# Minds experiment roadmap Suggested sequence, 2026-09-15. Start with small systems whose behavior we can inspect, then ask whether they can predict, choose useful work, reuse skills, and change their connections. Build the missing mechanism just before the experiment that needs it. This is a suggested order, not a requirement to obtain a positive result at each step. Sorting can branch after experiment 02. A valid negative result can change the next question; it is not a broken gate. Exact replay and numeric compiler guards are not prerequisites. ## 01. Two spaces exchange estimates Status: Explored. Understand the current behavior. **Already built:** Already built: the two-space runner, Gojo observations of the parts and pair, event records, and the step viewer. The five conditions reuse the existing calibration and exchange fixture. **Question:** What do agreement, disagreement and ongoing movement look like at two scales? **Try:** Compare half, full and quarter responses; let the second space receive a fresh message; then select the existing local references after peer agreement. **Observe:** Half responses reach 7 and 7. Full responses keep swapping 5 and 9 while the paired average stays 7. Quarter responses still differ at the end. Fresh delivery gives 9 and 9; the selected references give 2 and 2. **Controls:** Focused checks cover arithmetic, observation without mutation, preserved past readings, duplicate response refusal and fractional requests. These are apparatus checks, not evidence of learned response rules. **Ready when:** Complete as an exploratory six-round walkthrough. It can be revisited without treating its outcomes as required answers for later conditions. **Limit:** No learned policy, automatic idling, or advantage over conventional arithmetic has been established. ## 02. Find the limits of two-space feedback Status: Explored. Understand the current behavior. **Already built:** Built: parameterized fixture and application controllers, a declared event tape, no-response and missing-delivery controls, and an independent exact rational implementation with its own messages, fractions and counters. **Question:** Which effects come from response strength, message timing, resolution or the chosen reference? **Try:** Use a small stated matrix of initial gaps, equal and unequal native tick sizes, missing or delayed messages, changed references, and longer observation windows. Repeat conditions to expose any unexplained variation. Add a floating-point reference as a separate approximation condition if useful. **Observe:** Completed 24 conditions in three fresh worlds each. All 72 runs matched the independent exact reference. Quarter responses became quiet with a remaining gap; finer and coarser ticks changed that gap. Timing, missing messages, reference selection and controller limits remained distinguishable in the records. **Controls:** The exact reference should agree when the full contract is the same; a discrepancy calls for investigation. Show a conventional version retaining the same history and uncertainty before attributing a benefit to Gojo. Change one factor at a time before studying interactions. **Ready when:** Completed as an exploratory comparison under feedback-limits-v1. All cases, repetitions, schedules and outcomes are retained. Additional policies are new conditions, not fixes to these outcomes. **Limit:** Numerical agreement is a correctness check. Any Gojo benefit must be demonstrated in inspectability, composition, retained information, adaptation or measured cost. ## 03. Make predictions that can be wrong Status: Explored. Understand the current behavior. **Already built:** Built: an adapter over existing prospective forecast structure, issue/arrival/assessment records, a before/after viewer, independent exact line calculations, external hold-last and no-prediction controls, coverage and missing-outcome accounting. **Question:** Can a local model predict an observation before it arrives? **Try:** Constant signals, ramps, reversals, unequal sample intervals and missing samples. First keep the affine family and sample selection fixed. Use separate development and evaluation sequences. **Observe:** Twelve fixed cases completed with no differences from the exact reference. Line continuation predicted the unequal-time ramp exactly but lost to hold-last on alternating values and a sudden step. Old predictions survived new fits; five blackout forecasts remained unassessed. **Controls:** Forecasts must precede their outcomes. The evaluator's later data must not reach the participant early. Compare hold-last on the same targets and include sequences on which the affine model should lose. **Ready when:** Completed under predictions-v1 without changing kernel laws or tuning the policy after evaluation. All tapes, before/after journals, failures, coverage and exact results are retained. These evaluation tapes are now inspected data, not unused cases for a revised rule. **Limit:** Fitting a selected family is not discovering that family, and fitting past data is not predictive success. ## 04. Test when a broader view is useful Status: Next. Let observations guide work. **Build before this experiment:** Extend the existing Gojo pair observer with explicit goals and read requests, using focus, ports and timelines. Keep passive observation separate from a connected request for finer information. Preserve both paired snapshots and intermediate events; full world replay is unnecessary. **Question:** Can an observation of the whole help predict a disturbance or decide when to look closer? **Try:** Present quiet agreement, balanced internal motion and a later disturbance. Compare a coarse view, a fine view, and a coarse view that requests detail under a stated rule. Score prospective warnings or a specific intervention on new trajectories. **Observe:** Missed disturbances, false alarms, response delay, extra reads and unresolved detail. Ask whether the broader view improves a decision, not merely whether its plot looks simpler. **Controls:** Use the same inputs and work allowance; include always-fine and fixed-interval inspection. Removing the observer's action connection should remove its influence while preserving its read. **Ready when:** Observer access, goal, request and resulting work are visible in structure. Passive observation does not change the participant. **Limit:** A useful view is relative to its task and access. It does not establish a universally privileged scale. ## 05. Add a third source and test echoes Status: Planned. Let observations guide work. **Build before this experiment:** Minds: generalize exposed publication links to a small three-space network. Reuse source/event identity, provenance and shared-source bounded quantities. Add only missing correspondence reads needed to distinguish a known shared source from an opaque source; do not invent a universal trust score. **Question:** Can several views combine without treating a repeated message as new independent support? **Try:** Compare three separate observations, one observation arriving by two known routes, and an opaque source with unknown ancestry. Include conflicting references, a broken link and unequal resolutions. **Observe:** Agreement, reference mismatch, source coverage, unresolved dependence, circulation of old messages and communication cost. Keep observation fanout distinct from material transfer. **Controls:** Disconnect links and compare shared-source versus genuinely separate sample events. Distinct identities alone do not establish statistical independence. Conventional controls retain equivalent dependency information. **Ready when:** Known aliases are preserved; opaque sources remain usable without fabricated independence; transferred material, if introduced, is not duplicated across branches. **Limit:** Correlation-based trust and source reliability learning remain separate future laws. ## 06. Settle, do less work, and wake again Status: Planned. Let observations guide work. **Build before this experiment:** Express a local goal, tolerance, pending work and wake condition through Gojo repeat/control structures. Add work counters and a small scheduler that follows exposed requests. If an attached recurrent world is used, expose settlement pass/budget evidence without changing settlement laws. Idle must not mean no unresolved structure exists. **Question:** Can a system reduce its activity when its goal is met and resume useful work when conditions change? **Try:** Long quiet periods interrupted by small and large changes, including changes hidden by a constant average. Start with a stated wake rule; learning that rule is a later condition. **Observe:** Read and update counts, messages, wake delay, missed changes and goal satisfaction over time. Report wall time separately; changes in values are not measurements of physical energy. **Controls:** Compare always-active and fixed-period polling with the same input stream and information access. Include a coarse wake rule that can miss fine activity. **Ready when:** A quiet state with remaining residue is distinguishable from an exhausted work allowance. New input can reactivate work through an explicit route. **Limit:** A successful authored wake rule is not evidence that Gojo learned an attention policy. ## 07. Keep competing predictions and useful memory Status: Planned. Build and reuse skills. **Build before this experiment:** Follow the existing signal-learning direction: connect candidate models to their issued predictions, later outcomes, support, cost and selection or mixture rule. Define those local update laws before measuring them. Add a memory port with source, time and resolution so remembered observations remain distinguishable from new measurements. **Question:** Can past usefulness change which local models contribute, while preserving alternatives that may become useful again? **Try:** Offer a small stated set such as hold-last and affine continuation. Change the signal regime, restore an earlier regime, and introduce gaps. First retain exact recent history; test deliberate compression only as a separate condition with a declared loss. **Observe:** Future error, coverage, switching delay, cost, discarded alternatives and recovery when an old pattern returns. Keep ambiguous candidates visible when observations cannot distinguish them. **Controls:** Compare each fixed model, equal contributions, and the same selection rule without history. A retrospective best model can be an evaluator-only upper benchmark, never participant input. **Ready when:** Support changes have an explicit action/outcome basis; candidates cannot learn from outcomes not yet received; loss from forgetting is visible. **Limit:** Selection among supplied models is learning within a chosen vocabulary, not unrestricted model discovery. ## 08. Combine skills and reuse the combination Status: Planned. Build and reuse skills. **Build before this experiment:** Reuse Gojo patterns, ports, transforms and dependencies to connect calibration, prediction and response. Add missing application and revision witnesses. A reusable shortcut needs a stated domain, retained source correspondence, an equivalence check and a path back to detailed work when its premises fail. Crystallization alone does not discover a replacement equation. **Question:** Can a useful combination become an exposed operation that works in another setting? **Try:** Compose two already tested operations, expose their combined input/output, then connect that operation to a third. Transfer it to different units, ranges and signal timing. Test inputs outside the promised domain as well as compatible ones. **Observe:** Transfer accuracy, adaptation work, construction cost, retained uncertainty and reactions to stale dependencies. Compare the full structure with a proposed shortcut before replacing anything. **Controls:** Compare rebuilding from primitives, using a frozen combination, and an ordinary implementation with equally reusable modules. Match the allowed interface adaptations and count the cost of discovering the original skill. **Ready when:** The combination remains readable after removing its helper class. Unsupported inputs remain unresolved or trigger explicit revision, rather than receiving a memorized answer. **Limit:** Begin with authored composition. Discovery of which skills to connect belongs to the later search experiment; a new wrapper name alone is not a new skill. ## 09. Sort with local rules and damaged parts Status: Planned. Test organization in another task. **Build before this experiment:** Add a reusable Gojo movement contract: occupied sites, legal neighbors, cell identity, atomic swaps, conflicts and ties. Define separately a part that cannot initiate movement and a part that cannot be moved. Pin the sorting study's implementation before claiming replication; a changed scheduler is a stated variant. **Question:** What ordering can local moves achieve, and how does damage change the reachable outcomes? **Try:** Start with small exhaustive arrays and one local rule. Add each defect separately, then larger cases. The runner schedules legal attempts; it must not sort a host array and write answers back into Gojo. **Observe:** Attempted and completed moves, adjacent descents, total inversions, source-study measures and unreachable goals. Retain full trajectories and cost, including failed attempts. **Controls:** Use a conventional implementation with the identical movement and defect rules. Check reachability in tiny cases. A temporary worsening in one measure may coincide with improvement in another, so it does not by itself demonstrate foresight. **Ready when:** Moves preserve cell identity and occupancy; conflict and immobility checks pass independently of whether sorting succeeds. **Limit:** This is the most direct link to the sorting work discussed in Ingressing Minds. It can branch after step 02 if that comparison becomes the priority; it does not depend on all memory and skill work being finished. ## 10. Mix local sorting rules Status: Planned. Test organization in another task. **Build before this experiment:** Extend the tested movement contract to a stated mixture of local rules. Add independent clustering measures, inert comparison labels and matched label-count randomizations. Keep the rule each part executes separate from labels used only by observers. **Question:** Do useful patterns appear when parts following different rules share one space? **Try:** Mix rules at several stated proportions, vary starting order, and repeat with each defect law. Inspect clustering, sorting, deadlock and temporary reversals of progress. **Observe:** Ordering and clustering over time, dependence on initial layout, work and the cases where mixtures perform worse. Distinguish a predicted pattern on new arrays from a pattern noticed afterward. **Controls:** Compare pure-rule arrays and identical rules with inert random labels. Match movement opportunities and schedule; labels must not become a hidden attraction rule. **Ready when:** The observer can detect clustering in a constructed positive check and avoid reporting it as a rule effect in the inert-label controls. **Limit:** Clustering alone does not establish cooperation, a shared goal or a new mind. Its contribution must be tested by intervention. ## 11. Repair and change connections Status: Planned. Let organization change. **Build before this experiment:** Define legal proposals to attach, detach or redirect exposed relations, plus local capacity, cost and ownership. Connect proposals to goals, accepted/blocked changes and later outcomes. Start with a small fixed edit vocabulary; general growth and decay laws are not yet established. **Question:** Can local connection changes recover useful behavior after damage, and at what cost? **Try:** Remove a link or participant from a working small network. Compare recovery through a spare route, local edit attempts, and combinations of previously tested skills. Include damage for which the allowed edits cannot achieve the goal. **Observe:** Recovery of the selected task, unnecessary edits, cost, oscillating reconfiguration and remaining disconnections. A local condition such as two neighbors per part is not proof of one connected whole. **Controls:** Compare no repair, a fixed repair rule and a random legal edit policy with the same information and work budget. Keep any central observer's extra knowledge out of local repair decisions. **Ready when:** Edits preserve structural contracts, do not create free conserved participation, and retain failed proposals. Later utility has a traceable connection to the change being assessed. **Limit:** Recovery must be measured against the task; merely restoring a familiar-looking graph is insufficient. ## 12. Compare ways of discovering useful structures Status: Planned. Let organization change. **Build before this experiment:** Use the legal construction and edit vocabulary from earlier experiments. Separate candidate proposal, evaluation and selection; record every attempted candidate and its cost. For a confirmatory comparison, freeze the heuristic definitions, task splits and assessment before evaluation. Neither replay nor a numeric compiler guard is required unless the specific claim depends on them. **Question:** Do the proposed heuristics find transferable combinations more efficiently than simpler searches? **Try:** Compare reuse of prior skills, search guided by unresolved local work, random legal proposals and a stated conventional search. Train or develop on one set of feedback, prediction or sorting tasks; evaluate on unused tasks and changed combinations. **Observe:** Useful new operations found per budget, held-out performance, transfer cost, failure rate and the cost of maintaining the search itself. Report the whole search, not just the best surviving example. **Controls:** Give each search the same construction vocabulary, observations, legal actions and evaluation allowance. Count prior skill-building work or report it separately. Test removal of each proposed advantage. **Ready when:** A run can end with no useful discovery and still be valid. A changed rule creates a new condition whose claim is tested on fresh cases. **Limit:** This tests practical research heuristics and structural learning. It does not distinguish metaphysical explanations that make the same observable predictions. ## Shared rules - Keep the question, selected rules, information access and observation window with each run. Save surprises and amendments; use unused cases to evaluate a revised claim. - Separate agreement, goal satisfaction, ongoing activity, unresolved work and the end of observation. A system can meet a goal while parts continue moving. - Use Gojo observers where practical and independent small-case checks to validate them. A broad observer may inspect more than an individual participant, but that access must be declared. - Decimal calculations are allowed at declared interfaces. Exact replay and compiler guards can wait until a particular question needs them. - Compare the same information, actions, resolution and resource budgets. Conventional exact arithmetic and conventional graphs can represent uncertainty and history too. Floating-point arithmetic is an additional condition, not a stand-in for all ordinary mathematics. ## Later, when needed - Exact continuation and replay when diagnosing a difficult divergence or promising a restorable checkpoint. Gojo numeric graph transport is a starting point; host queues and external inputs require their own capture. - Compiler-aware numeric boundary checks when making a claim about an entire calculation path. Keep legitimate sensor and external-tool adapters explicit rather than banning decimal interfaces. - Probability, learned reliability, continuous dynamics and richer geometry only when a question needs their additional laws. Bounds, tolerance, output resolution and confidence must remain distinct. - Preregistered confirmation once exploratory behavior supports a specific claim. Keep earlier runs and amendments; do not redefine success to rescue a disappointing comparison. ## Research context - [Ingressing Minds](https://www.mdpi.com/2409-9287/11/5/161). The broader prompt for studying behavior across implementations and observational scales; this roadmap proposes operational tests, not new axioms for Gojo. - [Classical Sorting Algorithms as a Model of Morphogenesis](https://arxiv.org/abs/2401.05375). The primary sorting study behind the damaged-part and mixed-rule comparisons. - [Sorting authors' implementation](https://github.com/Zhangtaining/cell_research). Inspect and pin a revision before a reproduction claim; preserve its scheduler and defect distinctions or label the changes. Generated from `ExperimentRoadmap.json` by `Gojo.Demos.Web/tools/build-minds.mjs`. Edit that source to keep the site and document aligned.