Consciousness / Agency and free will
We can revise
the rules
by which we choose.
A choice depends on what an agent can perceive, retain, assess and do. Reflective freedom enters when the agent can examine a rule of decision and install a different rule that changes what it does next.
Bounded Agency and Reflective Freedom
Evaluative Revision, Information Limits, and Faithful Realization
Jeremy Rodgers · Independent Researcher · 6 October 2026
doi:10.5281/zenodo.23202999 ↗
The published contribution
A mechanism.
A limit. An exact advantage.
The change survives the next case.
A finite agent reads retained commitments, assesses candidate charters and installs a decision rule. Fresh cases separate the new disposition from a changed command. Interventions test the actual read and installation routes.
Information has to reach the decision.
Accurate endorsement places a lower bound on the distinctions the inquiry can carry. A causal response can still install the wrong rule. The information, mediation and endorsement requirements must all be met.
The implementation keeps the contract.
A native realization preserves the specified operations, paid access restrictions and fresh-case behaviour. Protected returns and predictive structure coexist with evaluative agency under the declared assumptions. Recurrence by itself is insufficient.
Information limits faithful revision.
Here H₀ and H₁ are the inquiry-transcript laws, P₀ and P₁ are the installed-amendment laws, and ε₀ and ε₁ are endorsement errors. The two contexts endorse disjoint amendment sets and share the same downstream installation kernel. TV measures how distinguishable two probability laws are.
Inspect the rule-revision mechanism and prove the bound ↗Adaptive inquiry has a sharp first winning budget.
In the paper’s latent-sensor and endorsement model, L is the first calibration count that can improve evidence accuracy, b is the uncalibrated accuracy, and CL is the optimum after L calibration queries. V permits adaptive query counts; F fixes those counts in advance. Their first strict separation is exactly L + 2 paid queries. The chapter develops the full family, optimal policy and proof.
Follow the exact model, optimal policy and proof ↗What this means for freedom
A bounded perspective
can govern its next move.
The agent never needs a complete copy of reality to act competently. What matters is whether its available distinctions support the task, whether its evaluation reaches the rule it installs and whether that rule governs a later case.
This gives a precise form to local self-government. It also makes the next questions sharper: where do retained standards come from, how can conditioning lose its authority, and what can an audit establish about a history of change?
The extended inquiry brings those philosophical questions together with further results on source access, reversible control, information bottlenecks, capability construction and provenance. Each added construction keeps its own assumptions. Revising a decision rule and originating the standards used to assess it remain different achievements.
Follow the deeper inquiry ↓Complete published argument / 11 chapters
The paper, developed for the web.
Every substantive definition, construction, result, proof, counterexample and appendix is included in the reading sequence. The mathematics sits beside the question it answers.
- Section 1Agency as a physical capacity↗
Physically caused decisions, evaluative mediation and the three constructive contributions.
- Section 2What self-governance requires↗
Executable options, retained evaluation, information access and implementation invariance.
Inside this chapter
- 2.1 State, information and executive targets
- 2.2 Executable options and actual decisions
- Definition 2.1: evaluative mediation
- Definition 2.2: bounded represented-rule self-revision
- 2.3 The evaluative role
- Assessment has no need for an infinite sequence of selectors
- Inward acts, inhibition and tools
- 2.4 Freedom has a profile
- 2.5 Testing the actual causal pathways
- Proposition 2.3: transport of a finite certificate
- Finite implementation error
- Section 3A rule can become an object of assessment↗
An editable charter, fresh-case tests and the sharp information–mediation–endorsement inequality.
Inside this chapter
- 3.1 The charter
- The retained commitments
- Proposition 3.1: selective represented-rule revision
- A changed disposition, demonstrated on a new case
- Every stage has a cost
- 3.2 Testing reading, installation and endorsement
- Proposition 3.2: information, installed contrast and endorsement
- A causally responsive evaluator can still be wrong
- Section 4Keeping revision available↗
Exactly when bounded inquiry can preserve endorsed amendment and future auditability.
- Section 5When inquiry is fallible↗
Finite zero-error limits, exact noisy-read frontiers and repeated-audit risk.
Inside this chapter
- 5.1 Why finite noisy inquiry cannot guarantee certainty
- Proposition 5.1: the finite full-support obstruction
- 5.2 A noisy charter audit
- Theorem 5.2: the exact bounded noisy-revision frontier
- The cost of a warranted tolerance
- What the intervention test measures
- Repeated audits and their joint risk
- Two ways to misread an improvement
- Section 6Evidence and adaptive inquiry↗
The complete sensor family and proof of the first adaptive advantage at B = L + 2.
- Section 7Faithful native realization↗
Protected returns, actual interventions, paid access and explicit physical schedules.
Inside this chapter
- 7.1 The physical contract comes first
- Carriers, operations and the apparatus boundary
- Native records must determine the right successor classes
- 7.2 A protected source carries the return
- Theorem 7.1: protected-source compilation
- Corollary 7.2: return organization does not entail evaluation
- 7.3 An eighteen-cell machine that changes its charter
- Test the paths that do the work
- 7.4 Stochastic closure and paid executive information
- Theorem 7.3: a successor-revealing continuation
- A native record is not automatically an executive read port
- Lemma 7.4: preservation of executive information
- 7.5 A physical realization of the exact inquiry advantage
- Two flags are enough at the first useful threshold
- Why the statistical result survives implementation
- Keep the source and the boundary complete
- How much implementation error can the advantage tolerate?
- Three witnesses, three established capacities
- Sections 8–9Ownership, history and freedom↗
What the results establish about local freedom, developmental history and consciousness.
Inside this chapter
- A decision can belong to a caused process
- Present control and historical ownership
- Proposition 8.1: the limit of present-state provenance tests
- Follow commitment formation as it happens
- Plasticity does not exclude manipulation
- Conflict, steadfastness and the time of a choice
- Consciousness requires its own physical account
- Experiments that distinguish the claims
- What has been established
- Publication, funding and reproducibility
- Appendix ARevision kernels and finite-horizon risk↗
Complete query recurrences, writable-validator kernels and finite-horizon risk calculations.
- Appendix BPhysical carriers and costs↗
Smooth bit dynamics, Gaussian refinement and detailed time, gate and work bounds.
Inside this chapter
- B.1 Why a covering identity tour needs twice the path length
- B.2 Explicit continuous bit dynamics
- A paid clock gives each gate its place
- Decode the well sign without disturbing the source
- A Boolean SWAP with no additional sampling cell
- Restoration closes the induction
- B.3 Gaussian noise gives a finite-error refinement
- Writes and restoration tolerate bounded forcing paths
- The rotating SWAP has a separate martingale bound
- Couple the complete finite episode
- B.4 The inquiry machine's full cost ledger
- Conditional time and energy bounds
- ReferencesReferences and publication record↗
The complete 24-entry bibliography and publication record.
Inside this chapter
- 1. Abstracting causal models
- 2. Certified self-modifying code
- 3. Games with imperfect information
- 4. Self-modification of policy and utility
- 5. Responsibility and manipulation
- 6. Responsibility and control
- 7. Sequential testing under deadlines
- 8. Adaptive submodularity
- 9. Transitive factorisations
- 10. Smooth simulation of computation
- 11. Selecting computations
- 12. Adaptivity gaps in Boolean evaluation
- 13. Discovering agents
- 14. Can AI systems have free will?
- 15. Unsupervised reliability calibration
- 16. Chance, choice and control
- 17. General agents need world models
- 18. Relational boundaries and awareness localization
- 19. Shadow Theory and Consciousness
- 20. Learning effective interfaces
- 21. Identifying binary realizations
- 22. Causal consistency
- 23. Risk-aware partially observed planning
- 24. Gödel machines
Extended inquiry / 6 chapters
Freedom, control and the limits of a perspective.
Additional philosophy and technical results developed from Bounded Agency and Reversible Control. These chapters extend the website treatment beyond the published paper and retain their separate models and assumptions.
- Extended inquiry 1Bounded perspective, real control↗
Why incomplete knowledge can coexist with exact action, and the precise point at which an aperture obstructs control.
- Extended inquiry 2Where uncertainty goes↗
Exact receiver capacity, calibration and causal deadlines reveal the physical resources behind successful control.
- Extended inquiry 3The operations that make control possible↗
Equal knowledge and equal memory can support unequal control. Exact affine and nonlinear frontiers show why.
- Extended inquiry 4Building a new way to act↗
A controller can acquire an unknown interface, compress its retained structure and gain a lasting capability under fixed laws.
- Extended inquiry 5A chooser with a history↗
How self-revision can be real without erasing its causes, and what bounded agency does and does not settle about authorship.
Inside this chapter
- Extended inquiry 6The work behind an available option↗
A compact result can require substantial inquiry and computation. Search guarantees and physical transfer make that cost explicit.
Inside this chapter
- One changed bit can hide a difficult decision
- A supported coordinate supplies a control instruction
- Identification, action and positive verification have different costs
- A sound certificate is not a completed search
- Progress must be measured at the level a proof needs
- From a logical controller to a physical experiment
- Transfer the whole experiment
Search the complete consciousness programme
Find passages in the monograph, Papers 2–4, relational development, the agency article, guides, glossary and FAQ.