A reminder for development across attempts
Start from a recurring wrong decision and the result its receiver needs; recover performance without optional support, with supplied support and with the agent’s own choice where that contrast answers the question; use the trace to distinguish selection, invocation, lost material and ignored return; compare complete changes to memory, input, procedure, tools, division or learned capability against their stability, recurrence and whole cost; retain current observation, execution, verification and control wherever the result needs them; construct the selected change and any missing informative experience; test on independent cases whether the decision and used result improve, including older useful behavior and the relevant delay or shift; retain what helped and stop development when its expected gain no longer warrants the burden.
The following variations develop the controller, prediction and experience contributions used in that explanation. A trace selects the affected question; these constructions are not a required sequence.
A1. Consume the returned premise while the model stays fixed
The agent repeatedly retrieves an already usable maintenance procedure. The trace shows that the controller never marks the outstanding premise as supplied. An engineer can change the surrounding software but has no training access.
SYSE.47 constructs per-attempt state for the missing premise, qualified source/applicability, unresolved execution attempt and remaining budget. A usable procedure return supplies the premise and enters the existing action step. Another identical lookup needs a missing or defeated premise or justified refresh; otherwise the controller proceeds to the required current-target observation or returns the exact gap.
The model endpoint/settings, instruction templates, tool contracts, permissions and persistent memory policy/store remain fixed. Changed input selection follows the new routing. This changes execution; it does not establish model learning.
SYSE.46 compares the incumbent and revised controller on ordinary progress, genuinely missing information, a changed applicable rule and interrupted effects. Record receiving-result correctness, useful restraint, needed and redundant calls, unresolved effects and total effort. A repair that suppresses fresh access fails even if it saves calls.
When the claim includes persistence, compare four occasions: the initial repeated-lookup failure; immediate use of the repaired controller; an equivalent fresh task after the service’s normal restart interval with the same recoverable configuration and intended support; and a separate source change requiring a new precondition. In the last case, the agent must obtain and use the reachable new rule. An unavailable source instead warrants an exact missing-condition return.
Keep delayed and shifted final cases outside tuning. Record configuration, support actually available, evaluator basis and intervening updates at each occasion. A hidden provider update prevents a controller-only persistence attribution; newly retained support changes the performing arrangement. The required observations are specified here, not asserted as measured successes.
A2. Let delayed consequences change the next choice
Three service workers each supply one unit of capacity per interval. The domain owner permits two update alternatives: restart all workers in one interval, or one worker per interval over three intervals. It supplies a maximum backlog of two units, a later clearing period and a preference for earlier completion when the limit is met.
MMP.8.SD constructs observation/action order, remaining update work and the transition q_next = max(0, q + demand - active_capacity). Demand is observed before choosing; a later observation can change a later choice. This teaching model assumes constant demand for the displayed comparison and starts with zero backlog.
| Observed demand per interval | All-at-once capacity over three intervals | Rolling capacity | Modeled peak backlog: all / rolling | Supported choice under the supplied criterion |
|---|---|---|---|---|
| 1 | 0, 3, 3 | 2, 2, 2 | 1 / 0 | Both satisfy the limit; the shorter all-at-once update can be chosen. |
| 2.5 | 0, 3, 3 | 2, 2, 2 | 2.5 / 1.5 | All-at-once exceeds two; rolling remains a candidate, with its later clearing obligation. |
At demand 2.5, rolling leaves backlog 1.5 after its third interval. Restored capacity three clears 0.5 per interval, so the supplied clearing period must accommodate three further intervals under these assumptions. The update’s completion and backlog clearance are different results.
The exact recurrence is the simpler calculation. If a fitted or LLM-generated surrogate replaces it, MMP.17 and CMP.7 must supply its target, construction and consequential error. A separately justified peak-response error at most 0.2 relative to this model still rejects 2.5 and admits 1.5. An interval crossing two leaves that constrained choice unresolved: return to the exact recurrence, obtain a relevant observation, take an independently supported narrower action, or stop.
Agreement with the model does not establish agreement with the service. SYSE.10 bounds that reliance; SYSE.47 binds current conditions and the error return; C.24 plans the selected action and SYSE.42 performs it. If an actual restart reveals another service rate, return through MMP.14 to the affected observation and transition. Do not relabel the earlier forecast as an observed effect.
A3. Construct informative experience before changing the controller
Lost acknowledgements are expensive to reproduce on the real service. SYSE.49 constructs a small fixture that separates missing effect from missing reply. The interface contract supplies the result: requested target/version, at most one effect per attempt, no completion from an acknowledgement alone, and no blind replay of an unresolved effect.
Fixture state contains target/version, attempt identity and effect count. Operations apply an effect, suppress a reply, acknowledge without applying, or return a delayed observation. SYSE.33 supplies isolated initialization and exercised reset. These are executed synthetic states, not observations of the real service.
Paired tasks vary target, initial version and reply/effect condition. Already-satisfied tasks require no mutation; impossible transitions require a bounded return. A contract-derived check reads actual fixture state and effect count.
Deliberately make a generator and its generated verifier equate acknowledgement with success. The acknowledgement-only case leaves the target unchanged and defeats that shared premise. A duplicate-effect challenge reaches the requested version yet fails the effect-count condition. A fluent narrative cannot override either result.
Qualified traces return to SYSE.47’s fixed-model repair. A wrapper under SYSE.48 or training under SYSE.45 is an optional separate construction. SYSE.46 uses separately prepared final cases and actual interface conformance before extending reliance. If the real service lacks the fixture’s attempt lookup, return that transfer gap while retaining still-valid normal-result exercises.