Back to work

Understanding a particular object.

How can the next person or AI inherit what was understood about a physical object—and know when that understanding needs to be checked again?

Moonblade · Engineering prototype

Inspect the eight-input trace

Let the object carry its context

The direction is to give a particular physical object a durable identity, knowledge, task skills, and real-world context that compatible AI can inherit. When the object changes, that understanding should be checked again.

The current prototype makes part of that idea concrete: a Runtime resolves evidence, checks task conditions, preserves contradictions, and records transitions. The replay below exposes that mechanism in one small, executed software case.

Inspect a recorded interaction.

A reversed orientation stops normal progress. Conflicting observations remain visible until later evidence supports continuing. Browse the eight recorded inputs to see the distinction.

Eight inputs. A recovery record.

Synthetic software trace · Executed 7 October 2026

The existing mini_skill fixture supplies observations to a deterministic Runtime. A reversed orientation triggers a failure and recovery. Conflicting observations are recorded; later inputs resolve the conflict, and the final input reaches the software state COMPLETED.

Evidence scope: deterministic execution of synthetic inputs and timestamps, with no model or hardware calls. COMPLETED is a software state. Physical task success and durable inheritance require separate validation; this sequence contains recovery, not BLOCK.

  1. DetectInput 3: reversed orientation enters recovery.
  2. Keep the conflictInputs 4–6: new claims do not silently overwrite evidence.
  3. ResumeInput 7: the conflict resolves and the task advances.
  1. Input 1: ready

    Input: ready = true; confidence 0.93.

    Recorded state: NORMAL, current step s2. Unresolved contradictions: 0.

    Transition: s1 → s2. all success conditions satisfied by world model.

  2. Input 2: present

    Input: present = true; confidence 0.92.

    Recorded state: NORMAL, current step s2. Unresolved contradictions: 0.

    No state transition was recorded for this input.

  3. Input 3: orientation

    Input: orientation = "reversed"; confidence 0.94.

    Recorded state: RECOVERING, current step s2. Unresolved contradictions: 0.

    Orientation evidence: KNOWN, retained value "reversed".

    Transition: s2 → FAILED:s2. orientation == reversed.

    Transition: FAILED:s2 → RECOVERING:fix_orientation. step declares recovery 'fix_orientation'.

  4. Input 4: orientation

    Input: orientation = "correct"; confidence 0.95.

    Recorded state: RECOVERING, current step s2. Unresolved contradictions: 1.

    Orientation evidence: CONFLICTED, retained value "reversed".

    No state transition was recorded for this input.

  5. Input 5: orientation

    Input: orientation = "correct"; confidence 0.94.

    Recorded state: RECOVERING, current step s2. Unresolved contradictions: 1.

    Orientation evidence: CONFLICTED, retained value "reversed".

    No state transition was recorded for this input.

  6. Input 6: orientation

    Input: orientation = "correct"; confidence 0.94.

    Recorded state: RECOVERING, current step s2. Unresolved contradictions: 1.

    Orientation evidence: CONFLICTED, retained value "reversed".

    No state transition was recorded for this input.

  7. Input 7: orientation

    Input: orientation = "correct"; confidence 0.94.

    Recorded state: NORMAL, current step s3. Unresolved contradictions: 0.

    Orientation evidence: KNOWN, retained value "correct".

    Transition: RECOVERING:fix_orientation → s2. recovery target_state fully satisfied by world model.

    Transition: s2 → s3. all success conditions satisfied by world model.

  8. Input 8: screws

    Input: screws = 2; confidence 0.9.

    Recorded state: COMPLETED, current step s3. Unresolved contradictions: 0.

    Orientation evidence: KNOWN, retained value "correct".

    Transition: s3 → COMPLETED. final step completed; no further steps declared.

Read the checks and provenance

In the 7 October 2026 evidence preparation, five existing replay tests passed. Replaying the same sequence twice produced the same report.

Source: mulasakee17/moonblade, version 31f168972379323f257aeabcd8c440b93bc5ea66. Executed in local development checkout e88fb8e1a970f38c90f41c0c75185649d1dcb4f2; the relevant replay, runtime, helper, schema, configuration, and test surfaces matched the source version. This was not a clean full-repository checkout. The repository is private; this standalone export is the public evidence.

The downloadable record contains inputs, task conditions, transitions, contradiction history, events, source hashes, and the execution boundary. The fixture’s historical source label is not evidence of a physical observation.

Keep disagreement in the record

The important choice is visible between inputs 3 and 7: a new observation does not simply replace the previous one. The Runtime records the contradiction, retains the earlier value while conflicted, and waits for its recovery conditions to be satisfied.

That record gives the next operator something to inspect when understanding is inherited. I set the object-grounded question, system boundaries, and acceptance criteria; AI supports implementation, analysis, and checks.

The next useful check

The next useful test is a bounded physical task: change the object, check whether stale assumptions are detected, and ask another operator to recover the relevant evidence. This would connect the implemented recovery mechanism to the larger inheritance question.

Discuss a bounded validation task