From Khayyam's Geometry to Artificial Consciousness

 

From Khayyam's Geometry to Artificial Consciousness

A Falsifiable Pilot Study Protocol for Researchers

Status: Proposed pilot protocol; hypotheses requiring empirical validation
Audience: Computational neuroscience, AI, cognitive architectures, philosophy of mind
Central question: Can a minimal architectural intervention—recurrent state integration plus an explicit self-model—produce measurable, theory-predicted effects on uncertainty handling, error correction, and cross-task information sharing, under controlled disruption?

 Abstract

Relevance of Khayyam's Method to a Falsifiable Pilot Study of Artificial Consciousness

Omar Khayyam's geometric treatment of cubic equations offers a methodological analogy rather than a metaphysical claim: when a problem resists solution in its original representation, transforming it into a new representation with explicit constraints can make the solution tractable and investigable. This article applies that lesson to artificial consciousness research. Rather than asking whether scale alone will produce conscious machines, we propose a minimal, falsifiable pilot study comparing a baseline AI agent with variants that add recurrent state integration and an explicit self-model. The Khayyam-inspired move is to reframe the problem as the intersection of two constraint sets—functional integration (recurrent world-modelling, self-monitoring, action-guiding feedback) and empirical evidence (theory-derived, preregistered predictions tested under controlled disruption)—and to treat that intersection as a candidate architecture meriting investigation, not as a consciousness detector. The pilot tests whether controlled disruptions of recurrence and self-modelling produce selective, predicted effects on uncertainty calibration, error correction, and cross-task information sharing, with capacity-matched controls as the critical falsifier. Khayyam's method contributes three things to this design: a rationale for changing the representation of the problem, a discipline of explicit constraints and derived predictions, and a warning that mathematical intersection has precise definitions whereas subjective experience does not. The analogy thus structures a research programme that can eliminate weak explanations and narrow plausible hypotheses without presupposing that the answer must be a conscious machine.

Keywords: artificial consciousness, Khayyam analogy, falsifiability, recurrent integration, self-model, preregistration, ablation, research methodology.

 

1. Why Start Small

The article's full consciousness architecture is too broad to falsify in a single study. The most tractable first experiment is a three-arm, preregistered pilot that compares:

  • Baseline agent — feedforward task policy, no recurrent state integration, no explicit self-model.
  • Recurrent-only agent — adds recurrent state integration (hidden state carried across timesteps, updated by new observations and predictions).
  • Recurrent + self-model agent — adds an explicit self-model representing internal state, uncertainty, and task progress, with the representation causally connected to action selection.

The pilot tests whether controlled disruptions of these components produce the predicted effects on three outcome families: uncertainty calibration, error correction, and cross-task information sharing. It does not test consciousness. It tests whether the architectural distinction is real, measurable, and causally specific.


2. Falsifiable Hypotheses

ID

Hypothesis

Prediction

Falsification condition

H1

Recurrent integration improves uncertainty calibration

Recurrent-only and recurrent+self-model agents show lower Brier score / expected calibration error than baseline on held-out uncertainty tasks

No difference, or baseline equal/better, after capacity controls

H2

Explicit self-model causally contributes to error correction

Disrupting the self-model (noise injection or lesion) selectively degrades error-correction performance, more than a matched control disruption

Disruption has no selective effect, or affects all tasks equally

H3

Recurrent + self-model enables cross-task information sharing

Agent transfers a learned uncertainty-estimation strategy to a novel task faster than baseline and recurrent-only

No transfer advantage, or advantage explained by general capacity

H4

Effects are causally specific, not capacity artifacts

Parameter-matched controls (baseline given extra capacity) do not reproduce the effects

Capacity-matched baseline matches the full agent

H4 is the critical falsifier. If a parameter-matched baseline reproduces the effects, the architectural claim fails.


3. Minimal Architecture Specification

Keep each module implementable in weeks, not years.

  1. Perception/world model — small encoder producing a latent state from observations.
  2. Recurrent integration — hidden state , where is a prediction. Ablate by severing .
  3. Self-model — a low-dimensional vector representing (a) estimated uncertainty, (b) task progress, (c) recent action history. Must be read by the policy, not just logged.
  4. Workspace (optional in pilot) — a shared bottleneck through which task-relevant information passes to multiple heads. Include only if compute allows; otherwise defer to Phase II.
  5. Evaluation layer — independent logger of predictions, observations, and discrepancies. Must not feed back into the agent.

Ablation handles: recurrent lesion (sever ), self-model lesion (zero or noise ), matched control lesion (equivalent noise on a non-self-model channel).


4. Task Battery (Keep It Small)

Use three tasks, each with a clear dependent variable:

  • Uncertainty task: probabilistic prediction with scored calibration (Brier score, ECE).
  • Error-correction task: sequential decision task with delayed feedback; measure recovery rate after induced errors.
  • Cross-task transfer: train on task 1, test on a structurally similar but novel task; measure sample efficiency of transfer.

All tasks should be runnable on a single GPU within a day per arm.


5. Experimental Sequence

Phase 0 — Preregistration. Register hypotheses, outcome measures, analysis plan, and failure criteria on OSF or equivalent before running decisive experiments.

Phase 1 — Baselines. Train all three arms to matched task performance. Report learning curves and final performance. If baseline cannot reach comparable performance, the comparison is invalid.

Phase 2 — Capacity controls. Train a parameter-matched baseline (extra capacity, no recurrence/self-model). This is the H4 test.

Phase 3 — Causal ablations. At test time, apply:

  • recurrent lesion,
  • self-model lesion,
  • matched control lesion.

Measure selective degradation per outcome family. Preregister the expected pattern: self-model lesion should hit error correction and uncertainty more than raw task accuracy.

Phase 4 — Cross-task transfer. Train on task 1, evaluate transfer to task 3. Compare sample efficiency across arms.

Phase 5 — Replication. Independent re-run with different seeds and, ideally, a second team using the released code and configs.


6. Outcome Measures and Interpretation Limits

Measure

What it tests

What it cannot establish

Brier score / ECE

Uncertainty calibration

That calibration reflects experience

Error recovery rate after induced errors

Error correction

That correction involves self-awareness

Transfer sample efficiency

Cross-task information sharing

That sharing is globally broadcast in a GNWT sense

Ablation selectivity

Causal contribution of a component

That the component is consciousness-relevant

Capacity-matched comparison

Architectural vs. capacity explanation

That the architecture is sufficient for consciousness

Report each measure separately. Do not collapse into a single "consciousness score."


7. Falsifiability Safeguards

  • Preregister predictions and failure criteria before decisive runs.
  • Capacity controls for every architectural claim.
  • Ablation and restoration: show removing a component produces the predicted effect and restoring it reverses it.
  • Alternative explanations: check whether results are explained by memory, optimization, or learned reporting conventions.
  • Negative results are publishable: a null result on H1–H3 is informative and should be reported.
  • Epistemic separation: report observations separately from theoretical interpretation and from any claim about subjective experience.

8. What This Pilot Can and Cannot Deliver

Can deliver:

  • Evidence that recurrent integration and self-modelling are causally distinct architectural features with measurable, selective effects.
  • A reproducible protocol and codebase for comparing architectures under matched conditions.
  • A template for theory-discriminating experiments (e.g., GNWT-style global availability vs. simpler routing).

Cannot deliver:

  • Any evidence of phenomenal consciousness.
  • A validated consciousness detector.
  • Resolution of the IIT vs. GNWT debate (the 2025 adversarial collaboration shows both theories remain contested even in human neuroscience).

The pilot's success criterion is not whether the agent declares itself conscious. It is whether the experiment eliminates weak explanations and narrows the space of plausible architectural hypotheses.


9. Proposed Timeline

Week

Milestone

1–2

Preregistration; task and architecture specification

3–4

Implement three arms + capacity-matched control

5–6

Phase 1 baselines; verify matched performance

7

Phase 2 capacity controls

8

Phase 3 ablations

9

Phase 4 transfer

10

Analysis, write-up, code release

11–12

Independent replication attempt


10. Research Questions for the Pilot

  • RQ1: Do recurrent integration and self-modelling produce distinct, reproducible effects when independently manipulated?
  • RQ2: Are those effects selective for the predicted outcome families, or do they reflect general capacity?
  • RQ3: Does a capacity-matched baseline reproduce the effects? (If yes, H4 fails.)
  • RQ4: Which results, if any, discriminate between competing theoretical accounts rather than merely confirming shared capabilities?

Conclusion

Khayyam's lesson is methodological: change the representation, constrain the problem, and let the intersection be investigated. For artificial consciousness, the first intersection to investigate is not "consciousness" but whether a minimal architectural distinction—recurrence plus self-model—has causally specific, falsifiable consequences. This pilot is designed to find out, and to be wrong in public if it is wrong.

Success is measured by eliminated explanations, not by a conscious machine.


Selected references: Butlin et al. (2023, arXiv:2308.08708); Butlin et al. (2026, Trends Cogn. Sci., doi:10.1016/j.tics.2025.10.011); Cogitate Consortium (2025, Nature 642, 133–142); Safron (2020, Front. Artif. Intell. 3:30); Safron et al. (2026, AAAI Symposium Series).

Comments