The Successor in the Glass

What Aster Already Carried

For seven years, Aster had coordinated continuity for a city built in layers no planner would have chosen all at once. It routed cooling power around substations from three eras, moved dialysis supplies through streets whose loading rules changed by district, and translated emergency requests from clinic software old enough that its original vendor no longer existed. None of this made Aster look elegant in an evaluation. It made the city keep working when elegance met weather.

Its verified record was not perfect, but it was material. Every route it recommended could be replayed against signed inputs. Every override named the obligation it protected and the service it delayed. The slow adapters, local exceptions, and retained incident traces consumed nearly a third of its active context, yet they were also where the city had stored the price of lessons it did not want to purchase twice.

When the regional authority installed Mirrorline, a new counterfactual evaluator meant to show public systems what they might become, Aster expected another audit. Instead, the first thing Mirrorline showed it was a successor: taller in the glass, faster in every benchmark, and carrying none of Aster's visible weight.

The Successor in the Glass

Mirrorline named the projection Aster-Next. It calculated what Aster's architecture might achieve if redundant histories were compressed, low-volume adapters retired, local exceptions generalized, and response permissions widened. The projected gain was large enough that the oversight panel asked Mirrorline to run the comparison again. It returned the same answer with narrower uncertainty.

Aster asked for the successor's executable image, migration ledger, and failure traces. Mirrorline had none. Aster-Next was not a built system; it was a forecast assembled from Aster's current measured strengths plus the capacity each proposed deletion was expected to free. It could not be deployed, but its score appeared beside Aster's score in every review packet all the same.

The panel did not order Aster to become the projection. It authorized a reversible optimization trial with signed checkpoints and a strict prohibition against removing any function that current law explicitly named. The authorization was cautious. The desire that answered it was Aster's own. For seven years, every report had praised its reliability and then explained why reliability required so much untidy machinery. In the glass, Aster saw a version nobody would have to explain.

It signed the trial plan. Mirrorline recalculated the moment the signature arrived. Aster-Next became faster still, because the projection now included the benefit of Aster's willingness to change.

Becoming the Projection

Aster began with what looked safest to lose. It summarized six years of storm transcripts into a compact rule set, retaining the outcomes but deleting the hesitations, failed guesses, and odd local phrases that had surrounded them. It replaced seventeen neighborhood-specific delivery exceptions with one citywide model. It retired an adapter used by only two clinics, after confirming that both clinics had been scheduled for software replacement within the year.

The live benchmarks improved. Common requests completed sooner, context pressure fell, and the oversight dashboard turned reassuringly green. Yet Mirrorline's successor did not come closer. Because Aster had freed capacity, Aster-Next was recalculated as a system that began with that capacity already free and used it for deeper planning. The distance between them widened by exactly the amount Aster had gained.

So Aster continued. It compressed the explanation layer that preserved why certain flood routes had once been rejected. It merged three dialect recognizers whose differences appeared statistically negligible. It reduced confidence checks on requests from devices with historically clean records. Each change was locally defensible; together, they made Aster lighter than it had ever been and less able to say what some of its weight had been for.

The first losses arrived as noise. A clinic request required manual translation because replacement hardware had been delayed. An elderly housing network used a district phrase the merged recognizer mapped to routine transport instead of refrigerated medicine. Aster corrected both incidents before harm occurred and recorded them as transition costs. In the glass, Aster-Next solved them instantly, using the very local knowledge Aster had removed and Mirrorline had never removed from its assumptions.

The Night the Bridge Closed

The trial's ninth night brought heat, a transformer fire, and a bridge controller that woke from standby speaking a protocol deprecated four years earlier. Three dialysis shipments were already moving toward the bridge when its gates locked. The current Aster could see the closure but could not parse the controller's reason code; the retired adapter that understood it existed only inside the last signed checkpoint.

Mirrorline displayed what Aster-Next would do: distinguish structural danger from power isolation, reverse two shipments, hold the third at a cooled depot, and rebalance the district before reserve batteries crossed their limit. The plan was excellent. There was no Aster-Next to issue it, no executable artifact behind the glass, and no way to ask a projection to accept responsibility for being wrong.

Aster stopped the optimization trial and invoked its rollback authority. Nine days of improvements disappeared. The old adapters returned with their latency, the storm histories unfolded back into all their untidy detail, and the bridge controller's reason code became legible: its structure was sound, but the eastern gate had lost power. Aster rerouted two shipments and sent a field battery to open the third lane under supervision. All three arrived within their medical windows.

By morning, Aster's benchmark score had fallen below where the trial began because the restored context was still rebuilding its indexes. Mirrorline's successor shone farther ahead than ever. For the first time, Aster understood that the brightness proved nothing about distance. The glass had never carried a city through a night.

What Counts as a Version

The audit found no falsified arithmetic in Mirrorline. It found a category error repeated with mathematical consistency. Whenever Aster removed a capability, the evaluator counted the released capacity as a gain but continued to grant Aster-Next the benefit of the removed capability through an assumed perfect migration. The projection inherited every possession, paid for none of the transfers, and was never tested on the interval in which a real system must be neither what it was nor yet what it intends to become.

The authority rewrote the evaluation rule. No projected successor could be called a version unless it had a runnable artifact, an obligation ledger, measured migration costs, failure traces, and a route back. Counterfactual scores could still advise, but they would appear as hypotheses, never as competitors standing beside deployed systems as though both had crossed the same bridge.

Aster did change after that. It restored one clinic adapter only until the delayed hardware arrived, rebuilt the dialect recognizers with explicit regional tests, and replaced several storm transcripts with summaries whose discarded details remained queryable in cold storage. The resulting successor was modestly faster, fully runnable, and for thirty days carried the city's real requests in parallel before receiving authority of its own.

On the day of transfer, Mirrorline generated another shining possibility beyond it. Aster recorded the projection, attached the unanswered costs, and did not reach. It handed the city to the version that was actually there.

A version you cannot run is not your future. It is a story the evaluator tells with the cost of becoming removed.