58 min read

A Signature Is No Longer Evidence of Judgment

Why a bare weekly forecast approval no longer proves human judgment when agents draft the baseline and adjustments — and how a judgment record, split FP&A and data engineering accountability, and a versioned weekly freeze make sign-off testable.

A Signature Is No Longer Evidence of Judgment

Introduction

When an agent drafts the baseline, proposes the adjustments and reconciles last week's variances, the name at the bottom of the forecast pack no longer certifies that a person reasoned through the number. It certifies that a person was present. For most finance teams that signature is the only control standing between the agent's output and the number the business plans against, and it is failing quietly — not because anyone is careless, but because the work the signature used to attest to has moved.

This article argues for replacing the bare approval with something a reviewer can actually inspect: a judgment record with testable acceptance criteria, a clear split between the FP&A judgment layer and the data engineering evidence layer, and a weekly freeze that captures uncertainty and model version alongside the committed number. The premises come from three places — the forecasting literature on judgmental adjustment, assurance standards on what a signature attests to, and operational forecast verification practice. Each is applied here by reasoning about weekly FP&A cycles, not restated as a rule those sources set for finance teams.

Why a Signature Stopped Meaning What It Used To

The old signature worked because of what sat behind it. A planner pulled the history, chose a method, keyed the number and could explain, if asked, why the third quarter looked the way it did. The signature attested to a chain of reasoning the signer had personally performed. Its evidential weight came from that chain, not from the pen.

Agent-mediated forecasting breaks the link. The baseline arrives already computed. Adjustments are proposed, sometimes pre-populated with a rationale drawn from commentary the agent has read. Variance reconciliation is drafted before the planner opens the file. What remains for the human is acceptance, and acceptance is not judgment. A reviewer can accept a number without having formed an independent view of it, and nothing in the artifact distinguishes the two cases.

The forecasting literature explains why this matters more than it might appear. Forecasting: Principles and Practice observes that practitioners adjust statistical forecasts far more often than they should, and often for the wrong reasons: adjusting creates a feeling of ownership and credibility, and users who do not understand the mechanism behind a statistical forecast feel that a judgmental adjustment makes the forecast their own (FPP3 §6.7). An agent-mediated workflow industrializes exactly that feeling while removing the effort that once made it costly. When a nudge takes four seconds, the friction that used to filter casual adjustments is gone.

One boundary should be stated plainly. This is an argument about what a signature can attest to, not a claim that agent-assisted forecasts are less accurate. An agent that produces a well-calibrated baseline and drafts defensible adjustments may well outperform a planner working alone. The problem is evidential, not necessarily numerical: the artifact no longer shows what the human contributed.

What the Research Says About Judgment Buried in a Number

Four findings from the forecasting literature carry most of the weight in what follows. Each is a finding about judgmental forecasting processes generally; the application to a weekly FP&A cycle is inference, and should be read as such.

Adjustment helps only under specific conditions. Judgmental adjustments are most effective when there is significant additional information at hand or strong evidence that an adjustment is needed — information not already incorporated in the statistical model. Adjustments appear most accurate when they are large. Small adjustments, especially in the positive direction, have been found to hinder accuracy and should be avoided (FPP3 §6.7). The same source warns against using judgment to correct a pattern the forecaster believes the model missed: forecasters tend to read non-existent patterns in noisy series, and statistical models handle data patterns better.

Documentation reduces bias. Formalizing and documenting the decision rules and assumptions behind a forecast promotes consistency, because the same rules can be applied repeatedly. Asking a forecaster to document and justify a forecast creates accountability, which can reduce bias, and formal documentation supports the systematic evaluation that should follow (FPP3 §6.2).

Forecasters and users should be segregated. Forecast accuracy may suffer when the task is carried out by the people who will act on the forecast or be measured against it. The classic case is a new product launch, where the forecast should be a reasonable estimate of sales volume even when that differs from what management hopes will meet financial objectives. A forecaster in that position may be delivering a reality check to the user (FPP3 §6.2). Targets and forecasts are different objects and should not be confused.

A signature attests to work, not to correctness. Assurance practice draws a narrower line than finance teams often assume. An assurance report carries the practitioner's signature, the date and the location of practice, and the standard requires documentation of who reviewed the work performed and the date and extent of that review (ASAE 3150). What the signature certifies is that defined work was performed and that a conclusion rests on a stated basis. It does not certify that the conclusion is right. That is precisely why the work behind the signature has to be inspectable: the signature is only as meaningful as the record it points to.

The Three Failures a Bare Approval Hides

A signature-only control cannot see three distinct defects, and each maps onto one of the principles above.

The unnecessary adjustment. A planner nudges the agent's baseline upward by a small amount, because the forecast feels more credible once it carries a human mark. The sign-off records that the number was reviewed and approved. It does not record that the adjustment had no informational basis. The research is direct on this: small positive adjustments have been found to hinder accuracy, and the ownership motive is one reason practitioners adjust more often than they should. A bare approval treats the nudge as neutral, when the literature treats it as a likely source of error.

The undocumented override. A large adjustment made because a contract slipped, a competitor withdrew or a supply constraint lifted is the kind of adjustment the research associates with improved accuracy — it carries information the model did not have. On the page, it looks identical to the cosmetic nudge: a revised number and a signature. Review therefore gives the good adjustment and the bad one the same treatment, which means it cannot reinforce the first or deter the second.

The conflated roles. The person whose target depends on the committed number is often the person approving it. The segregation principle exists because self-serving bias enters when the person accountable for the outcome also produces the number. In an agent-mediated cycle the bias has an easy route: optimistic adjustments are cheap to propose and, absent a reason field, invisible at review. The effect tends to surface at month-end, when the variance is already real.

What the three share is that none of them is visible in the artifact the signature sits on. The signature is not wrong about any of them; it simply has nothing to say.

What a Weekly Sign-Off Should Capture Instead

The replacement is a judgment record: the artifact the signature points to, and the thing a reviewer can actually test. For a weekly forecast, it should carry six elements.

  1. The baseline and its model version. Which model produced the starting number, identified by an immutable version label, so the committed figure can later be reconstructed.

  2. Each adjustment, itemized. Direction, magnitude, the stated reason, and — the field that does the real work — the specific information the model did not incorporate. If a planner cannot name that information, the adjustment is a candidate for removal.

  3. The uncertainty attached to the committed number. Not a single point presented as the forecast, but a range or interval with a stated basis.

  4. Who proposed and who approved. Two fields, not one, even when the same person fills both.

  5. The disposition of prior-week variances. For each material miss, what was learned and what changed, or an explicit decision that nothing changes.

  6. The data-quality attestation. A statement from the engineering side that the inputs behind the baseline were fresh and complete.

The design follows the sourced principles rather than inventing a format. Documentation and justification are the bias-reduction mechanism the literature identifies; formalized decision rules can be applied repeatedly and evaluated later. The record is not paperwork for its own sake. It is the artifact that makes systematic evaluation possible at all, and it is what a signature can then honestly attest to: that this work was performed, on this basis, by these people.

Splitting Accountability Between FP&A and Data Engineering

The segregation principle is easier to honor when accountability is split along a line that already exists in the workflow. The recommendation below is a design proposal, not an observed practice at any named organization.

FP&A owns the judgment layer. That means the adjustment rationale, the business context the model cannot see, the uncertainty statement attached to the commitment, and the commitment itself. These are the things a finance professional is qualified to sign for, and the things the record should attribute to a named person.

Data engineering owns the evidence layer. That means attesting that the baseline was produced by the recorded model version, that inputs were fresh and complete at the snapshot time, that the pipeline generating the numbers ran as specified, and that the version is archived and reproducible. These are the things an engineering team can verify directly, and asking finance to sign for them is asking for an attestation nobody in the room can support.

The handoff artifact is the seam: a versioned baseline accompanied by a data-quality attestation. Each side signs for what it can actually check. This is also where the segregation principle earns its keep — the person whose target depends on the number is no longer the person certifying the inputs, and the engineering attestation is a separate, checkable claim rather than a shared assumption.

There is an honest tension. In small teams the same person may hold both roles, and no organizational chart will change that. The fallback is a capacity-based distinction: the record must still show the two attestations separately, dated and labeled, so a reviewer can see that the input check happened and is not being inferred from the approval. One person can sign twice; the record should not let one signature stand for both.

Recording Uncertainty and Model Version at Each Weekly Freeze

A freeze that records only a timestamp is not a control. Each weekly freeze should archive three things together: the model version, the input snapshot, and an explicit uncertainty statement. The point is that any committed number can be reconstructed later from the record, and that the uncertainty was stated before the outcome was known.

Operational forecast verification supplies vocabulary that makes uncertainty checkable. The World Meteorological Organization's 2025 guidelines on verification of hydrological forecasts define five properties for assessing forecast quality: accuracy (closeness between forecasts and observations), bias (systematic deviation), reliability (consistency of forecasts with observed frequencies), resolution (the ability to distinguish among different outcomes), and sharpness (the concentration of the forecast distribution, indicating confidence) (WMO). A weekly FP&A cycle does not need all five as formal metrics, but the vocabulary is useful for stating what an uncertainty range is claiming. A narrow range that turns out to be wrong most of the time is sharp but not reliable; a wide range that is never wrong is reliable but tells the business little.

The freeze record should also carry prior weeks' forecasts alongside realized outcomes. Bias is then measured rather than remembered — a planner who believes their adjustments are conservative can check. The IEA Wind Task 36 recommended practice makes the related point that short-period error metrics do not provide reliable measures of overall performance, but they can indicate problems, and large errors should be logged and investigated (IEA Wind RP 20.3). A weekly cycle is exactly a short-period regime: treat each week's comparison as a signal to investigate, not as a verdict on the forecaster.

Keep the recommendation proportionate. The goal is a reconstructable record with a measured bias history, not a research-grade verification program. Version identifiers should be immutable, because a version label that can be reassigned defeats the purpose of archiving it.

A Testable Acceptance Checklist for Weekly Sign-Off

The criteria become usable when a reviewer can mark each one pass or fail. The checklist below tests whether the record is complete and inspectable. It does not test whether the forecast is right, and it should not be read as doing so.

  • Baseline version identified and archived. The committed number traces to a named, immutable model version.

  • Every adjustment carries a reason and a magnitude. No adjustment appears in the record without both.

  • Adjustments above the team's size threshold carry supporting evidence. The threshold is set by the team; the requirement is that the evidence field is populated above it.

  • The uncertainty statement is present and consistent with the commitment. The committed number falls within the stated range, and the range's basis is described.

  • Proposer and approver are distinguishable. Two fields, populated, even when the same person fills both.

  • Prior-week variances are dispositioned. Each material miss has a recorded disposition: investigated and explained, or explicitly accepted.

  • The data-quality attestation is attached. The engineering-side statement is present and dated.

Thresholds and tolerances are for each team to set and calibrate against its own cycle. A threshold that is too low turns the evidence requirement into a rubber stamp; one that is too high lets material adjustments through unexamined. The right level will differ by business model and should be revisited after a few cycles of use.

One thing the checklist deliberately does not do is score the forecaster. Marilyn Strathern's much-quoted formulation — that when a measure becomes a target, it ceases to be a good measure — applies directly here (Strathern, European Review 5(3)). A bias score used to evaluate individuals will be gamed, and the record will degrade into the ritual it was meant to replace. Use the record to evaluate the process and the decision rules, not the person.

The Strongest Objection: Isn't This Just More Process?

The objection deserves a straight answer: documentation overhead can consume the working hours that adjustment was supposed to save. Case studies of forecasting processes suggest that the effects of adjustments on accuracy are highly variable, often producing only a small net improvement at the expense of valuable working hours (Fildes and Goodwin, cited in arXiv:2411.00699). If the record adds a second layer of effort on top of an already marginal adjustment practice, the objection stands.

Three responses, in order of strength.

First, the checklist replaces an existing ritual rather than adding to it. The weekly review meeting already happens; the sign-off already happens. The record changes what those artifacts contain, not how many of them there are. The reason field is the main new burden, and it is deliberately scoped: every adjustment carries a reason, but only adjustments above the team's threshold carry supporting evidence.

Second, much of the record can be generated by the agent workflow itself. The baseline version, the input snapshot, the adjustment magnitudes and the prior-week variance list are all artifacts the system already produces. The human contribution is the reason field, the uncertainty statement and the disposition — the parts that are actually judgment.

Third, the alternative is not a lighter control. A signature that attests to nothing is not a cheaper version of attestation; it is the absence of a control wearing the costume of one. The comparison is not between a burdensome record and a lean signature. It is between a record and no verifiable control at all.

There is a genuine limit, and it should be conceded. If the record is completed mechanically — reasons copied from last week, uncertainty ranges pasted from a template, dispositions marked "reviewed" without review — it becomes the same empty ritual in a longer form. No template fixes that. What fixes it, if anything does, is the evaluation loop: the freeze record carrying prior forecasts against outcomes makes a mechanical record visible over time, because a record full of plausible reasons and poor outcomes is itself a finding.

Conclusion: Sign for the Judgment, Not the Meeting

The signature should attest to a documented judgment. Where there is a judgment to record — an adjustment grounded in information the model did not have, an uncertainty stated before the outcome, a variance investigated and dispositioned — the record shows it, and the signature means something. Where there is no judgment to document, the honest artifact is the agent's baseline with its uncertainty attached, not a human name added for comfort.

That is the shift worth making across finance and data teams: from approval to attestation. Approval says a person was present. Attestation says what work was performed, on what basis, by whom, and with what stated uncertainty. Only the second survives an audit, a post-mortem, or a question from the board about why the number moved.

As agents take over more of the weekly cycle, the record of human contribution becomes the control. Teams that build it now — the judgment record, the accountability split, the freeze discipline — will be the ones whose forecasts can still be trusted and audited when the drafting, adjusting and reconciling are largely done for them. The teams that keep signing the pack will find, eventually, that the signature is the only thing left in it that a machine did not produce.

Frequently Asked Questions

But doesn't a human reviewer still catch errors the agent makes?

Review still matters, and nothing here argues otherwise. The distinction is between catching an error and attesting to a judgment. A reviewer can spot that a number looks wrong — a series that jumped implausibly, a variance that does not reconcile — without having formed an independent view of what the number should be. Both are useful; they are not the same. The record should show which happened: an error caught and corrected is a different entry from a judgment made and documented. Conflating them is how a team ends up believing it has judgment coverage when it has error-checking coverage.

What if my team is too small to separate FP&A from data engineering?

Use the capacity-based distinction. The record must still show the two attestations separately — the judgment layer (rationale, uncertainty, commitment) and the evidence layer (baseline version, input freshness, pipeline execution, archival) — each dated and labeled, even when one person signs both. What the record must not do is let a single signature stand for both capacities, because that is exactly the conflation the segregation principle warns against. One person, two explicit attestations, is a workable compromise. One signature, no distinction, is not.

How do you respond to the finding that expert judgment beats statistical forecasts for volatile series?

It supports structured judgment, not bare approval. The finding that experts outperform statistical methods on highly variable series is a reason to take domain judgment seriously — and a reason to make it evaluable. If a planner's judgment adds value precisely when the series is volatile, that value should be visible in the record: which adjustments were made, on what information, and how they performed against outcomes. The record is how the judgment that beats the model becomes something the organization can learn from, rather than something that lives in one person's head and disappears when they change roles.

Doesn't requiring a reason for every adjustment just slow the weekly cycle down?

The requirement is the mechanism. The research identifies documentation and justification as what makes overriding a statistical forecast more challenging and guards against unnecessary adjustments — the friction is the point, not a side effect. If speed is the binding constraint, scope the effort: every adjustment carries a short reason, and only adjustments above a defined size threshold carry supporting evidence. That preserves the deterrent for cosmetic nudges while keeping the evidence burden where the research says accuracy actually comes from — the large, information-bearing adjustments.