CN-023 — Intentional Irrecoverability

Concept Note · When Reconstruction Is Not Safe · June 2026
Continues: CN-022 · Related: SM-007 · SP-007

Abstract CN-022 established that AI can reconstruct a legacy system's operational ontology from system artefacts. This memo addresses the inverse question: under what conditions is that reconstruction not safe — not because the AI performs poorly, but because the conditions for valid interpretation no longer exist? The argument is that reconstruction safety depends less on technical capability than on the availability of tacit knowledge to adjudicate the output. Where that knowledge is absent, accurate reconstruction produces a precisely wrong map. The instrument is a tool for participants, not a substitute for participation. Core claim: The risk of legacy reconstruction is not technical failure but interpretive failure — and interpretive failure is most likely precisely when reconstruction is most urgently needed.

§ 01 — The Inverse Problem

CN-022 argued that AI can reconstruct the implicit ontology of a legacy system — what the system knows — from code, schemas, logs, and documentation. The instrument is powerful. The limitation identified was the distinction between structure (recoverable) and intention (partly irrecoverable).

This memo addresses a more specific question: when is the reconstruction not merely incomplete but actively unsafe?

The question is not rhetorical. Reconstruction is most urgently needed in the situations where it is also most dangerous: organisations that have lost institutional memory, that face large-scale modernisation under time pressure, or that have inherited systems whose design rationale is genuinely unknown. These are the cases where reconstruction is hardest to adjudicate and most likely to produce drift.

The instrument is most needed where it is least safe to use without supervision. This is the inverse problem of legacy recovery.

§ 02 — The Temporal Gap

A reconstructed ontology is static. It shows what concepts exist in the system and how they relate. It does not show which structures are ontological residue — historically contingent, no longer governing real decisions — and which are operational core concepts still actively controlling outcomes.

This distinction requires temporal knowledge: why a structure appeared, under what pressure, to solve what problem. A field that was added as a workaround in 1987 and never removed looks identical to a field that encodes a fundamental business rule. Both are present in the code. The reconstruction instrument treats them equally.

Without temporal context, the reconstructed map has no elevation. It renders terrain accurately but cannot show which features are load-bearing and which are geological remnants from a prior landscape.

§ 03 — The Dialogic Condition

The instrument is safe when used dialogically — when the reconstruction output is returned to a trained interlocutor who can adjudicate it against tacit knowledge.

The trained interlocutor performs a function the instrument cannot: they recognise the difference between ontological residue and operational core. They know which exceptions became permanent because no one removed them, and which exceptions are permanent because the domain actually contains that edge case. They hold the temporal dimension in memory that the system cannot encode.

In this mode, the instrument externalises the system's implicit model and makes it available for structured critique. The experienced practitioner does not need the instrument to tell them what the system does — they know that. They need the instrument to surface what they have forgotten or never articulated. The dialogue between the reconstruction and the expert is where the value is generated.

The instrument externalises institutional memory. It does not create it. An organisation that has lost its tacit knowledge does not recover it through reconstruction — it receives an accurately detailed rendering of what it no longer understands.

§ 04 — Conditions for Unsafe Reconstruction

Three conditions, individually or in combination, make reconstruction unsafe regardless of the instrument's technical quality:

Condition 1 — Tacit knowledge severance The organisation has lost the practitioners who understood the system's design rationale. Key staff have retired or departed without knowledge transfer. The reconstruction output has no qualified interlocutor to adjudicate it. Accuracy of reconstruction increases the danger: a precisely wrong interpretation is more likely to be acted on than an obviously incomplete one.
Condition 2 — Time pressure without validation cycles The modernisation timeline does not allow for iterative validation of the reconstructed ontology against operational behaviour. The reconstruction is treated as a deliverable rather than a starting hypothesis. Wrong reconstructions propagate into the new system before they can be corrected.
Condition 3 — Domain opacity The system's domain is not legible to the modernisation team — legal, actuarial, or regulatory structures that the system encodes are not understood by those performing the reconstruction. The instrument cannot supply domain knowledge it was never given. It will reconstruct a coherent model of a domain it does not understand, and that model will be internally consistent and externally wrong.

These conditions are not hypothetical. They describe the modal modernisation project: inherited systems, departure of original architects, time-constrained delivery, and technical teams without deep domain knowledge. The instrument is being applied most widely in precisely the conditions that make it most dangerous.

§ 05 — What Safe Reconstruction Requires

Safe reconstruction is not a property of the instrument. It is a property of the organisational conditions surrounding the instrument's use.

Minimally, safe reconstruction requires:

The third requirement is the most commonly omitted. Reconstruction instruments produce outputs that appear equally authoritative across the full confidence spectrum. Confidence stratification must be imposed externally.

A reconstruction without confidence stratification is a map without a legend. It shows everything at the same resolution, which means the critical distinctions are invisible.

Conclusion

The question is not whether AI can reconstruct legacy ontology. CN-022 established that it can. The question is whether the conditions exist for that reconstruction to be interpreted correctly.

Those conditions are organisational, not technical. They depend on tacit knowledge, time, and domain legibility. Where those conditions are absent, the instrument's accuracy becomes a liability: it produces a confidently rendered map of terrain the organisation can no longer read.

The appropriate response is not to avoid the instrument but to assess the conditions before deploying it. An organisation facing Conditions 1–3 does not need better reconstruction. It needs to reconstruct the interpretive capacity first — and that is a human problem that no instrument resolves.

The instrument is a tool for participants. It is not a substitute for participation.

Related documents
CN-022 — Legacy Ontology Recovery (the prior memo this continues)
SM-007 — The Convergence Finding (local correctness / global drift)
SP-007 — Biological Constraints on Institutional Continuity (narrative as default assumption)
CN-002 — Human–LLM Cognitive Space (dialogic instrument model)