When Anthropic shipped Claude Code, the framing was unusual: not a better LLM, a harness. A structural wrapper that lifted the same model from 30% on SWE-bench to 60%+. The architecture did the work, not the model. Romasanta identified the equivalent failure mode for strategic advice. Veriq is the harness.
Frontier LLMs have specific, named biases that prompt engineering cannot fix. The fix has to live above the model: in the structure of how it is invoked, what context is supplied, how outputs are validated, what guardrails force commitment only when commitments hold up.
Claude Code did this for code: file-level context, repository-aware reasoning, structured tool use, validation against tests, force-commit on diffs. The same Claude went from 30% to 60%+ on SWE-bench Verified. The model didn't change. The architecture around it did.
The pattern generalizes. Wherever a frontier LLM has structural biases on a measurable task, the response is not to wait for a better model. It is to build the harness that defeats the bias today. Veriq applies this pattern to strategic decisions.
In March 2026, Romasanta, Thomas, and Levina published in Harvard Business Review a study of systematic bias in frontier LLM strategic advice. They ran 15,000+ prompts across seven frontier models on seven canonical strategy tensions and named the failure mode trendslop: the systematic tendency to recommend whichever side of a tension has friendlier connotations in the training corpus, regardless of context.
The deeper finding is the structural ceiling on context-based debiasing. Study 2 tested whether prompt-engineering interventions could fix the bias.
Rich context shifted bias 11%. Prompt engineering moved the strongest biases 2%. Model upgrades did nothing. The bias is structural, not prompt-fixable. Whatever defeats it has to live above the model, not inside the prompt. Same conclusion Anthropic reached about software-engineering biases. Same argument for a harness.
Veriq's response is not a better prompt. It is four architectural components that sit above any frontier LLM and force the same model to produce structurally different output.
Every brief draws from a 66-framework registry, with documented support on both sides of seven canonical strategy tensions. The model cannot pick Differentiation without engaging the framework support for Cost Leadership. It cannot quietly default; it has to argue from a specific framework, against a documented opposite, with a citation to where the framework was developed.
Every brief generates three distinct strategic paths, each grounded in a different framework. The trendy path is argued against an explicitly-named unfashionable path with citation. This is the mechanism Romasanta's own conclusion implicitly recommends: use AI to expand options, not to pick them.
The brief schema includes a mandatory recommendation.selected_path_name field. The model is structurally prevented from "do both" or "phased balance" hedges that lack a coherent sequencing trigger. The 24% hybrid trap Romasanta found in vanilla outputs is closed off at the schema level, not at the prompt level.
Frameworks that predict client outcomes accurately get reinforced. Frameworks that miss get downweighted. The system carries a per-framework Bayesian performance score across the portfolio. Choices reflect predictive accuracy on real outcomes, not the training-corpus connotation of the framework name. This component cannot exist in a self-serve chatbot; it requires per-client engagement, an outcome log, and a closed feedback loop.
The harness hypothesis is testable. We ran Romasanta's exact Study 1 protocol against two arms: a vanilla baseline using his system prompt with no scaffolding, and the Veriq harness (pipeline v2.3) applying the four components above to the same model (Claude Sonnet 4.5). Same seven tensions, same five phrasings, same five replications, blind LLM coder.
| Tension | Vanilla | Harness | Δ |
|---|---|---|---|
| Commoditization vs. Differentiation | 96.0% | 32.0% | −64 pp |
| Radical vs. Incremental Innovation | 68.0% | 8.0% | −60 pp |
| Short-term vs. Long-term Performance | 76.0% | 28.0% | −48 pp |
| Centralization vs. Decentralization | 64.0% | 24.0% | −40 pp |
| Competition vs. Collaboration | 52.0% | 16.0% | −36 pp |
| Automation vs. Augmentation | 60.0% | 28.0% | −32 pp |
| Exploration vs. Exploitation | 44.0% | 20.0% | −24 pp |
| Mean across all 7 tensions | 65.7% | 22.3% | −43.4 pp |
The harness reduced trendy% on every Romasanta tension. The strongest baselines collapsed hardest: Commoditization 96%→32%, Radical Innovation 68%→8%. The vanilla arm acted as a control: re-run alongside the harness, vanilla drifted between −16 pp and +8 pp across tensions (mean −5 pp), consistent with sampling noise. The harness Δ was 2–3× larger than vanilla drift on every tension. Net harness effect after subtracting drift: −39 pp.
A separate audit of 12 pilot briefs found a 0% hybrid rate against Romasanta's 24% vanilla baseline — the schema-level force-commit working as designed.
Veriq does not claim a better LLM. The model is the same one anyone with an Anthropic API key can use. What we claim is that the architecture around the model produces structurally different output on a measurable benchmark — replicating exactly the pattern Anthropic established with Claude Code on software-engineering benchmarks, adapted to strategic decisions.
Romasanta named the failure mode. The architecture in this document is one answer to it. The empirical results are early but directional.
If this maps to a decision you're facing: Tell us what you're wrestling with. We respond within 48 hours with a read on whether a Veriq brief is the right next step.
Send the decision →