Better Business Decisions Require Better Information — Part 4

Better Business Decisions Require Better Information

Part 4 — The Technical Review Standard

Roger E. Henley II · 2026

Purpose

The series exists because unstructured human decision-making is biased, and because enterprises need better infrastructure if they want to raise the odds of a good call. Better systems for analyzing data are not a technical preference. They are how an organization increases the likelihood of a positive outcome when the people making the decision cannot see their own error.

This paper specifies the technical review as a control. Procedure does not depend on the participant's ability to detect that error. It is a specification of the same kind the plant already funds for a pressure transmitter, a drawing, and a physical change, applied to the decision those instruments exist to inform.

What Parts 1–3 Established

Part 1. Incomplete information architecture hides the structure of the problem. Each function's map is internally coherent and therefore feels complete. Acting on partial data feels right, and often is not. The cost shows up late, and the wrong local fix adds a second bill.

Part 2. Human interpretation then distorts the evidence that does exist. Confirmation, framing, and unearned confidence produce a conclusion that feels settled. We become confidently wrong. The original frame absorbs contrary data instead of being tested by it.

Part 3. Even a correct analysis, sitting on better information, can still fail in the exchange that is supposed to turn it into a decision. The room scores fluency, substitutes an easier question, and treats adjournment as evidence that the search is over. Four practices help an individual tomorrow: make the condition inseparable from the answer; give dates the same treatment; rank the doubts; tag the provenance of every material figure. They are real. I use them. They are also collectively insufficient. An individual can bring a bounded answer into a room that was not built to receive one. Education confers no immunity. If the remedy cannot be educational, it has to be procedural.

This paper does not add a fourth diagnosis. It specifies the procedure.

A Problem Well Defined

A review cannot close a search that was never opened as a problem. It can only ratify a preference.

If the problem is not stated, and if what counts as solving it is not stated, the team will work in circles. The goalposts will move because there were never goalposts. Activity will look like progress. It will not be.

A problem well defined is a problem half solved. That is not a slogan. It is a control. The definition does the first half of the work: it names the gap between the current state and a state worth paying for, and it names how anyone would know the gap had closed.

Three artifacts, before analysis, before options, before a date:

1. The problem, in one sentence, including who absorbs the cost if it is left alone.
2. What solved looks like, in the world, including the measure and the threshold that would count as done. Not "improve yield." "Restore lot yield to the 18-month baseline and hold it there for three consecutive lots."
3. The value of solving it, stated as the cost of leaving it unsolved, in the units the business already uses to allocate capital. If that number cannot be produced, the problem is not yet defined. It is a complaint.

A well-defined problem that is not worth the work will still consume a team. Establishing true value is what keeps a real problem from becoming an endless one. Without these three, later controls have nothing to close against. If the solution cannot be recognized, the meeting will keep meeting.

The Decision Class

Not every decision deserves this treatment. Putting it on all of them would be as wasteful as a Management of Change package on a light-bulb replacement.

The new work is to draw the line in writing, before the meeting, and to suspend fluency-based judgment above it. The test is not "does this feel important." The test is structural:

• Does the decision commit capital, product, or safety in a way that is expensive to unwind?
• Will the evidence that it was wrong arrive after the people in the room have moved on?
• If it is wrong, can other functions see that from inside their own maps, or does the error stay hidden until the cost lands?

If yes, the expert is not ignored. The expert's confidence is not admissible as evidence, including when the expert is certain, including when the expert is you.

The plant already draws this kind of line for lockout and for MOC. The failure mode is irreversible enough to specify in advance. Decisions in this class have the same failure mode. They have not had the same line.

The class has to be narrow enough to fund. If everything is a Class A review, nothing is.

What the Meeting Is For

Above that line, the meeting is not convened to produce a yes. It is convened to receive a decision rule.

A decision rule has four parts, stated in the same breath: the recommendation; the governing variable and its threshold; where we are relative to that threshold today; what happens to the recommendation if the input moves.

"Above 15.5% moisture, option A. Below, option B. We are at 14.2% today." That is a complete answer. What remains open is a monitoring question: has the input crossed the trip point?

A date is the same artifact. A bare date is a promise the speaker cannot keep. Conditions without a date is a refusal to answer. The bounded form is: "August 14, contingent on the switchgear shipping by June 1. Every week it slips past that moves the date a week. I will confirm or revise on June 1 either way." When the date moves, the movement is self-explaining.

The derivation belongs in the file. Leading with it treated complexity as the message. The decision rule is the message. A specification sheet carries the tolerance. The calculation package sits in the file.

You cannot train a room out of scoring fluency. You can change what the room is scored for producing.

Entry Criteria

A permit-to-work system does not begin because the crew arrived. It begins when the required conditions are on the page. If the following are missing, the meeting has not yet become a review.

• The problem, what solved looks like, and the value of solving it.
• The alternatives considered and the one-line reason each was set aside. A review with only one option is a ratification.
• The decision rule in bounded form.
• Provenance on every material figure: measured, modeled, vendor-asserted, assumed, or inherited. Untagged figures are treated as assumed.
• The two doubts that could change the decision, ranked, with the observation that would promote a third.
• The residual-risk owner, named as a person, not a function.
• The next proof-test: what will be loaded, by whom, by when, that would show the rule is wrong.

The decision frame in Part 2 is the derivation behind these artifacts. It belongs in the file.

Admissible Uncertainty

Undifferentiated uncertainty is not honesty. It transfers the ranking problem to people who cannot perform it.

Uncertainty is admissible only when it is ranked, thresholded, tagged by provenance, and capable of changing the rule, the date, or the residual-risk owner. Everything else stays in the file.

A fluent analysis looks finished because it required nothing of the room. An unranked register is an alarm flood. The room will stop looking.

Independence That Is Real

A reviewer who shares the capital request, the reporting line, the milestone, or the original frame can still be competent. They cannot be the calibration standard. A gauge cannot confirm itself.

Real independence has three properties:

• Separate reporting. The person who can stop the decision does not report through the person who needs it to proceed.
• The ability to load the claim. Challenge is a proof test, not an opinion about the presentation. If the reviewer cannot load the claim, they are scoring fluency with better vocabulary.
• A different map. At least one person looks at adjacent territory: the function that will absorb the delayed cost, the operator who will run the change, the maintainer who will inherit the asset.

Closure

Agreement is not closure. A review is closed only when these are on the record: the defined problem and what solved looks like; the decision rule; the trip points and current values; the residual-risk owner; the next proof-test, dated; the reassessment triggers.

If any of those is missing, the meeting ended. The review did not.

Reassessment Triggers

A decision rule that cannot be reopened is a drawing that cannot be redlined. Specify the triggers in advance, or "we will revisit if needed" means we will not.

The minimum set: the governing variable crosses its trip point; a dated dependency slips; a proof-test fails, or its date arrives unrun; a contrary observation arrives from outside the map that produced the rule; the decision class itself changes; "solved" is quietly redefined.

When a trigger fires, the review reopens at the same standard. It does not reopen as a conversation about whether the original people still feel confident. The original definition of done is what makes a moving goalpost visible.

Manufacturing the Collision Early

Where natural feedback arrives too late to correct anyone, the review has to manufacture the collision while there is still time to act. These are not delay. They are the proof test.

Pre-mortem. Before close, write the history of how it failed as if the failure has already occurred: what we believed, which tagged figure was wrong, which trip point we did not have, who owned the residual risk, and what observation was available and unused. If the room cannot write that history, it does not yet understand the decision.

Shadow-running. Where the decision can run in parallel with the current state, it does, without authority, and is compared. Where that is available and skipped, the skip needs an owner.

Staged validation. Where it cannot be shadowed: limited scope, a hold point, a walk-down against as-built, a trial with a kill criterion that does not require reconvening the original room. If only the people who needed the decision to proceed can exercise the kill criterion, it is not one.

None of these manufactures certainty. They distinguish a claim that has been loaded from a claim that has only been presented.

What This Looks Like in Practice

An executive asks whether a date will be met. Under the current design there are two acceptable-feeling answers: yes, and a paragraph. The yes closes the meeting. The paragraph is heard as a non-answer. Four months later the inputs have moved, and the executive has two different answers with no visible reason. Both people are trying. The control does not exist.

Under this standard the work starts further back. What problem does the date serve? What does solved look like, other than a date that has not yet slipped? What is hitting it worth versus missing it? Then the date never arrives bare: dependency, trip point, confirmation date, residual-risk owner, next proof-test. When the switchgear slips, the date moves and the movement is self-explaining.

That is a smaller meeting and a shorter answer. It survives compression. The listener can keep only the rule and still have a decision they can use when the world moves.

The same shape applies to a yield loss blamed on a shift. "Second shift is the problem" is a fluent story, not a definition. The defined problem is a yield loss that tracks a substrate change. Solved is the baseline returning when the substrate is restored. The value is the original process loss plus the cost of the wrong intervention. Until those are named, every function has a different goalpost, and each one is internally coherent.

What Has to Change on the Receiving Side

A standard that only constrains the presenter will be gamed, including by presenters who mean well.

• Refuse to start if the problem, what solved looks like, and the value are not on the page.
• Refuse a bare yes. A recommendation without its envelope is incomplete the way a setpoint without units is incomplete.
• Ask what earned the confidence: measured, modeled, vendor-asserted, assumed, inherited. If the speaker cannot tag it, treat it as assumed.
• Do not score adjournment as evidence.

What This Is Not

This is not a requirement to slow every decision. The decision class exists so that most decisions keep moving.

This is not a requirement that technical people become less precise. Precision has to take a form that survives the listener. The bounded answer is the analysis, stated as a rule.

This is not a claim that procedure makes people rational. Procedure is what you use when you have already admitted they will not be, including you, including on a day you are sure you are right.

Closing Thought

When the plant disagrees with the drawing, the plant is right. You walk it down, you field-verify, you redline, you reissue as-built. The drawing was never the point. It was an approximation maintained for as long as it remained useful.

The technical review is a drawing of a decision. Under the current design, organizations defend it. They preserve the original frame and treat the close of the meeting as as-built.

A problem well defined is a problem half solved. The other half is a review that can recognize the solution when it arrives, and refuse a moving target when it does not. The review is the same kind of object as an instrument, a drawing, and a physical change. It has not been specified.

What the review requires the organization's systems to remember, so that it runs on evidence rather than memory, is the subject of Part 5.

---

Roger E. Henley II — Part 4 of a five-part series on decision quality in complex operations. Preceded by Part 1: "Why Acting on Partial Data Feels Right and Often Is Not," Part 2: "Why and How We Become Confidently Wrong," and Part 3: "When Being Right Is Not Enough." Continued in Part 5: "The Decision Infrastructure Standard." github.com/reh3376 · linkedin.com/in/rogerehenley