Goal fragmentation: signals distinguishing salami-slicing from legitimate…

Research Question

When a Goal is a fragment of a larger intent (salami-sliced deliberately or accidentally), what signals distinguish it from a legitimately scoped sub-goal?

Findings

(Populated from §6 Synthesis above.)

Executive Summary

Salami-slicing (artificial goal fragmentation to evade approval thresholds) and legitimate sub-goal scoping produce identical-looking specifications in isolation; the distinction is detectable through a three-tier signal battery: structural, semantic, and contextual. The strongest single signal is semantic: a goal that cannot deliver a customer-observable outcome in isolation fails the INVEST "Valuable" criterion and the EIA standalone justifiability test, regardless of whether its specification text is structurally complete. Technical prerequisites and infrastructure enablers are the main false positive class: they fail the "Valuable" criterion yet are legitimate sub-goals when the dependency on a declared parent is explicitly stated. A reliable minimum check requires at least three conditions to be satisfied simultaneously: (a) the goal carries an explicit parent reference, (b) its success criterion is verifiable in isolation, and (c) it delivers at least one customer-observable outcome without requiring other goals to be completed first.

Key Findings

1. Salami-slicing is the splitting of a single coherent goal into multiple fragments, each below a governance or approval threshold, such that the aggregate intent is never visible in any single specification artefact, as established in Environmental Impact Assessment (EIA) case law and programme governance practice.

2. KAOS AND-refinement completeness provides the most formally precise structural signal for legitimate sub-goal scoping: the conjunction of sub-goals must bidirectionally imply the parent goal, and a goal with no declared parent cannot pass this test, making absent parent reference a necessary precondition for applying any fragmentation check.

3. INVEST criteria (Wake, 2003) encode the two strongest semantic fragmentation signals as questions usable without formal logic competence: "Valuable" asks whether the goal delivers customer value in isolation, and "Testable" asks whether acceptance can be determined without completing another goal first.

4. Ashchurch three-part test: the UK Court of Appeal (2023) establishes a functional interdependence test for detecting artificial project splitting: a goal is a fragment if it cannot be justified on its own merit, if documentation reveals it forms part of a wider scheme, and if it would not be used until the wider scheme is complete.

5. Technical prerequisites and infrastructure enablers are the primary false positive class for fragmentation signals: they fail the INVEST "Valuable" criterion yet are legitimate sub-goals when the dependency on a declared parent goal is explicitly stated in the specification.

6. Conditional acceptance criterion: a goal whose acceptance depends on another goal's completion fails the independent testability test and is a semantic fragmentation signal; in Planning Domain Definition Language (PDDL), such a mutual precondition chain may render the entire goal set unsolvable at plan-search time, making the fragmentation computationally detectable.

7. Horizontal slicing (scoping a goal entirely within one technical layer without delivering an end-to-end customer-observable outcome) is a documented anti-pattern in agile product management that fails both the INVEST "Valuable" criterion and the EIA standalone justifiability test.

8. Contextual signals (approval threshold proximity, temporal clustering of related fragment submissions, absence of disclosure of the wider scheme) are the most specific indicators of deliberate salami-slicing but require governance metadata absent from the goal specification text and cannot be evaluated from specification content alone.

9. Benefit Dependency Network (BDN) orphaned-node concept in Managing Successful Programmes (MSP) encodes a structural fragmentation signal: any benefit or capability node with no traceability path to a strategic objective is structurally fragmented and must be re-linked or retired.

10. No single structural, semantic, or contextual signal is sufficient to determine fragmentation; the literature across requirements engineering, product management, programme governance, and EIA case law consistently supports a combined minimum check requiring an explicit parent reference, an independently verifiable success criterion, and at least one customer-observable outcome delivered without requiring other goals to be completed.

Assumptions

• Assumption 1: The EIA functional interdependence test is applicable to requirements governance beyond planning law. Justification: The logical structure of the test is domain-agnostic; its criteria (can this artefact be justified on its own merit? does documentation reveal it is part of a larger scheme?) apply wherever a governance threshold exists.

• Assumption 2: The INVEST criteria, designed for agile user stories, are applicable at the level of Goal specifications as used in this item's scope. Justification: The "Valuable" and "Testable" criteria address the semantic properties of any well-formed goal specification regardless of delivery methodology.

• Assumption 3: Reinertsen (2009) connects legitimate batch-size reduction to parent-goal traceability. Justification: This claim is drawn from secondary summaries; primary access was not obtained. It is consistent with the documented batch-size optimisation principle but is treated as an assumption pending direct source verification.

Analysis

The three signal tiers form a detection ladder in which each tier catches a different fragmentation mechanism. Structural signals cost the least to evaluate but carry the highest false-positive rate because legitimate prerequisites and infrastructure goals routinely lack standalone value. The semantic tier is more discriminating because the "Valuable" test is explicitly about customer-observable outcomes rather than implementation structure, which is harder to fake without revealing the wider scheme. Governance metadata unlocks the contextual tier, which is the most specific to deliberate intent but cannot be evaluated from specification content alone, making it complementary rather than primary.

A rival interpretation is that structural signals alone are sufficient: a goal without a declared parent cannot pass the KAOS AND-refinement entailment check, which is itself conclusive. The structural interpretation is weaker for two reasons: (a) a legitimate root goal also lacks a parent, so the signal is not specific to fragments; and (b) a declared but fictitious parent can be added to a fragmented goal, satisfying the structural check while preserving the fragmentation effect at the semantic level. Semantic signals are therefore the most robust primary layer; structural and contextual signals serve as supporting evidence.

The GORE completed item (2026-05-31-gore-strategic-intent-to-delivery-decomposition) establishes that the abstraction gap in GORE decomposition occurs in roughly 40-60% of industrial applications. This finding sharpens the current item's conclusion: at the abstraction boundary, GORE formal rules stop providing guidance, making semantic signals (INVEST, EIA) particularly valuable as the primary detection layer precisely where formal checks break down.

Risks, Gaps, and Uncertainties

• The Horkoff et al. (2017) systematic mapping was not directly accessed; its content is drawn from the completed prior item on GORE decomposition.
• Primary access to Reinertsen (2009) was not obtained; the batch-size/parent-goal traceability characterisation is from secondary summaries.
• The Firesmith (2004) article was retrieved as a binary PDF only; scope-coherence claims are characterised from abstract-level search results.
• The AXELOS MSP source was inaccessible in useful content; the BDN orphaned-node claim is based on secondary characterisation.
• No empirical study was found measuring false-positive and false-negative rates of the proposed signal taxonomy on a labelled dataset of goal specifications.

Open Questions

1. Can the combined three-check minimum be encoded as a formal schema validation rule that an automated planner can apply at goal submission time? This would be a natural extension of the five-field minimum schema from the completed item 2026-05-31-goal-specification-completeness-schema.
2. Is there empirical evidence on false-negative rates for semantic signals in industrial requirements databases?
3. Can the EIA functional interdependence test be operationalised as a machine-readable check without requiring natural language understanding of the "wider scheme"?

sources

• [x] Van Lamsweerde et al. (2001) Handling Obstacles in Goal-Oriented Requirements Engineering: Knowledge Acquisition in Automated Specification (KAOS) AND-refinement completeness and bidirectional entailment requirement
• [x] Wake (2003) INVEST in Good Stories, and SMART Tasks: Independent, Negotiable, Valuable, Estimable, Small, Testable (INVEST) criteria; horizontal slicing as anti-pattern; "Valuable" and "Testable" signals
• [x] Agile Alliance (2023) INVEST Glossary: canonical definition of INVEST criteria
• [x] Pinsent Masons (2023) Bridge ruling: warning for English planning applications: Environmental Impact Assessment (EIA) Ashchurch ruling; functional interdependence test; three-part salami-slicing detection framework
• [x] planning.wiki PDDL Problem Reference: Planning Domain Definition Language (PDDL) goal block; mutual precondition chains
• [x] Reinertsen (2009) The Principles of Product Development Flow: batch-size and scope fragmentation; secondary access only
• [x] AXELOS Managing Successful Programmes (MSP): Benefit Dependency Network (BDN) orphaned-node concept; product page access only
• [x] Firesmith (2004) Specifying Reusable Security Requirements: goal granularity and scope-coherence; PDF retrieved as binary only
• [x] Yu (1997) Towards Modelling and Reasoning Support for Early-Phase Requirements Engineering: i-star means-end links; structural orphan in strategic dependency models