Minds · project notes

Download Markdown

Preparing Minds for trustworthy experiments

Status: preparation review, September 13; priorities revised September 15, 2026. Inspected repository revision: 13596f2d4f353c952c90c7d9e7522bf6fab272a4.

Recommendation

The experiment roadmap is now the current suggested sequence, interspersing each experiment with its required development work. The numbered E0–E4 proposals and audit gates below remain background design notes; they are not additional prerequisites on that roadmap.

The feedback-limits experiment now adds the independent exact comparison, timing/resolution/reference controls and retained repetitions. These are selected exploratory checks; broader learning and replay claims remain open.

The prediction experiment now exercises prospective issuance and later assessment on twelve fixed sequences. It preserves losses against hold-last, missing outcomes and an independent exact comparison. The selected family and sample window remain authored; no learned selection was added to improve the outcomes.

Small experiments are already possible. Gojo has exact selected arithmetic, explicit read remainders, local feedback, prospective forecasts, and restorable numeric graphs. The first task is to make runs of those mechanisms accountable, not to finish a general learning system.

Start with the existing two-space response policies and make what happened visible. The first exploration now runs those policies, retains event evidence, and includes Gojo observers of the individual views, their average, disagreement and movement. Prospective prediction can follow.

The user's September 15 priorities supersede the prerequisite language in the original review below. Its broader audit gates remain future options:

• Decimal calculations and conversions are legitimate at declared interfaces to sensors, external tools and the world. A compiler guard is not urgent.
• Settlement can mean that a goal is met at a given scale while finer activity continues. Observe changes and remaining work, not just whether execution stopped. Activity reduction and crystallization into an equation are intended directions, not behaviors this first experiment implements.
• Exact replay is useful later. Different results across runs are information to examine. Retain enough observations and selected conditions to investigate them without requiring a perfectly reconstructible world now.
• Keep authored policies, participant responses, external inputs and observer reads distinguishable. This information can itself inform later settlement.
• Broader observations should also be expressible inside Gojo. A constant average and moving parts can both be valid readings. External checks help validate those observers; they need not own the system's understanding.
• Describe rules and amendments now; use full frozen protocols when appropriate for confirmatory experiments. Surprising valid behavior invites investigation.

Here, neutral means that laws, information access, policies, and assessment rules are explicit and applied consistently, including when results disappoint. It does not mean a system without assumptions. Locality, selected resolution, available operations, and a finite resource budget all affect what it can do.

Three questions must remain separate:

1. Did the mathematics and software follow their contracts?
2. Was the experiment valid under its declared conditions?
3. What did the experiment show about the hypothesis?

A failed numerical invariant invalidates a run. A correctly executed run that does not learn, converge, or outperform a baseline is a useful scientific result.

What we are taking from the paper

Levin's *Ingressing Minds* proposes studying mathematical and behavioral patterns across embodiments, including unfamiliar systems and different observational scales. Section 6.3 and Figure 14 discuss sorting behavior, temporary losses in sortedness, and clustering of cells using different sorting rules. Section 7.3 proposes a research program and judges its value by the investigations and capabilities it enables. The paper reports no new dataset.

For this lab, those are prompts for controlled comparisons, not additional Gojo axioms. Our proposed questions are:

These are operationalizations proposed here. The experiments below do not distinguish nonphysical causation from an ordinary account of the same executed rules. Such a claim would need competing models that predict different observations under a specified intervention. Gojo can help expose the rules and evidence; naming the project Minds adds no such prediction.

Current capability and its limits

The signal learning plan already points toward competing continuations, ambiguity, composition, driven systems, and later response/utility. Minds should exercise that direction, with stricter experimental records, rather than introduce a competing theory.

Six areas to develop (broader audit, not prerequisites for exploration)

1. A declared and checked numeric boundary

Finding. A source search of Gojo/**/*.cs found no executable uses of the searched floating/decimal types or transcendental calls. This is a preliminary source audit, not a compiled dependency or data-flow proof.

The same is not true across the solution:

• GearShapeGeometry uses double, Math.PI, trigonometry and roots, then rounds coordinates in FrameBasis.Point before constructing Gojo points.
• GraphicAreaSettlementSurface uses floating-point geometry and fixed numerical thresholds to select and construct area results. These calculations are not just pixels.
• GraphicOrthogonalFillSurface also has floating-point geometry and rounded construction paths.

These are relevant limitations of those paths, not evidence that the core feedback/forecast fixtures require floating point. Start Minds with a direct dependency on Gojo only. Do not use the gear demo as evidence for an entirely Gojo-derived geometry calculation until its actual path has been addressed.

Later audit work:

1. List the participant's builders, readers, mutators, dependencies and input adapters. Record every place a host calculation can affect a decision.
2. Add a compiler-aware guard for this declared closure: primitive and aliased floating types, decimal, Half, inferred floating literals, conversions, generic math and calls into unreviewed numeric helpers. Include generated code and transitive project/package changes. Reflection/native/dynamic paths need explicit review; an unknown path cannot receive a clean certificate. Define a small trusted boundary for exact runtime operations such as integer arithmetic and the used BigInteger operations. Do not reject an entire framework assembly merely because unrelated APIs in it support floating point.
3. Test the guard with deliberately contaminated examples, including a helper that returns an integer after internally rounding a floating result.
4. For each scientifically relevant calculation, retain the Gojo inputs, chosen basis, operation/read, result and remainder. For each decision, identify its structural rule or label it as a selected host policy.

A token search is a useful alarm, not the final guard. A type guard also cannot detect a hidden optimizer written entirely with integers: the structural derivation review is a separate requirement.

Allowed exact implementation work. Gojo uses ordinary host integer instructions, Int128 and BigInteger to implement justified structural reads. ExactScalar is an internal exact rational calculation aid; it encodes results back into grounded relations. Reimplementing these integer instructions in a second arithmetic engine would not improve experimental neutrality. The important distinction is whether the connected premises justify the result, not whether C# ever performed arithmetic.

Decimal notation is also not the issue: an exact input 1/10 may be legitimate structural participation. It is different from silently computing an approximate result in a host numeric type and inserting only the answer into Gojo.

Gate. An intentionally contaminated path is rejected; unsupported exact operations stay unresolved; changing a plot or reference evaluator cannot change participant actions. External measurements may enter through a declared adapter with stated resolution and uncertainty. They must not be relabeled as internally derived Gojo answers.

2. Explicit update order, settlement and observation limits

Finding. Carrier.Add, Remove, Set, and Attach settle immediately. SettleWorld runs at most 256 passes, compares carrier value/tension snapshots, and returns without reporting whether it reached the cap. Carrier identity order and relation traversal affect the execution order. Some route allocation also orders candidates by relation identity in Carrier.RouteSettlement.

Consequences:

• A returned call does not prove convergence, zero tension, or goal completion.
• Equal snapshots at pass boundaries do not prove that nothing moved within the pass. They also do not compare every structural or host state component.
• Several sequential perturbations are not a simultaneous update.
• Reordering construction to obtain favorable identity order is a policy change, even if the final graph drawing looks the same.

Build in two steps. For the first response experiment, use the existing free-actuator contract and an explicit event sequence: publish A, publish B, deliver the captured observations, then apply each response. Record the chosen order and retain its difference from sequential publication or delayed delivery. That experiment does not require a new global scheduler.

Before studying recurrent attached worlds, add a small general settlement report with pass count, changed-state evidence and distinct quiet/capped/error outcomes. Define what state its stopping check covers. Preserve unresolved tension; do not call all remaining tension failure. A later one-step/batched interface must distinguish receiving changes, local absorption, route settlement and time. Adding an observation/report must not secretly change those laws.

Gate. Demonstrate a quiet result with unresolved work, a budget stop, a declared oscillating example, and order-sensitive behavior. Repeat a declared event sequence exactly. For cycle detection, include all state affecting future transitions, including messages and controllers; repeating one scalar is not proof of a cycle. If complete cycle evidence is unavailable, report a suspected repeat or budget limit, not an established attractor.

3. Complete run records and replay (deferred)

Finding. The numeric graph format is a strong starting point, but the lab does not yet have a complete run bundle or general replay runner.

For a future exact-replay mode: a small versioned host record containing protocol/code hashes, runtime and dependency versions, exact initial roots and graph bytes, selected numeric basis/resolution, input tape, event order, interventions, budgets, and all recorded outcomes. If there is randomness, retain its generator/version, seed/state and realized draws or event tape; a seed alone is not portable replay. Record setup actions too: attaching a relation may already cause settlement.

Store graph snapshots with GojoNumericFormat, not a new serialization of participant concepts. JSON or similar metadata is appropriate for run hashes and filenames; the graph remains the participant's source of meaning. Store immutable bytes at capture points, not live references presented as past state.

Gate. Restore into a fresh world/process, continue from a checkpoint with messages in flight, and obtain the same declared observations and actions. Use an explicit identity correspondence for graph comparison. Byte equality of separately restored graphs is not required when process identities change. Changing original live objects must not rewrite stored evidence. Audit roots for detached recipes, open slots and pending messages; omitted state must not be reconstructed from a remembered expected answer.

Keep exact experiment archives separate from any later deliberately lossy memory model. Successful replay tests only what the archive actually retained.

4. Separation of participant, environment, policy and evaluator

Finding. The present fixtures explicitly select reader calls, response strengths, candidate families and event order in host code. That is useful experimental control. It becomes misleading only when described as something the participant learned or discovered.

Build: an information-access table for each experiment. Identify what each participant can observe, which actions it can take, which private internals are unavailable, and which decisions are authored. The runner can schedule declared actions but must not search for the best next action on the participant's behalf.

For example, a fixed half response is an intervention parameter, not learned confidence. A supplied affine family is a hypothesis class, not discovered mathematics. An environment may contain an unreachable target. An evaluator may know that target without giving it to the participant. If reward is later fed back, make its source, timing and response route explicit.

Gate. Refuse future observations, stale plans, duplicate consumption and lookalike identities; test that a participant cannot consult the reference answer through a callback or shared mutable object. Reordering diagnostic observers must not create new actions. Rename incidental labels without changing behavior. Separately test changes to meaningful relation order; that order is not an incidental label.

5. Gojo observers, independent checks and fair comparisons

The current priority is a hierarchy of local and broader Gojo observations. The two-space observer now calculates average, disagreement, changes and movement using existing bounded-quantity reads and calibrated composition. None of those readings is a universal verdict. A coarse stable result can coexist with finer movement or disagreement. Independent checks validate these reads without requiring all future assessment to live outside Gojo.

Finding. A conformance assertion and a scientific measurement serve different purposes. Existing unit tests establish many local contracts; they do not decide which experimental score fairly measures learning or agency.

Build: read-only measures with definitions fixed before evaluation. Retain raw exact counts, signed mismatches, nonnegative losses, missing outcomes and measurement denominators. Do not let positive and negative errors cancel into an apparently perfect predictor. A system that answers fewer cases must show that coverage difference. Open, incompatible and unavailable reads are not zeros.

Use small exhaustive cases or independently derived exact arithmetic to check measurements. An evaluator that invokes the same production reader is useful for consistency, but is not an independent correctness oracle. Statistical summaries and floating-point plots can be produced afterward in a named evaluation boundary, with retained exact inputs and no feedback into the run.

Compare at least:

• The declared Gojo mechanism.
• A conventional exact integer/rational implementation with the same rules, observations, retained state and schedule. This checks behavior without a floating-point confound.
• A conventional implementation with the information bookkeeping Gojo makes explicit, when the claim concerns representation or composability.
• A controlled removal of one Gojo facility when that tests the hypothesis. Hiding a remainder from the evaluator and physically discarding incoming change are different interventions and must not be conflated.

Where useful, add a floating-point implementation as another declared condition. It tests the effect of that approximation, not all of ordinary mathematics. Conventional mathematics can represent exact fractions, graphs, uncertainty and history too. Gojo's possible advantages include inspectability, reliable composition, explicit unresolved states and cost; none requires predicting a different answer to the same fully specified exact calculation.

Gate. Controls share initial problems, admissible actions and information. Record local work, messages, memory and wall time separately; counting only swaps or successful changes may hide expensive searches or failed attempts. Count experimental runs as independent units, not every correlated step of one trajectory. Predeclare how multiple measures and parameter comparisons will be reported; retain unfavorable cases and resource costs.

6. A rule for changing the rules

Finding. Preparing an experiment will expose missing mechanics. The risk is not adding a useful mechanic; it is repeatedly adding just the mechanic needed to rescue the current example and then calling the outcome emergent.

Build: use the protocol template with a version identifier. Record the question, selected laws, parameters, candidate space, metrics, controls, budgets, exclusions and both favorable and unfavorable outcomes before confirmatory evaluation. Record exploratory runs as exploratory.

If a missing law appears:

1. Save the failed or unresolved case.
2. Classify it: software defect, absent operation, poor chosen policy, invalid premise, or legitimate negative result.
3. For a proposed kernel addition, write its Gojo shape and independent law. Include cases unrelated to the desired experimental outcome, especially refusal, conservation and range boundaries.
4. Version the code and protocol. Results across versions belong to different conditions; do not silently combine them.
5. Evaluate the revised claim on unused cases. Do not keep refining against the same supposedly held-out examples.

Gate. The experiment can finish as valid with an unfavorable result. There is a complete ledger of attempts, including stalls, invalid runs and amendments. Unexpected patterns can motivate a new hypothesis, but their discovery trace does not also serve as its independent confirmation.

Proposed experiments and their prerequisites

E0. Establish that the lab measures honestly

This is apparatus validation, not a test of cognition. Reuse the existing projection, bounded-quantity and restoration fixtures as specifications. Exercise observation without mutation, source-basis reconstruction, same-source versus distinct-source uncertainty, refused fractional import, overflow and checkpoint continuation. Include deliberately corrupted histories and a deliberately contaminated numeric helper.

For the future strict apparatus audit: the numeric guard, capture/replay boundary and independent checks. These are not gates on the first exploratory run. Do not copy the entire kernel test suite into Minds. Add only tests of the new lab boundary and failures that the original tests cannot catch there.

E1. Agreement, disagreement and oscillation in two spaces

Start here with focused checks. Reuse the current local response/exchange contracts. Freeze policies with response fractions 0, 1/4, 1/2, and 1; implement zero as a declared no-response control if the existing response reader refuses zero strength. Do not change that reader merely to admit the control.

Use a small predefined set of starting discrepancies, exact basis conversions, consistent/conflicting local references, missing messages and delivery orders. The current selected law already predicts agreement in the captured half/half case and swapping in the full/full case. Reproducing those outcomes validates the apparatus, not discovery of intelligence. Quantization may preserve a nonzero disagreement; delayed delivery can change trajectories.

Measure disagreement, mismatch to independently supplied reference observations, unapplied change, messages and attempts. Peer agreement and agreement with the reference are separate measures, and reference observations are still premises.

For initial exploration: retain observed events and explain the selected local attempt limits. Exact replay and a numerical dependency guard can follow later. A general network scheduler, automatic trust and autonomous rewiring can wait.

E2. Forecasts that can be wrong

Use the current prospective forecast path with a frozen affine family and readout. Compare against hold-last on finite exact sequences: constant, ramp, reversal, missing observations and unequal sample intervals. A hold-last participant needs its own small justified structural construction if the existing temporal API cannot express it; do not sneak it in through a scalar fallback. A conventional hold-last reference can be evaluated externally now.

Freeze the selection of past samples. Issue every forecast before its outcome is available. Measure exact absolute error, coverage and cost over the same targets; retain signed mismatch separately. Later extend to competing families along the existing signal plan. A model that fits past observations can still fail prospectively, and that is an intended result.

Required: E0, complete issue/outcome records and a fixed loss. Automatic family selection, confidence learning and reward attribution are not prerequisites.

E3. Same behavior, different observations

Replay the same frozen trajectory through several predeclared read surfaces. Ask which view predicts a future event or helps choose a useful intervention on unused trajectories. Start with exact scalar/time reads and no geometric solver. Count complexity of the observer as part of the proposed explanation.

Required: immutable snapshots, separate observational and intervention phases, and at least one task on which the proposed read could fail. Additional views must not mutate the participant. A different description of a trace is not by itself evidence of a new skill or of a privileged perspective.

E4. Sorting and mixed local rules

The underlying sorting study distinguishes cells that cannot initiate moves from cells that cannot be moved at all, uses threaded local implementations, and measures sortedness and clustering. Consult its authors' implementation and pin a revision before attempting replication. A changed scheduler or defect law is a variant, not automatically a reproduction.

Gojo can already represent sites, relations, comparisons and histories. A tested local movement contract is still needed: occupied sites, legal neighbors, ordering, movement of cell identity, swap atomicity, defect behavior, tie rules and conflict handling. These must remain visible without a host-side array sorter that writes its answers back into Gojo.

Use a conventional exact implementation first as a reference. Small exhaustive arrays can expose unreachable target states under the declared move set. Preserve all attempted and completed swaps and the complete trajectory. Record the paper's measures alongside total inversions and adjacent descents, with duplicates and boundary conventions fixed in advance. Control clustering with identical rules carrying randomly assigned inert labels, preserving label counts; rule labels must not be input to movement decisions.

An independent arithmetic example shows why the interpretation needs care:

[2, 4, 1, 3] -> [2, 1, 4, 3]
adjacent descents: 1 -> 2
total inversions:  3 -> 2

Swapping one inverted adjacent pair worsens one measure while improving the other. A dip followed by recovery therefore does not establish foresight or planning. Test the operational claim about the trajectory first; stronger claims need additional memory, goal or intervention controls.

Required: E0 plus an explicit movement/scheduling law and independently checked measures. Do not require a positive sorting result before freezing those laws.

Later: transferred skills, repair and evolving connections

These remain worthwhile, but should not be included in the first readiness gate. They require explicit proposal/selection rules, legal structural edits, local resource costs, action/outcome correspondence and credit over time. Memory compression requires a declared retained promise and measurable loss.

For transferred skills, freeze the reusable operation and permitted interface adaptations before trying a new context. Compare transfer with reconstruction under equal budgets. For repair, distinguish restoring a local condition from achieving a global one: degree two at every site does not imply a single ring. For heuristic comparisons, give search policies the same construction vocabulary and evaluation access, count all candidates tried, and use a held-out set of tasks. A heuristic is an experimental condition, not a kernel law.

Geometry-heavy and continuous-dynamics experiments have additional requirements: grounded constructions, supported operations, explicit approximation bounds and resolution studies. An unsupported root or transcendental expression must remain unresolved or gain a separately justified refinement mechanism. Finite ticks alone do not supply exact real analysis.

Current implementation order

1. Run the existing two-space fixture and retain observations (implemented).
2. Observe at both local and pair scales through Gojo structure (implemented).
3. Show messages, changes, plan residue and the observation window (implemented).
4. Explore unexpected behavior and additional schedules or goals.
5. Add stronger auditing, replay and formal protocols when a question needs them.

The earlier audit work breakdown remains useful as a backlog, not a dependency chain that must finish before experiments:

Minds.Tests now checks the new observer and driver boundaries, including observation without mutation, retained past readings and fractional requests. Keep experiment reports separate from test pass/fail: conformance tests assert contracts; scientific runs record what happened. A narrow authored event tape is enough initially. A universal scheduler, confidence algebra, distribution library, advanced geometry, automatic rewiring and a definition of consciousness are not blockers for E1 or the existing affine forecast experiment.

Review evidence and reproducibility

The baseline solution run on September 13, 2026, using .NET SDK 10.0.401, passed 1,653 tests: 782 Gojo, 695 encoding, 155 CAD, and 21 gear-ladder; zero failed or skipped. These are dated inspection results, not a permanent inventory or evidence for an unimplemented Minds runner.

Commands used or useful when updating this review:

dotnet test Gojo.slnx --verbosity minimal
rg -n '\b(double|float|decimal|MathF)\b|Math\.(Sqrt|Pow|Sin|Cos|Exp|Log|Round)' Gojo --glob '*.cs'
rg -n '\b(double|float|decimal|MathF)\b|Math\.(Sqrt|Pow|Sin|Cos|Exp|Log|Round)' GojoCAD Gojo.GearShapeLadder --glob '*.cs'

The search is deliberately documented as a preliminary audit: comments may match, and aliases, inferred types and indirect calls need stronger analysis. The current code/tests take precedence over earlier dated status counts.

Paper review used the published 53-page publisher PDF, especially sections 6.3 and 7.3, and visually revisited Figure 14 on printed page

1. Its inspected bytes have SHA-256 D347F3B9043C49D07DB72D1C9F2F265C8D380C23292524FC82DE48158FA017A3. The publisher web page was rate-limited during this pass; the locally retained publisher PDF from the earlier review was used. The sorting article is a separate primary source, not new data supplied by Ingressing Minds.