Minds · project notes

Download Markdown

Two-space feedback: first exploration

September 15, 2026. This is an exploratory walkthrough documented with its results, not a preregistered scientific comparison.

What the spaces do

A and B estimate the same quantity from different local reports and calibrations. A initially predicts 5. B predicts 18 in its own unit, equivalent to 9 in A's unit. The link explicitly maps two B ticks to one A tick. The reports, subjects, clocks and calibration actuators remain distinct.

Each space publishes a copied observation. Its peer receives that observation through the link, compares it with its own prediction and requests a local actuator change. Existing LocalCalibration readers do this work. No host code calculates the next estimate or supplies an answer such as seven.

The host selects the response fraction, message order, and whether to use the existing local reference. These are authored conditions, not learned policies. The source fixture and its six-attempt controller are reused from the existing exchange fixture.

What is observed

The same state has several simultaneous observations:

• Each space's current estimate and last published message.
• The pair's average and signed disagreement, B minus A.
• Each local change and the sum of their absolute magnitudes.
• Change in the average.
• Each estimate's mismatch to its separate reference observation.
• The latest exact actuator request, whole change, retained fractional request, carrier tension and number of applications.

All pair quantities use A's comparison unit. Displayed actuator requests use each actuator's native ticks. Fractional plan residue is not automatically carrier tension, confidence or uncertainty about the sensor's truth.

The observer uses the existing BoundedQuantity affine reads, calibrated composition, and distance-to-zero reads. In Gojo notation its construction is:

native estimate -> exact affine observation
B observation -> authored B-to-A calibration -> common-unit observation
ports(A observation -> half scale,
      B observation -> half scale) -> P(comparison,1)      average
ports(A observation -> negative unit scale,
      B observation -> unit scale) -> P(comparison,1)      disagreement
previous/current observations -> signed change reads
change reads -> distance from zero -> sum                 movement
before observation -> event evidence -> after observation timeline

The selected sources, maps, results and remainders remain connected if the C# observer helper is removed. Its read does not act on the participants. It has access to both local prediction witnesses for this diagnostic experiment; it is not pretending to be a third isolated participant or a universal observer. There is no observer feedback route yet.

The viewer can inspect each individual event or only the state after both responses. Movement is calculated relative to the previous observation in the selected view. Consequently, the broader view can show a constant average while the fine view shows intermediate shifts as A and B act in order. Neither view makes the other false.

Selected runs

Each condition begins with a fresh parse of the same fixture and runs for six observed rounds. The driver records all six, including zero changes. There is no success-based early exit and no assertion that the participant has stopped.

The quarter case has not become completely quiet within this observation window. Its unequal native resolutions and whole-tick actuator responses also allow the average to drift. That is a result worth inspecting, not something the driver corrects. A longer run would need a separately declared larger attempt controller; exhausting the existing controller is not the same as reaching an intrinsic equilibrium.

What this tells us so far

The broad observation is already useful: unchanged average does not imply unchanged parts, and local peer agreement does not establish agreement with another source. Conversely, disagreement does not itself prove bad arithmetic.

The message-order comparison shows why it helps to retain what each side actually received. The full-response rules can give swapping or agreement depending on that schedule. None of these examples discovers the response policy or establishes a general convergence law.

The two reference values both convert to 2, but they are supplied premises, not an external guarantee of truth. Selecting them in round 3 is an explicit intervention. The example does not yet learn when peer agreement deserves closer inspection or when a reference deserves more weight.

The current driver also continues making selected reader calls after numerical agreement, and the existing response history counts zero applications. It does not implement reduced computation/energy, adaptive waking, or automatic replacement by a stable equation. Those are sensible next questions, and should be made visible as changes in activity and retained structure when implemented.

Running and viewing

dotnet run --project Minds -- Minds/results/two-space
dotnet test Minds.Tests/Minds.Tests.csproj

The step viewer reads exported observations, not a browser version of the Gojo algorithm. The export is a diagnostic projection; it is not an exact replay checkpoint. Keep another run by choosing another output directory. Decimal conversion is confined to display coordinates. No decoder or renderer feeds values back into the spaces.

Focused tests check the observer's arithmetic against simple independent expectations, observation without mutation, copied earlier readings, duplicate response refusal, and retained fractional requests. These tests guard the selected implementation contracts; they do not require every future experiment to produce agreement or the same result.

Useful next variations are a changed local reference, delayed or missing messages, and different local response resolutions. A later Gojo observer could use the difference between coarse quiet and fine activity to request a closer read. That request and its goal would need their own explicit connection.