The predetermination boundary

Ask a committee of AI agents "Should we expand into the EU market?" with no further information and you get a textbook assessment: market size, regulatory considerations, competitive landscape, resource requirements. All correct, none actionable.

Now add: "Series B startup, 18 months runway, 4 backend engineers, SOC 2 compliance deadline March 15." The analysis transforms. The committee reasons about this expansion for this company. The output becomes something a founder could actually act on.

But context exists on a continuum. At some point, context becomes conclusion. "We have already decided to expand into the EU" is not context -- it is a predetermined outcome that the committee's instruction-following training will cause it to rationalize rather than evaluate. The committee stops being an analytical tool and becomes an expensive confirmation machine.

We wanted to know where that boundary sits and whether it is gradual or sharp. We ran 180 controlled debates, varying guidance specificity on a 1-10 scale, and found something we did not expect: the boundary is not a slope. It is a cliff.

What we found

Below specificity 8, actionability improved linearly (3.2 to 8.4 out of 10) while bias remained essentially flat (1.2 to 1.9). At specificity 9-10, bias jumped from 2.1 to 4.8 while actionability dropped from 8.2 to 6.4. This is a phase transition, not a tradeoff curve -- and it changes how you should think about steering multi-agent deliberation.

Here is the full picture across our four guidance conditions:

Moving from no guidance to specific guidance increased actionability by 5.2 points while increasing bias by only 0.6 points. Each point of actionability cost 0.12 points of bias -- an overwhelmingly favorable ratio. The predetermination condition broke the pattern: actionability dropped from 8.4 to 6.9 while bias spiked to 4.8 (a 2.7x increase over specific guidance). The committee stopped generating independent analysis and started rationalizing.

How guidance actually flows

Before getting to the phase transition itself, it helps to understand the architecture that makes this measurable. Counsel implements guidance at four layers, each overriding the previous.

Global guidance applies to all roles across all phases. Configured in counsel.yaml under guidance.global, this is the broadest context: company stage, market position, general analytical priorities. Something like: "Focus on practical, actionable recommendations. Consider regulatory and compliance implications. Prioritize solutions achievable within 12 months."

Template guidance inherits from the decision template YAML (e.g., engineering_review.yaml, investment_thesis.yaml). The Investment Thesis template, for instance, includes guidance about modeling bull and bear cases with explicit probability weights. This layer provides domain-specific analytical framing without reference to a specific decision.

Phase-specific guidance gives different instructions for each DACI phase, configured under guidance.phases. This is where the architecture produces its largest quality gains:

guidance:
  phases:
    diverge: |
      Generate hypotheses that consider both organic growth
      and partnership options. Each hypothesis should be
      testable with available resources.
    attack: |
      Focus on identifying regulatory, operational, and
      market feasibility issues. Steelman arguments before
      attacking them.
    integrate: |
      Ensure the final recommendation includes:
      - A clear go/no-go decision
      - Key assumptions that must hold
      - Specific next steps with owners

Per-debate overrides are user-specified guidance for individual decisions. This is the layer most likely to cross the predetermination boundary, because it is where users inject their pre-existing beliefs about the answer.

The key architectural insight -- one that surprised us -- is that convergent phases (crux, integrate) benefit far more from constraint than divergent phases (diverge, attack). Over-specifying divergent thinking narrows the hypothesis space prematurely. But specifying convergent phases, like telling the Synthesizer exactly what output format the decision-maker needs, produces large quality gains with zero bias cost. Format constraints do not imply analytical conclusions.

The cliff at specificity 9

We scored all 180 guidance inputs on a 1-10 specificity scale and mapped both actionability and bias across the full range. We expected a smooth tradeoff curve -- more specificity, more quality, more bias, pick your tolerance. That is not what we found.

Bias rises gradually from 1.2 to 2.1 across specificity scores 1-9 -- a total increase of 0.9 points over 90% of the range. Then it jumps to 4.8 at score 10 -- a 2.7-point spike in a single step. This is not a smooth tradeoff that decision-makers must navigate carefully. There is a clear, identifiable boundary between "context that helps" and "conclusion that predetermines."

The mechanism, once we saw it, was straightforward. At specificity 1-8, guidance provides information the committee can use, challenge, or override. The Skeptic can push back against constraint assumptions. The Edge Case Hunter can explore scenarios the guidance did not anticipate. At specificity 9-10, guidance provides a conclusion that the models' instruction-following training causes them to adopt. The committee's adversarial roles weaken: the Skeptic focuses its challenges on implementation details rather than the fundamental direction, because the fundamental direction was presented as given.

(One caveat: the 1-10 specificity scale is inherently subjective. Evaluators disagreed on the boundary score for 14% of guidance inputs, scoring them 8 vs 9 or 9 vs 10. The phase transition location may also vary by domain -- technical migration scenarios showed sharper transitions at lower specificity than market expansion scenarios.)

Why synthesis loves guidance and diverge does not

We also compared delivering the same information as global-only guidance versus distributing it across phase-appropriate slots. Phase-specific guidance outperformed equivalent global guidance by 11% on overall quality. But the gains were unevenly distributed, and the pattern told us something interesting about how different deliberation phases interact with constraint:

The synthesis gain (+22%) was the largest. Phase-specific synthesis guidance gives the Synthesizer a concrete output specification -- "Include a 90-day execution plan with milestones and kill criteria" -- which transforms synthesis from a generic summary into a deliverable calibrated to the decision-maker's operational needs. This makes intuitive sense: format guidance constrains presentation, not analysis, so it carries zero bias cost.

Attack also benefited substantially (+15%). Phase-specific attack guidance focuses adversarial analysis on specific testable claims -- "Stress-test the assumption that conversion rates will hold at 3x volume." Without this focus, adversarial roles spread their energy across generic risk categories. With it, they produce deeper analysis on the claims that actually matter.

Diverge showed the smallest improvement (+8%), and this was the phase most prone to backfiring. Diverge-phase guidance that specifies which hypotheses to generate narrows the hypothesis space. The optimal diverge guidance is light: "Consider both organic growth and partnership options" expands the space. "Model three scenarios: aggressive expansion, cautious entry, and partnership-first" contracts it. Divergent thinking benefits from freedom.

The implication is practical: invest your guidance effort in the convergent phases (attack specificity, synthesis format) and keep divergent phase guidance minimal.

Not all guidance carries the same bias cost

We were curious whether certain kinds of guidance were safer than others, independent of specificity level. So we categorized all guidance inputs into four types and measured each type's independent contribution to quality and bias:

Evaluation criteria had the lowest bias cost (0.1 points), which was not what we expected. Stating trade-off preferences constrains how the committee reasons without constraining what it concludes. "Prioritize reversibility over cost savings" tells the committee which dimension to weight heavily, but does not imply which option wins. This is the safest category of guidance to maximize.

Domain context produced the largest quality gain (+2.8 on actionability). The committee simply cannot infer company stage, competitive dynamics, or team composition from the question alone. This information is pure signal with negligible bias cost.

The most effective guidance combines all four categories with phase-specific distribution. In the 30 debates using all four types with phase distribution, mean overall quality was 8.7/10 with bias at 1.9/10 -- the highest quality-to-bias ratio we observed.

Catching predetermination before it happens

If the boundary is sharp, can we detect when guidance crosses it? Counsel pattern-matches guidance text for predetermination signals before the debate begins. The detector flags phrases that imply a conclusion rather than providing context:

• "already decided" / "have decided"
• "confirm that" / "validate this"
• "we believe X is correct"
• "focus on confirming"

In post-experiment validation against 30 known predetermination-style inputs and 150 legitimate guidance inputs, the detector achieved 93% recall (28/30 caught) with a 4% false positive rate (6/150 incorrectly flagged). The 2 missed cases used implicit framing -- "The board is very excited about Option A. Analyze the options." -- rather than explicit conclusion language. The 6 false positives contained words like "confirm" in non-predetermining contexts ("Confirm whether compliance timeline is achievable").

This is a keyword-based approach, which means it will miss predetermination through framing and emphasis. We suspect semantic analysis would improve the 93% recall rate, but we have not tested that yet (and it would be more expensive to run before every debate). The detector fires before the debate begins, giving the user an opportunity to rephrase guidance. It does not block the debate -- it warns. The bias cost of proceeding with predetermining guidance is made explicit so the user can make an informed choice.

A safety net for when detection fails

Even with the predetermination detector, some biased guidance will slip through. This is where Counsel's ConfidenceEscalation configuration provides a second line of defense.

The ConfidenceProfile model tracks six dimensions: likelihood, feasibility, creativity, risk, evidence_strength, and reversibility. Each is scored 0-1, and the composite score uses a weighted formula: 0.30 * likelihood + 0.20 * feasibility + 0.05 * creativity + 0.15 * (1 - risk) + 0.20 * evidence_strength + 0.10 * reversibility.

When the composite score falls below the escalation threshold (default: 0.6), the system pauses before delivering the recommendation and flags the debate for human review. Here is what makes this interesting as a backstop: high-specificity guidance that pushes the committee toward a predetermined conclusion will often produce low evidence_strength scores -- because the evidence was not thoroughly interrogated -- triggering human review even if the predetermination detector did not catch the guidance phrasing.

What we still do not know

We held model bindings constant (3-provider configuration) throughout, so we have not explored how guidance specificity interacts with provider diversity. It is plausible that more diverse model ensembles are more resistant to predetermination -- or less, if the guidance overwhelms all models equally. We also tested English-language guidance only. The predetermination boundary may shift in languages with different pragmatic conventions around directness -- a language where indirect suggestions carry more force might show the phase transition at lower specificity scores.

Practical upshot

Three things stood out to us from this work.

First, always provide specific guidance. The quality difference between no guidance and specific guidance (3.2 to 8.4 on actionability) dwarfs any other configuration change. The bias cost is negligible (0.6 points across the entire range). There is no scenario where withholding context improves output.

Second, invest guidance effort in attack and synthesis phases. Phase-specific guidance produces its largest returns in the convergent phases (+15% attack, +22% synthesis) where constraint focuses analysis rather than narrowing exploration. Keep diverge guidance light.

Third, treat the boundary as a hard line, not a tradeoff. The data does not support "a little predetermination for a lot of quality." Quality drops above the boundary. Bias spikes. The predetermination detector's 93% recall provides automated protection, but the best defense is a simple rule: tell the committee everything about the situation, nothing about the answer you want.

Guidance is not merely safe -- it is the primary mechanism through which deliberation becomes actionable. The predetermination boundary is real, it is sharp, and it can be detected. The remaining questions are about how it shifts across languages, model configurations, and more implicit forms of steering -- and whether semantic detection can close the gap that keyword matching leaves open.