What we stopped counting in the replacement estimate

One of the first corrections to HALO's internal Homebridge replacement-cost model was an addition error. The component rows did not sum to the subtotal we had given them.

Correcting the arithmetic was necessary. It was also the easier problem. A table can add up perfectly while counting work it has not justified.

We were trying to estimate the professional effort required to reproduce defined Homebridge capabilities. The work crossed software, integrations, data, cloud operations, testing and documentation. Breaking it into components made the question more manageable: what existed, what would have to be recreated, which skills would be involved and how much effort might that require?

The resulting detail looked substantial because the underlying work was varied. But the hour estimates were generated assumptions, not measurements of another team's labor. A detailed description of a component did not independently establish how long a competent person would take to reproduce it.

That left a circular argument available to us. The system appeared sophisticated. We assigned substantial effort to its components. The total then appeared to confirm that the system was sophisticated. The calculation could return our judgment to us with more authority than it had acquired.

The internal challenge therefore had to reach beyond arithmetic. What had we allowed into the scope? Were we counting a working capability, a design for one or a future possibility? Did the components require distinct effort, or had our descriptions divided the same work several times?

The clearest change concerned Homebridge's Knowledge Observatory.

The Observatory contained useful work around entities, claims, sources, relationships and structured representation. That did not automatically make the whole envisioned system a completed capability within the replacement scope. The existence of an interesting demonstration could not settle the maturity of everything it suggested.

We removed the Observatory from this analysis.

That decision did not say the project had no intellectual use or that its implemented artifacts ceased to exist. It said this particular replacement model would no longer count it. Exclusion from a defined estimate is a narrower judgment than declaring a project valuable or valueless.

Other changes occurred inside what remained. The AI discoverability scope was brought closer to concrete catalog, feed, structured-information and observed surfacing work. Future possibilities were not to receive credit as completed implementation. Within Growth Engine, the model distinguished stronger operational evidence from partial Google proof and mixed-maturity persistence or reconciliation work.

Those were modeling decisions about what could be counted at which maturity. They did not independently certify every retained capability or supply the exact effort required to reproduce it. They made the assumptions more exposed to challenge.

Overlap required a different kind of scrutiny. Architecture, integration, deployment, testing and documentation can describe distinct responsibilities. They do not necessarily imply separate people performing wholly separate blocks of work. A capable engineer may cover several of them in the same engagement.

Some activities also belonged to more than one system description. Understanding a product source did not become two independent tasks because two outputs used it. Attribution work could not be counted in full again merely because another channel needed it. Naming the work twice did not make the effort occur twice.

We challenged the component model with a broader question about a plausible replacement team. That was a useful check on whether the allocation made sense. It remained an internal reasonableness exercise, not an independent estimate or a market quotation. Agreement between two calculations can reflect shared assumptions as easily as external confirmation.

The revised model became smaller. That did not make what remained independently validated.

The final report still described estimated professional replacement effort under a chosen scope and set of assumptions. Its evidence appendix recommended a component-to-artifact schedule. In preparing this account, we did not recover the complete final trail connecting every adjusted effort assumption and deduction to its supporting artifact. We cannot turn a recommended audit trail into a completed one by writing confidently about the result.

Observed work does not measure another team's replacement effort. Calendar time is not labor time, and historical spending is not replacement cost.

Replacement cost establishes neither market value nor usefulness to a particular buyer. Hypothetical expenditure cannot become realized savings or productivity merely by subtraction.

The arithmetic correction fixed an earlier subtotal. Scope, maturity and overlap changes narrowed what we counted. None validated the surviving assumptions.

What changed was more than the total. The model became a narrower account of work we believed would need to be reproduced, with a clearer explanation of what had been left out. The next serious challenge could still change it.