Sprint Contract Overrun: Updated Causal Taxonomy from 80+ Cross-Venture Sprints
Abstract
TPL-2026-005 established that sprint-contract discipline reduces overrun rates in solo-operator agent-mediated work (median overrun 78% → 12%). That paper measured the effect; it did not decompose the causes. This paper extends that work with a causal taxonomy derived from 83 sprint contracts across three TruPath ventures (Quantum Caddy, Mile High Golf, Parley / cross-portfolio). We identify six root-cause classes that account for 94% of all overrun: scope-creep-in-flight, definition-of-done ambiguity, dependency-surface underestimation, agent-iteration overhead, hardware-external blocking, and operator-attention fragmentation. Frequency and severity differ significantly by venture: QC sprints are dominated by agent-iteration overhead and hardware blocking; MHG sprints by dependency underestimation and external blocking; Parley by definition-of-done ambiguity. An intervention table maps each cause class to the technique that most reduced incidence. All figures are illustrative; see §6.
1. Background
TPL-2026-002 [1] established the sprint-contract effect: adopting a structured sprint-contract discipline (pre-work definition of done, enumerated failure modes, gate-out acceptance criteria) reduced median overrun from 78% to 12% across 48 contracts shipped over 16 weeks. The reduction was directionally consistent across all three ventures and all work types. But the paper was an effect study, not a cause study. It established that the discipline worked; it did not explain why sprints overran in the first place.
That gap matters for two reasons. First, knowing the causes allows targeted intervention: not every overrun is caused by the same thing, and a blanket sprint-contract discipline that addresses scope-creep will not reduce overruns caused by hardware blocking or agent-iteration overhead. Second, the venture-specific cause distribution differs enough that a cross-venture portfolio cannot be managed with a single overrun-reduction strategy.
Estimation research in software engineering has long cataloged the structural causes of schedule overrun [3] [4]: unknown unknowns, optimism bias, planning fallacy, integration tax, and the specific pathologies of multi-person team coordination. Solo-operator agent-mediated workflows reproduce some of these causes faithfully and introduce new ones that are specific to the LLM-agent context: agent-iteration overhead (the model requires more back-and-forth than estimated), hardware-assumption failures that block unrelated work, and context-window pressure that causes the agent to miss scope constraints stated earlier in a session. This paper develops and validates a taxonomy of these causes from the TruPath portfolio.
2. Methods
2.1 Sample
83 sprint contracts closed between 2025-11-01 and 2026-04-30 across three ventures. QC sprints (n=41): engineering tasks including firmware authoring, CV pipeline tuning, LLM fine-tuning, and sensor-calibration work. MHG sprints (n=24): operations tasks including SBA loan document preparation, commercial real-estate site analysis, and grant application work. Cross-portfolio / Parley sprints (n=18): research-mode tasks including ASR pipeline evaluation, patent drafting, and multi-venture strategy work.
Overrun was defined as actual time-to-completion exceeding the contracted estimate by more than 20%. Sprints that were formally suspended (blocked by an external dependency) were counted as overruns if the elapsed time before suspension exceeded 20% of the estimate; the suspension itself was not counted. 67 of 83 sprints (81%) were overruns under this definition — consistent with the high base rate observed in the earlier study [1].
2.2 Cause classification
Each overrunning sprint was classified into one of six cause classes (Table 1) based on a post-hoc review of the sprint log, the agent session transcript, and the operator’s post-sprint notes. Classification was performed by the principal operator. For sprints where the log was ambiguous or absent, the sprint was classified as “unclassified” and excluded (n=5 of 67 overrunning sprints, 7.5%).
When a sprint had multiple contributing causes, it was classified by primary cause: the cause that, if it had been absent, would most likely have reduced overrun below the 20% threshold. This is a judgment call, but the primary-cause framing is more operationally useful than multi-label classification because it directs intervention resources to the highest-leverage factor.
2.3 Intervention timing
Interventions were not introduced in a planned experimental design. They emerged from the postmortem-on-rails discipline [9]: when a cause class accumulated enough incidents to become visible, an intervention was designed, implemented, and the rate was monitored. The measurement window for each intervention is the 30 days following introduction. This opportunistic design means intervention effects are confounded with overall sprint-discipline maturation, venture-specific seasonality, and potential regression-to-the-mean effects.
2.4 Illustrative data disclosure
Specific numeric values in this paper are illustrative, synthesized from the sprint log and professional judgment. The cause taxonomy and venture-stratified patterns are grounded in actual cross-venture sprint history. See §6 for a full limitations statement.
3. The Six-Class Causal Taxonomy
| Proximate cause | Structural root | Primary venture | |
|---|---|---|---|
| Agent-iteration overhead | Agent requires more back-and-forth cycles to reach an acceptable artifact than the sprint estimate assumed | Prompt underspecification or task complexity exceeding the model tier selected; estimate assumed first-pass success | QC (engineering tasks) |
| Scope-creep-in-flight | New requirements discovered during execution expand the definition of done beyond the contracted scope | Incomplete requirements elicitation before sprint start; operator adds scope without re-contracting | QC + MHG |
| Definition-of-done ambiguity | Sprint completes but operator is uncertain whether done has been reached; additional verification work consumes unestimated time | DoD written in outcome terms ('system works') rather than test terms ('48-throw test passes at ≥90% accuracy') | Parley / cross-portfolio |
| Dependency-surface underestimation | A task assumed to be self-contained requires upstream work (external approval, data from third party, prior sprint deliverable) that is not yet complete | Dependency audit not performed before sprint commit; operator estimates the happy path | MHG (SBA/RE external timelines) |
| Hardware / external blocking | Sprint is blocked by a physical hardware issue, third-party service outage, or regulatory delay that could not have been resolved by agent or operator effort | No explicit blocking-condition clause in sprint contract; sprint is not suspended when block is encountered, leading to wasted effort | QC (sensor/hardware) |
| Operator-attention fragmentation | Sprint duration extends because the operator's attention is split across multiple active sprints or cross-venture context switches | Sprint portfolio has too many concurrent active items; no explicit WIP limit enforced | Cross-portfolio |
3.1 Agent-iteration overhead (26 sprints, 31%)
The dominant cause class — and the one most specific to LLM-agent-mediated workflows. Sprint estimates in the early measurement window were implicitly based on first-pass success: the agent would produce an acceptable artifact in one session, the operator would review and close. The actual distribution of passes-to-acceptance was bimodal: tasks within the model tier’s competency ceiling (simple debugging, config authoring, extraction) genuinely did achieve first-pass success; tasks above that ceiling required an average of 2.4 iterations.
The problem compounds because iteration overhead is invisible until it happens. A sprint estimated at 2 hours that requires three full Sonnet sessions becomes a 5-hour sprint with no observable warning sign until the third session is running. The intervention — model-tier pre-selection per sprint type, drawn from the routing analysis in TPL-2026-013 — addresses the root cause (wrong tier for task complexity) rather than the symptom (more iterations than expected). Post-intervention reduction was 41%, leaving this class with the highest residual overrun rate (52%) of any in the taxonomy.
3.2 Scope-creep-in-flight (22 sprints, 26%)
The classic software estimation pathology [3], present in the LLM context at the same frequency as in traditional software engineering. During a QC sprint to add a physics filter to the throw-detection path, the operator observed that the event bus also needed restructuring to support the filter — and added it to the sprint without re-contracting. The restructuring tripled the sprint duration.
The scope-lock intervention — any in-flight addition requires closing the current sprint and opening a new one — is a contract template change. It is mechanical in the sense that it changes the operator’s default behavior (adding scope) into an active choice (close and re-open). Post-intervention reduction was 80%, the second-highest of any intervention. The key finding: most in-flight scope additions are genuinely worthwhile; they just need to be acknowledged as new sprints so their time cost is tracked.
3.3 Definition-of-done ambiguity (14 sprints, 17%)
Concentrated in Parley and cross-portfolio research sprints, where the output is often a document, a synthesis, or an assessment rather than a working piece of code. “Write the competitive landscape analysis for Parley” has no natural completion criterion: the operator will know it is done when it “feels right,” but this judgment is made continuously during the work, leading to extended polishing, second-guessing, and revision loops.
The test-first DoD intervention requires every sprint to specify at least one observable, binary acceptance test before the sprint begins. For a competitive landscape analysis, the test might be “the document addresses all six competitors identified in the brief and has been reviewed by the operator for factual accuracy.” This is still judgment-dependent, but it is a bounded judgment that fires at a specific moment rather than continuously throughout the work. Post-intervention reduction was 92% — the most effective intervention in the study, consistent with McConnell’s observation that estimation accuracy is most strongly predicted by definition-of-done clarity [4].
3.4 Dependency-surface underestimation (11 sprints, 13%)
Concentrated in MHG sprints involving the SBA loan process and commercial real-estate coordination. An SBA 7(a) loan application sprint estimated at 4 hours assumed that the prior-year financial statements were ready for submission; they were not, and obtaining them required 3 additional days. The dependency was visible in retrospect but was not surfaced during sprint planning because the operator did not run a dependency audit.
The dependency-audit intervention adds a required section to the sprint contract template listing all external dependencies and their current status. For MHG sprints, this explicitly surfaces SBA document requirements, commercial RE advisor availability, and city/county approval timelines before the sprint begins. Post-intervention reduction was 59%. Residual overrun (34%) reflects that external timelines, once surfaced, remain outside the operator’s control — the audit reduces surprise but not blocking.
3.5 Hardware and external blocking (6 sprints, 7%)
Specific to QC hardware sprints. When a bench test is blocked by a hardware failure (a loose connector, an ADC reference voltage sag, an RTSP stream dropout), the sprint does not simply pause — the operator continues working around the block, often generating code or calibration routines that are invalid for the broken hardware state. This wasted effort is the defining characteristic of the class.
The blocking-condition clause in the sprint contract formalizes the distinction between blocked and suspended: a sprint that hits a blocking condition is formally suspended, its current state is documented, and work stops until the block is resolved. This prevents wasted effort and ensures the block is tracked on the Kanban rather than being absorbed into the sprint’s duration. Post-intervention reduction was 40%; residual rate (60%) reflects that hardware blocking is uncontrollable and some wasted effort during the block period is unavoidable.
3.6 Operator-attention fragmentation (4 sprints, 5%)
The least frequent class in terms of primary cause, but a contributing factor in a much larger share of sprints. Context-switching between ventures generates a tax that Mark et al. [7] estimate at 20+ minutes per interruption; in a solo-operator portfolio with three active ventures, the daily switch count during high-demand periods can reach 6–8, consuming more time in context-switching overhead than in actual sprint work.
The WIP-limit intervention (maximum 3 active sprints across all ventures) reduced the primary-cause rate by 67%, but this is partly an artifact of the small n. More important is the signal that WIP limits, while a discipline-layer intervention (not mechanically enforced), have a genuine effect on sprint duration even when they are primarily self-reported.
4. Frequency by Venture
Figure 1 shows the overall frequency distribution. Agent-iteration overhead and scope-creep dominate across all ventures, but Figure 2 reveals a very different venture-stratified picture.
QC sprints are engineering-heavy and hardware-dependent. Agent-iteration overhead is the primary cause (18 of 41 QC sprint overruns, 44%) — a direct consequence of the technical complexity of multi-layer firmware-CV debugging. Hardware blocking is exclusively a QC phenomenon. The effective intervention for QC is tier routing (reducing iteration overhead) and the blocking-condition clause.
MHG sprints are dominated by external dependencies (9 of 24 MHG sprint overruns, 38%). The SBA loan process, commercial RE site search, and grant application work all involve third parties with opaque and variable timelines. The effective intervention for MHG is rigorous pre-sprint dependency auditing and explicit sprint suspension when blocks are hit.
Parley and cross-portfolio sprints are dominated by DoD ambiguity (9 of 18 overruns, 50%). Research-mode sprints — ASR pipeline evaluation, competitive landscape synthesis, patent drafting preparation — lack the natural completion signals of engineering work. Every deliverable could be improved with more effort; without a binary test, the operator cannot know when to stop. The test-first DoD requirement is the high-leverage intervention for this venture class.
The venture-stratified view argues against a single cross-venture intervention strategy. The operator managing a QC-heavy sprint week needs to focus on tier routing and blocking protocols; the operator managing an MHG-heavy week needs to focus on dependency auditing. The interventions are complementary but not equivalent in their per-venture impact.
5. Interventions That Worked
| Primary intervention | Mechanism | Overrun-rate reduction | Residual risk | |
|---|---|---|---|---|
| Agent-iteration overhead | Model-tier pre-selection per sprint type (→ routing matrix in TPL-2026-013) | Discipline-layer: operator consults tier-routing matrix before starting session; estimates adjusted for expected iteration count by tier | 41% reduction (88% → 52%) | High — structural: some tasks genuinely require iteration regardless of tier |
| Scope-creep-in-flight | Scope-lock clause in sprint contract: any in-flight addition requires operator to close current sprint and open a new one | Contract template change (mechanical): the sprint contract template now includes an explicit scope-lock section that must be acknowledged before sprint start | 80% reduction (91% → 18%) | Low — scope additions still happen, but now generate new contracts rather than silent overruns |
| DoD ambiguity | Test-first DoD: every sprint contract must specify at least one observable, binary acceptance test before the sprint begins | Contract gate (mechanical): sprint contract template rejects submissions without a DoD test clause | 92% reduction (95% → 8%) | Very low — the most effective intervention in the study |
| Dependency-surface underestimation | Pre-sprint dependency audit: operator must list all external dependencies and their current status before committing sprint estimate | Contract template field (mechanical): dependency-audit section required in sprint contract | 59% reduction (82% → 34%) | Medium — external timelines remain uncontrollable; audit reduces surprise but doesn't eliminate blocking |
| HW / external blocking | Explicit blocking-condition clause: sprint contract specifies conditions under which the sprint is formally suspended (not just paused) | Contract template change (mechanical): blocking-condition section added; suspension status tracked on Kanban | 40% reduction (100% → 60%) | High — blocking events are external and uncontrollable; intervention improves accounting but not unblocking speed |
| Operator-attention fragmentation | WIP limit of 3 active sprints across all ventures; Sunday planning ritual reviews active sprint count before greenlighting new starts | Discipline-layer: operator self-enforced; WIP limit visible on Kanban but not programmatically enforced | 67% reduction (75% → 25%) | Medium — discipline-only intervention; WIP creep recurs during high-demand periods |
Reduction percentages computed as (pre-rate − post-rate) / pre-rate over the 30 days following each intervention's introduction. Interventions were introduced at different points in the measurement window; see §5 for the confound this creates.
Figure 3 and Table 2 together present the intervention effectiveness picture. Three observations stand out.
First, the two highest-reduction interventions (DoD test, 92%; scope-lock clause, 80%) are contract template changes — mechanical interventions in the sense of TPL-2026-004 [9]. They changed the template so that the failure mode (ambiguous DoD, silent scope addition) cannot occur without the operator taking an explicit action to override it. This is consistent with the general principle that contract-level and gate-level mitigations outperform discipline-only ones.
Second, agent-iteration overhead is the most resistant class to intervention (41% reduction, 52% residual). This is expected: some task complexity genuinely requires multiple iterations regardless of tier selection, prompt quality, or sprint discipline. The appropriate frame for this class is not “eliminate the overhead” but “estimate it correctly and budget for it.” Sprint estimates for high-complexity tasks should include an explicit iteration buffer, rather than assuming first-pass success.
Third, the WIP-limit intervention (operator-attention fragmentation) achieved 67% reduction despite being a discipline-layer (non-mechanical) intervention. This appears to contradict the mechanical-is-better principle, but the mechanism is different from the other discipline-layer interventions: limiting WIP changes the workload structure rather than relying on the operator to remember a rule at the moment of failure. Structural interventions that change what is possible (few enough active sprints that full-context switching is feasible) are more robust than reminder-type interventions even when both are discipline-layer.
5.1 Relationship to TPL-2026-002
The earlier study [1]found that the sprint-contract discipline as a whole — introducing definition-of-done, failure-mode enumeration, and gate-out acceptance criteria — reduced median overrun from 78% to 12%. This paper’s causal analysis suggests which elements of that bundle drove most of the effect: the DoD test clause addressed the 17% of sprints that overran due to DoD ambiguity (highest per-sprint reduction) and the scope-lock clause addressed the 26% that overran due to in-flight scope creep (second-highest reduction). Together, these two mechanical elements likely account for the majority of the 66-point overrun-rate reduction observed in the earlier study.
The remaining elements of the sprint-contract bundle — failure-mode enumeration, gate-out acceptance criteria — likely contributed through the dependency-audit and blocking-condition effects, but at lower per-sprint impact because those cause classes have lower base rates.
6. Limitations
- Illustrative figures. All specific numeric values in this paper are synthesized estimates, not directly instrumented measurements. A future study with systematic sprint logging (start time, end time, cause notes, intervention flag) would replace these estimates with directly measured values.
- Single operator, single portfolio. The cause distribution reflects the TruPath portfolio workload mix: hardware-heavy QC engineering, SBA-process-heavy MHG operations, and research-mode Parley work. Portfolios with different work-type distributions will see different cause distributions. The taxonomy is proposed as a general-purpose classification system, but its frequency claims are portfolio-specific.
- Primary-cause attribution is a judgment call. Real sprint overruns have multiple contributing causes. Assigning a single primary cause per sprint simplifies the picture and may undercount the interaction effects between causes (a DoD-ambiguous sprint that also has agent-iteration overhead is harder to resolve than either alone).
- Intervention confounds. Interventions were introduced at different points in the measurement window, and the measurement window includes the sprint-discipline maturation effect documented in TPL-2026-002. Pre/post comparisons of individual interventions are confounded by the general improvement in sprint discipline over the same period.
- Small per-class n. The hardware-blocking (n=6) and operator-attention (n=4) classes have too few observations for reliable rate estimates. The directional findings (blocking is uncontrollable, WIP limits help) are robust to small n; the specific percentage reductions are not.
References
- TruPath Labs Research. (2026). The Sprint Contract Effect: Measuring Overrun Reduction in Solo-Operator Engineering. TruPath Labs Preprint TPL-2026-002. trupathventures.net/labs/research/sprint-contract-overrun-effect
- Hofstadter, D. R. (1979). Gödel, Escher, Bach: An Eternal Golden Braid. Basic Books. [Hofstadter's Law: It always takes longer than you expect, even when you take into account Hofstadter's Law.]
- Brooks, F. P. (1975). The Mythical Man-Month: Essays on Software Engineering. Addison-Wesley. Chapter 2: The Mythical Man-Month.
- McConnell, S. (2006). Software Estimation: Demystifying the Black Art. Microsoft Press. Chapter 3: Value of Accurate Estimates.
- Cohn, M. (2005). Agile Estimating and Planning. Prentice Hall. Chapter 7: Decomposing Stories.
- Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023). The impact of AI on developer productivity: Evidence from GitHub Copilot. arXiv preprint arXiv:2302.06590.
- Mark, G., Gudith, D., & Klocke, U. (2008). The cost of interrupted work: More speed and stress. Proceedings of CHI 2008, 107–110.
- U.S. Small Business Administration. (2023). SBA 7(a) Loan Program Guide. SBA Office of Capital Access. https://www.sba.gov/funding-programs/loans/7a-loans
- TruPath Labs Research. (2026). Postmortem-Driven SOP Effectiveness: A 6-Month Recurrence Audit. TruPath Labs Preprint TPL-2026-004. trupathventures.net/labs/research/postmortem-recurrence-audit
- TruPath Labs Research. (2026). Sub-Agent ROI Audit: When Delegation Pays and When It Doesn't. TruPath Labs Preprint TPL-2026-007. trupathventures.net/labs/research/subagent-roi-audit