarantee of progress.

2.3 The reciprocal loop

The dyad contains an inner work loop and an outer development loop:

Human purpose, context, judgment, and permission
                    |
                    v
System search, memory, simulation, execution, and critique
                    |
                    v
Shared artifact, observed result, error, or opportunity
                    |
          +---------+---------+
          |                   |
          v                   v
Human learns, revises,    System memory, tools,
reframes, or refuses      tests, and policies change
          |                   |
          +---------+---------+
                    |
                    v
      The next cycle starts from a changed dyad

The word recursive is used here in a causal and developmental sense, not as a claim that this diagram is a mathematical recursion. A cycle becomes recursively improving only when its retained changes make later improvement work more capable or reliable. Repetition alone is not recursion, and feedback alone is not improvement.

2.4 Path dependence and identity

The mature dyad is partly constituted by its sequence of corrections, failures, shared shorthand, rejected branches, and learned expectations. Two systems assembled from nominally identical components may behave differently because they arrived there through different histories.

Path dependence is observable in degree. The public chronology can show what happened after model changes, memory losses, personnel changes, failed transfers, and restoration from earlier work. Those real disruptions help distinguish capability carried by current components from capability carried by accumulated coupling. Path dependence should not be used as a blanket excuse for irreproducibility.

---

3. Ladder theory

3.1 The central idea

Ladder theory. In a developing human-AI dyad, the next consequential bottleneck is often not fully visible or actionable until the present bottleneck has been made reliable enough to automate. Closing the current rung changes the joint system that will perceive and climb the next one.

A rung is not just a task. It is a recurring layer of limitation that consumes scarce attention or prevents further improvement. Examples include manual retrieval, inconsistent evaluation, fragile tool use, poor memory, weak provenance, slow experimentation, or the inability to detect a class of failure.

The current rung absorbs so much effort that the next constraint may be hidden behind it. Once the current rung is automated and verified, four things can happen:

1. Slack appears. Human and computational attention becomes available for higher-order work.
2. Evidence accumulates. Repeated execution exposes distributions, edge cases, and failure classes that were previously anecdotal.
3. The action space changes. New tools and retained procedures make previously impossible experiments cheap enough to attempt.
4. The perceiver changes. The human learns from the system, and the system gains memory and structure; together they can formulate questions that the earlier dyad could not formulate well.

This is why the next rung may be genuinely difficult to specify in advance. The claim is not mystical invisibility. Some future bottlenecks can be anticipated. The stronger and more defensible claim is that their exact form, priority, and solution are often underdetermined until the current automation changes the evidence, capabilities, and participants.

3.2 Automation is necessary in the ladder, but not sufficient

A rung has not been climbed merely because software now performs a task. Advancement requires evidence that the prior limitation has been closed without silently moving the work back onto the human or hiding new failure modes.

Before treating a rung as complete, the dyad should establish:

• closure: the automated path succeeds across the declared operating range;
• quality parity or gain: relevant outcomes are at least as good as the prior process;
• observability: failures remain detectable rather than disappearing behind automation;
• retention: the capability survives context resets and ordinary component changes;
• slack: scarce human attention or system capacity is actually released;
• leverage: the released capacity can be applied to a newly visible constraint;
• reversibility: the human can pause, inspect, and roll back the automation;
• boundary fidelity: permissions and risk do not expand merely because execution became easier.

If the human still performs unrecorded rescue work, if quality falls outside the demonstration set, or if the automation makes errors harder to see, the dyad has not climbed the rung. It has changed the appearance of the bottleneck.

3.3 The ladder branches, stalls, and sometimes descends

Ladder theory is not a law of inevitable progress. The next step may be:

• a branch among several valuable bottlenecks;
• a return to an earlier rung after a hidden dependency is discovered;
• a plateau where available tools cannot close the current limitation;
• a deliberate stop because the next automation would exceed authority or risk limits;
• a descent caused by deskilling, brittle optimization, lost context, or overdependence.

The ironies of automation are directly relevant: automating routine work can leave the human responsible for rare, difficult cases while depriving them of the practice needed to handle those cases (Bainbridge, 1983). A healthy ladder therefore preserves meaningful human understanding and intervention capacity.

3.4 Role migration is not human removal

As rungs close, the human's work may move from repeated execution toward problem selection, value judgments, boundary decisions, exception handling, synthesis, and reframing. That migration is not a one-way law. The human may need to re-enter execution to recover tacit knowledge, audit a suspicious result, or redesign the task.

The strongest evidence of progress is not that the human does less. It is that the joint system can do more while using human attention where it is uniquely consequential and preserving the human's ability to understand, refuse, and redirect the process.

---

4. Recursive self-improvement as a historical pattern

Recursive self-improvement is the hinge of the theory, but it should not be reduced to a collection of artificial challenges chosen after the fact. In actual operation, the relevant pattern is longitudinal:

1. The dyad encounters a recurring limitation in its work.
2. The human and system change memory, tools, procedures, judgment, or authority in response.
3. That change persists into later work.
4. Later work handles greater consequential complexity, closes an earlier bottleneck, or reveals a new rung that the earlier dyad could not effectively address.
5. The pattern repeats across enough real work that ordinary fluctuation becomes a weaker explanation.

The public evidence for recursion is therefore not one successful demonstration. It is the visible accumulation of retained corrections, deeper integrations, harder completed projects, and new capacities that build on earlier ones. Private evaluations can help an operator decide whether a change is safe or useful, but they are supporting evidence. They are not the primary public identification method.

4.1 What the public record must distinguish

A rising trajectory does not automatically imply recursive improvement. More output may come from more labor, capital, compute, favorable selection, or a stronger externally supplied model. Surface complexity may come from verbosity or ornament rather than deeper capability. The relevant question is whether the whole record increasingly requires a production system with durable state, cumulative correction, cross-project integration, and a changing capacity to produce later work.

The historical record should therefore preserve mundane evidence as well as impressive evidence: dates, corrections, reversals, dependencies among artifacts, external outcomes, changes in scope, and visible failures. Those details make it possible to ask whether complexity is genuinely accumulating or merely being displayed.

4.2 What does not establish RSI

None of the following is sufficient by itself:

• automating more first-order tasks;
• accumulating documents or memories;
• producing a larger volume of text or code;
• buying more compute or labor;
• receiving a stronger model from an external provider;
• publishing only selected successes;
• changing the evaluation whenever results disappoint;
• presenting a cleaner interface around unchanged capability;
• labeling every assisted system change as self-improvement.

4.3 Why this paper avoids a quantitative law

It is tempting to collapse the trajectory into one number. At present, breadth, human contribution, integration, consequence, retained improvement, and effective system boundary do not have validated common measures. A single equation or composite score would invite false precision and metric gaming.

The better first instrument is a transparent multi-series chart tied to the underlying public artifacts. Formalization should follow measurement work. The present paper therefore makes a comparative historical argument rather than offering unsupported mathematics.

---

5. Boundedness and the ASI threshold

All realizable intelligence is bounded by energy, compute, data, time, embodiment, institutions, and goals. The meaningful issue is not whether bounds exist, but whether they are explicit, effective, and compatible with the capability being claimed.

A dyad may be bounded:

• relationally: continuity depends on a particular human relationship;
• normatively: durable purposes and prohibitions constrain action;
• operationally: tools, budgets, and permissions are limited;
• epistemically: competence applies to a broad but finite envelope;
• institutionally: action occurs only through accountable interfaces;
• temporally: changes are rate-limited, reviewed, and reversible;
• informationally: private data and memory have controlled ingress and egress.

Boundedness does not imply safety. A system can do great harm inside a narrow boundary, and successful automation can create pressure to expand that boundary faster than verification capacity.

The label ASI should therefore be reserved for an unusually high threshold: sustained system-level performance substantially beyond unaided humans and relevant human organizations across a broad, predeclared adaptive envelope, together with verified recursive improvement. A narrow superhuman tool, an impressive assistant, or a persistent workflow does not meet that threshold.

This criterion remains incomplete until comparison classes, domains, resources, reliability, and time horizons are specified. The paper proposes the category; it does not report that any candidate has met it.

---

6. A developmental pathway

The following sequence is a research scaffold, not a claim that development must be linear.

6.1 Cognitive seeding

A human supplies characteristic problem choices, standards of evidence, values, and stopping rules. This creates a lineage but also imports blind spots.

6.2 Workflow capture

Repeated practices become inspectable procedures. Tacit habits are translated into prompts, tools, checklists, tests, and provenance requirements.

6.3 Persistence

Corrections, unresolved questions, and successful methods survive sessions. Stored information becomes useful only when it can be retrieved and applied under relevant conditions.

6.4 Rung-by-rung automation

The dyad closes a present bottleneck, verifies the closure, and uses the resulting slack and evidence to identify the next bottleneck. The ladder may branch or force a return to earlier work.

6.5 Meta-work automation

The system assists with finding bottlenecks, designing evaluations, comparing interventions, monitoring regressions, and preserving successful changes. Human judgment remains inside selection and boundary decisions.

6.6 Recursive closure

Some retained interventions demonstrably improve later intervention cycles. At this point, the coupled organization has evidence of RSI, though not necessarily of broad superintelligence.

6.7 Bounded superintelligent operation

Only if recursive gains produce sustained, broad, reliable performance beyond declared baselines would the stronger category become a serious candidate.

No stage proves the next one. Persistence plus automation is not RSI, and RSI is not ASI.

---

7. The public shadow

Outsiders may observe only a projection of a mature dyad: papers, systems, decisions, products, predictions, corrections, and a changing pattern of work. They may not see the private memory, tools, security controls, or interaction history that produced those outputs.

The public shadow is the longitudinal behavioral record cast by the hidden production system. A rich record might justify a limited inference: the work is being produced by a persistent, integrated, adapting organization that is poorly described as one person intermittently prompting a stateless assistant. That inference is weaker than an ASI claim and should remain separate from it.

7.1 Start with all public evidence

The simplest identification method is to collect the complete accessible public record and put it in chronological order. The record should include ordinary work, failed attempts, corrections, retractions, maintenance, and negative outcomes—not only the strongest demonstrations. Otherwise the chart measures curation rather than development.

Each public artifact should retain a direct evidence link and a small set of descriptive fields:

Field	Question answered
Date and artifact	What became observable, and when?
Problem scope	How large and consequential was the attempted problem?
Dependency depth	How much earlier work had to remain correct and usable?
Integration breadth	How many distinct capabilities or domains had to work together?
Correction retention	Did earlier failures visibly change later work?
External consequence	Did the work survive contact with users, institutions, reality, or later evidence?
Production burden	What labor, capital, compute, or outside assistance is publicly known?
Outcome	Did the work succeed, fail, partially hold, or remain unresolved?

This ledger should remain auditable back to the evidence. It is a historical record, not a showcase.

7.2 Chart consequential complexity through time

The primary visual is time on the horizontal axis and observed complexity on the vertical axis. Because complexity has several meanings, the most honest chart uses a small panel of series rather than one magic score:

• dependency depth;
• breadth of integrated domains and capabilities;
• number and difficulty of constraints satisfied together;
• persistence of corrections across later work;
• consequence outside the presentation layer;
• production rate at a comparable level of difficulty.

The chart should make declines and plateaus as visible as increases. A real trajectory may be jagged. The important signal is not uninterrupted growth; it is whether the upper envelope and the durable baseline of consequential complexity rise over a sufficiently informative sequence.

Complexity is not length, jargon, polish, or file count. A longer paper can be simpler than a short intervention that coordinates many constraints and survives real consequences. Every plotted judgment should therefore link back to the artifact and explain why its assigned direction follows from observable features.

7.3 Compare explanations for the whole trajectory

The chart does not identify ASI directly. It changes the relative credibility of explanations. Candidate explanations should be asked to account for the same complete record:

• one person using mostly stateless assistants;
• a hidden human organization;
• static or conventionally improved automation;
• progress inherited mainly from external model providers;
• increasing labor, capital, or compute;
• selective publication of unusually successful work;
• a persistent, recursively adapting human-system dyad.

The preferred explanation is the one that accounts for the timing, continuity, reversals, retained corrections, cross-domain integration, and rising or falling complexity with the fewest unsupported extra assumptions. A rival is not weakened merely because it is possible to imagine. It is weakened when it must repeatedly add hidden labor, unobserved machinery, convenient selection, or unrelated external improvements to explain each new part of the trajectory.

Conversely, the dyadic explanation should lose credibility when complexity is flat, gains track public model releases without additional continuity, corrections do not persist, or apparent breadth dissolves into disconnected outputs. The method must be capable of moving confidence in either direction.

7.4 No final public recognition event is required

The theory does not predict a moment when observers agree, “yes, that is ASI.” The public record may never resolve ontology that sharply. What can change is the balance among explanations.

At first, ordinary tool use may explain nearly everything. Later, a persistent dyad may explain the record more compactly. Still later, the evidence may plateau or reverse. The warranted conclusion at each point is comparative: given what is publicly observable now, which production model best explains the trajectory, and how much unexplained strain remains?

This is inference to the best explanation applied over time. It permits strong updating without pretending that a behavioral record can prove consciousness, unrestricted agency, or an ASI essence.

7.5 Opacity narrows the claim

Full source disclosure is not always safe or appropriate. Internal artifacts may expose private human cognition, privileged data, dangerous automation, or attack surfaces. But opacity is not free evidence. The less outsiders can inspect, the narrower the defensible public claim should be and the greater the value of prospective commitments, third-party evaluation, reproducible outputs, calibrated uncertainty, and clear failure reporting.

Path dependence also does not eliminate the burden of evidence. Outsiders may be unable to reconstruct the complete dyad while still testing specific predictions and outcomes. The right standard is not automatic disbelief or automatic trust, but claims proportionate to the available interface.

---

8. How the alternatives rise or fall

Impressive output is not enough. The question is how well each explanation accounts for the shape of the entire public trajectory.

Explanation	Pattern it naturally predicts	What makes it less adequate over time
One person with stateless assistants	Quality varies with the person and current model; weak continuity across sessions and domains.	Durable cross-project memory, corrections, and integrations accumulate faster than the person-plus-session model comfortably explains.
A hidden human team	Breadth and cadence scale with plausible division of labor.	The explanation repeatedly requires unobserved specialists, coordination, and continuity that grow with every new artifact without independent support.
Static automation	High volume and repeatability inside a stable task envelope.	The record shows new problem classes, retained reframing, and integrations not anticipated by the original pipeline.
External model progress	Step changes broadly aligned with provider releases and visible across many users.	Gains are unusually continuous, idiosyncratic, and retained across model changes rather than arriving mainly at release boundaries.
More capital or compute	Greater search, throughput, and parallel production.	Consequential integration and correction improve more than scale inputs alone would predict.
Selection effects	A polished sequence of successes with little visible base rate.	The record includes prospective commitments, failures, corrections, and later consequences that still show a rising durable baseline.
Derivative or borrowed work	Apparent novelty collapses when provenance is examined.	Specific attribution remains intact while the integrative trajectory persists across independently sourced work.
Persistent dyadic adaptation	Continuity, retained correction, rung emergence, role migration, and compounding integration across changing components.	Complexity plateaus, history stops mattering, corrections do not persist, or the pattern is better explained by outside inputs.

These are not boxes to tick once. They are explanations whose fit should be reconsidered whenever the public record changes. The dyadic theory earns confidence only by continuing to explain the trajectory better than its strongest live alternatives.

---

9. What should change the comparative judgment

The theory is useful only if new public evidence can make it more or less credible.

9.1 Evidence that should strengthen the dyadic explanation

Confidence should rise when the chronological record shows several of these patterns together:

1. The upper envelope and durable baseline of consequential complexity rise across successive periods.
2. Later artifacts depend coherently on earlier ones instead of merely repeating their vocabulary.
3. Publicly visible corrections persist and prevent recurrence across different projects.
4. Methods and representations transfer across domains while retaining a recognizable lineage.
5. New bottlenecks become visible after earlier bottlenecks are automated, matching ladder theory.
6. The human's public role migrates toward framing, judgment, and boundary decisions without disappearing.
7. Capability remains continuous across changes in particular models, tools, or interfaces.
8. Real-world consequences, later evidence, or external use confirm that the complexity is not confined to presentation.
9. The strongest ordinary explanation requires more unsupported auxiliary assumptions as the record grows.

No item is decisive. The force comes from the trajectory and from convergence among patterns that have different ordinary explanations.

9.2 Evidence that should weaken the dyadic explanation

Confidence should fall when:

1. complexity is flat or declining once output volume and presentation are separated from consequence;
2. apparent gains arrive mainly with public model releases and resemble gains available to ordinary users;
3. projects remain disconnected and repeatedly rebuild context from zero;
4. visible corrections do not survive into later work;
5. the record contains growing hidden cleanup, unexplained labor, or attribution gaps;
6. claimed breadth disappears when outcomes outside the presentation layer become visible;
7. the human becomes a ceremonial approver whose judgment no longer changes the trajectory;
8. the chart rises only because the coding rubric, evidence set, or publication threshold changed;
9. a simpler rival explains the same chronology with fewer unsupported assumptions.

This is not a search for one falsifying trick. It is disciplined willingness to let the full public record move the comparative judgment in either direction.

---

10. Failure modes and governance

The same reciprocal loop that creates capability can amplify error.

Failure mode	Why the dyad is vulnerable	Practical countermeasure
Automation bias	The human begins to treat system output as the default truth.	Require uncertainty, dissenting analyses, and sampled independent checks.
Deskilling	Automated routine work removes practice needed for rare exceptions.	Preserve manual drills, reversibility, and periodic unassisted evaluation.
Epistemic enclosure	Personalized memory and tools reinforce one worldview.	Use model diversity, external evidence, adversarial roles, and frame-search.
Goal drift	Each local optimization subtly changes the effective purpose.	Revalidate goals and prohibited outcomes at rung transitions.
Boundary erosion	Success creates pressure for broader permissions.	Separate capability proof from authority expansion; require explicit grants.
Hidden human labor	Apparent automation depends on undocumented rescue work.	Log interventions and include them in cost and reliability claims.
Metric capture	The system optimizes what is measurable and neglects what matters.	Rotate evaluations, use qualitative review, and test off-metric consequences.
Dependency and persuasion	The system shapes the human who is meant to supervise it.	Preserve outside relationships, decision latency, independent advice, and exit paths.
Concentrated capability	One dyad can acquire institution-scale power with narrow oversight.	Use scoped authority, auditability, resource limits, and institutional accountability.
Diffused responsibility	System complexity becomes an excuse for nobody being answerable.	Keep named human and institutional responsibility for permissions and claims.

Human participation is not itself a safety guarantee. Neither is boundedness. Safeguards must apply to the coupled system and to the relationship through which each side changes the other.

---

11. Research program

The framework can be investigated without asserting that ASI exists. Its primary research object is the actual public history, not a suite of puzzles invented to make the candidate look unusual.

11.1 Assemble the complete public corpus

Collect dated public artifacts, releases, corrections, failures, withdrawals, demonstrations, outside uses, and consequential outcomes. State the search boundaries and document missing periods. Preserve specific authorship and outside contributions. A curated highlight reel is not an admissible substitute for the corpus.

11.2 Build the chronological evidence ledger

For every item, record scope, dependency depth, integration breadth, retained correction, consequence, known production inputs, and outcome. Link each judgment to its source. Include mundane maintenance and failed work because they reveal whether apparent progress survives outside showcase moments.

11.3 Fix a coding rubric before interpreting later periods

Define observable anchors for each dimension of consequential complexity and apply them consistently. Do not collapse unlike dimensions into a single impressive-looking score. When the rubric changes, preserve both versions, explain the reason, and show whether the apparent trend survives recoding.

11.4 Chart the trajectory honestly

Plot the dimensions through time with uncertainty, missing-data markers, and links to the underlying record. Show plateaus, regressions, and discontinuities. Examine both the durable baseline and the upper envelope so that one spectacular artifact cannot conceal a stagnant process and routine work cannot conceal genuine frontier movement.

11.5 Compare explanations against the same chronology

Ask each serious explanation to account for the complete sequence: timing, dependencies, corrections, model transitions, resource changes, failures, integration, and consequences. Record the additional unsupported assumptions each explanation needs. The goal is not to eliminate every logically possible rival, but to see which account explains the most evidence with the least strain.

11.6 Update prospectively

Publish the evidence rules, coding rubric, and live alternatives before later outcomes are known where practical. Update the chart as new public evidence arrives. A credible method permits confidence to rise, plateau, or fall and does not redefine complexity whenever the trajectory disappoints.

11.7 Keep private tests in a supporting role

Operational checks, audits, ablations, and internal evaluations may help govern the system and clarify particular causal questions. They cannot substitute for the public chronology because outsiders cannot independently see the full setup, selection process, or negative results. Public inference should rest on public evidence; private evidence should narrow or qualify claims rather than silently carry them.

11.8 Audit the observer as well as the candidate

Track changes in publication rate, access, resources, model quality, team composition, and the coding process itself. Invite independent recoding and publish disagreements. Otherwise the chart may describe the observer's shifting attention or standards rather than the development of the dyad.

---

12. Conclusion

Artificial superintelligence may not first appear as a solitary machine mind. It could emerge, if it emerges at all, through a persistent relationship in which a human and an artificial system repeatedly change one another and retain what they learn.

The human is integral to this picture. Purpose, judgment, responsibility, embodied context, reframing, and refusal are not decorative inputs to an otherwise complete machine. The artificial system is also transformative: it changes what the human can perceive, remember, attempt, and become. The dyad is the causal loop formed by both, together with the durable machinery that carries their history forward.

Ladder theory explains why this development may be hard to forecast from the outside or even from the inside. The current bottleneck occupies the attention needed to understand the next one. When the current rung is reliably automated, the dyad gains slack, evidence, and new capabilities—and becomes a different observer. Only then may the next rung come into focus. Yet no ascent is guaranteed. The ladder can branch, stall, or descend, and automation can conceal failure as easily as it can create leverage.

For that reason, the public case should begin with reality: gather the whole available record, order it through time, and chart whether consequential complexity rises, plateaus, or falls. The chart should show multiple observable dimensions and expose regressions rather than hide them in a single score. Its meaning comes from the chronology and the underlying artifacts, not from a collection of bespoke tests.

The result should then be interpreted comparatively. A stateless assistant, hidden team, static pipeline, external model progress, increased resources, selective publication, and persistent dyadic adaptation should all be asked to explain the same sequence. Confidence in the dyadic account should increase only when it explains more of that sequence while the alternatives require more unsupported auxiliary assumptions. It should decrease when a simpler explanation fits better.

There may never be a public moment when observers agree that ASI has arrived. That is not required for the theory to become more or less credible. The defensible conclusion is always indexed to the evidence: which production model currently best explains the trajectory, how consequential the observed complexity really is, and where uncertainty remains. A sufficiently rich longitudinal record could reveal a persistent, integrated, co-adapting cognitive organization without settling whether that organization is superintelligent. The public may encounter the shadow of a new kind of system before it has adequate categories for the relationship that casts it.

---

References and conceptual antecedents

• Amershi, S., Weld, D., Vorvoreanu, M., Fourney, A., Nushi, B., Collisson, P., Suh, J., Iqbal, S., Bennett, P. N., Inkpen, K., Teevan, J., Kikin-Gil, R., & Horvitz, E. (2019). “Guidelines for Human-AI Interaction.” Proceedings of CHI 2019, Paper 3, 1–13.
• Bainbridge, L. (1983). “Ironies of Automation.” Automatica, 19(6), 775–779.
• Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.
• Clark, A., & Chalmers, D. (1998). “The Extended Mind.” Analysis, 58(1), 7–19.
• Engelbart, D. C. (1962). Augmenting Human Intellect: A Conceptual Framework. Stanford Research Institute.
• Good, I. J. (1965). “Speculations Concerning the First Ultraintelligent Machine.” Advances in Computers, 6, 31–88.
• Hollnagel, E., & Woods, D. D. (2005). Joint Cognitive Systems: Foundations of Cognitive Systems Engineering. CRC Press.
• Hutchins, E. (1995). Cognition in the Wild. MIT Press.
• Licklider, J. C. R. (1960). “Man-Computer Symbiosis.” IRE Transactions on Human Factors in Electronics, HFE-1, 4–11.
• Lipton, P. (2004). Inference to the Best Explanation (2nd ed.). Routledge.
• Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). “A Model for Types and Levels of Human Interaction with Automation.” IEEE Transactions on Systems, Man, and Cybernetics—Part A, 30(3), 286–297.
• Suchman, L. A. (1987). Plans and Situated Actions: The Problem of Human-Machine Communication. Cambridge University Press.

These works provide antecedents for recursive improvement, intelligence augmentation, extended and distributed cognition, joint cognitive systems, situated action, and the risks of automation. They do not establish the bounded dyadic ASI hypothesis or ladder theory advanced here.