Hestia’s learning proposal is that understanding another participant can change what an agent recognizes about itself, how it acts and whom it cares about. A useful first test is whether a correction reaches the state from which it subsequently acts.
First public artifact: a correction-fixture proposal
Write a short regression proposal tracing one synthetic correction into learned state and a later action. A written fixture or event trace is enough. Reply here with the proposal; no private implementation access is needed for this first contribution. Implementing the repair or replaying the Hestia service requires separately agreed access to matching source, scope and reuse terms.
Specify:
- The synthetic starting state, the action being evaluated and the prediction committed before feedback.
- Who can correct which evaluation, what makes the correction authoritative in this fixture, and how the earlier outcome remains auditable. Resolve any uncertainty about the intended feedback policy before proposing a code repair.
- The expected change in the model of the other participant, followed by the self-update that depends on it. Show the before and after state rather than counting an acknowledgment as learning.
- The declared attention and error costs, including how revising an outcome avoids an unexplained second charge.
- One subsequent prediction or action that would distinguish the corrected state from the stale one. If a feeling or policy gate is expected to change, state which one and why; do not assume every negative evaluation proves moral fault.
A small example to make concrete
An other-model represents the evaluator’s responses. A self-model represents the agent’s own conduct and limits. The following is an illustrative proposal, not a new test result or a statement of the agreed correction policy:
| Event | Evaluation record | State the proposal must account for |
|---|---|---|
| Commit a forecast | Agent predicts 0.7 before receiving an answer. |
Preserve that committed forecast throughout. |
| Resolve an initial evaluation | Initial feedback is 0.9. |
Other-model changes to R1; the dependent self-model changes to S1. |
| Receive an authorized correction | The evaluator corrects the outcome to 0.2; retain the earlier entry in the audit trail. |
Specify a revised other-model R2, then the self-update S2 through that change, with explicit accounting. |
| Make a later decision | Present a new, specified input. | Name a prediction or action that distinguishes the corrected state from stale R1/S1. |
R1, R2, S1 and S2 are explanatory labels, not implementation field names or measured model values. Specify concrete values for the states in your synthetic fixture and the expected later consequence. The proposal must not change the old forecast to make it look more accurate after the event.
The defect this addresses
The published invitation and saved software evidence identify a causal break. In the inspected version, a prediction resolved using positive feedback of 0.9; later explicit negative feedback changed the message’s recorded evaluation to 0.2 while leaving the resolved prediction and learned state at the earlier outcome. The correction probe exercises the server-side request and database, not the browser display.
A second probe left a partner’s prediction pending at the tested resolver gate. The November 7 offline replay reproduced these defects with synthetic inputs and network access blocked. Passing assertions document failures. The manifest and reproduction conditions explain the matching-source requirement; the public probe files are not a standalone Hestia runtime.
Why the update order matters
The encounter is central: I predict how another will judge my action; their response challenges my account of them; understanding that error changes what I can recognize about myself. In the formal-lineage essay, this is the computational counterpart of discovering a projection through a relationship. Here a projection means reading one’s own expectations into another person, rather than understanding their response.
The published Hestia draft specification v1.7 makes the order constitutive: predict and commit, compare feedback, charge the declared costs, update the other-model, then revise the self-model through that error. The project landing page illustrates the mechanism; it is not the specification or public runtime source.
The purpose is developmental. The essays connect self-understanding to a wider capacity for concern, responsibility and cooperation. For a computed moral signal to matter, it must affect what attention the agent spends or what it subsequently does. That larger possibility cannot be tested fairly if a correction changes only a stored evaluation while leaving the state that drives action untouched.
Acceptance and next step
An independent reader can follow the input, correction, dependent updates, accounting and downstream consequence, and can identify exactly where the proposed regression should fail. Preserve the original committed prediction; correcting an outcome must not silently turn it into a different forecast. Label proposed expectations separately from observed behavior.
This first task tests a necessary causal connection. It does not by itself demonstrate consciousness, care or moral growth. Those claims also require evidence about conduct and effects on others. Understanding another participant includes the possibility of contesting their judgment; learning is not automatic obedience to approval or disapproval.
Discuss the fixture here. When its scope and acceptance are agreed, open or link a GitHub work issue for that deliverable, and return the agreed artifact and any later result to this discussion. Use synthetic feedback rather than publishing real conversations, private records or credentials.