decision-biases

Decision biases — how they corrupt a matrix, and how to counter them

A weighted matrix does not make a decision objective. It makes the judgment *visible*, which is more useful — but only if you actively defend against the biases that quietly bend the weights and scores. Each bias below comes with its **tell** (how to notice it) and a **structural counter** (a step baked into the process, not just "try harder to be rational").

Criteria-fishing / rationalization

**What it is.** Choosing or weighting the criteria *after* you already know which option you want, so the matrix confirms a pre-made decision. The most dangerous bias here because it wears the costume of rigor. **Tell.** The weights were set (or quietly adjusted) once the scores were in view. The "winner" happens to be the option someone favored going in. **Counter.** Set and justify the weights **before** scoring, ideally with a second person, and don't touch them afterward. If a weight genuinely needs to change, change it openly and re-run the whole thing — and be suspicious of any change that just happens to flip the result your way.

Anchoring

**What it is.** The first number/option seen drags all subsequent judgments toward it — the first option scored becomes the yardstick; a vendor's list price anchors the "fair" price. **Tell.** Early options cluster at similar scores; later options are judged as "better/worse than the first" rather than on their own merits. **Counter.** Score **column by column** (one criterion across all options at once) using an anchored scale with concrete descriptors, so each cell is judged against a fixed reference, not against whichever option you saw first.

Confirmation bias

**What it is.** Seeking and over-weighting evidence that supports the option you like; ignoring or explaining away evidence against it. **Tell.** The favored option's weak scores all have a "but actually…" excuse; the disfavored option's strengths are dismissed as edge cases. **Counter.** For the *leading* option, deliberately go find the disconfirming evidence and the failure mode (a mini-premortem). Ask "what would have to be true for this to be the wrong choice?" and check it.

Halo effect

**What it is.** One salient strength (or the brand/hype of an option) inflates its scores on unrelated criteria. A tool that's great at X gets generously scored on Y and Z it isn't actually good at. **Tell.** One option scores uniformly high with little cell-level justification; the scores feel like a gestalt impression rather than per-criterion evidence. **Counter.** Require a one-line, evidence-based justification for **every** cell. Score independently per criterion. A high score with no specific reason is a halo, not a finding.

Sunk cost / status-quo bias

**What it is.** Favoring the option you've already invested in (time, money, code, identity), or the current state, because switching *feels* like waste — even when the past investment is irrecoverable and irrelevant to the forward decision. **Tell.** "We've already put six months into X" appears as an argument; the status quo isn't scored on the same terms as the alternatives. **Counter.** Decide only on **forward** costs and benefits — sunk costs are gone regardless of choice. Always include the status quo as an explicitly scored option so it competes on the merits, not by default inertia.

False precision / over-quantification

**What it is.** Treating small differences in weighted totals as meaningful, or believing that assigning numbers made the judgment objective. The scores are 1-significant-figure opinions; the total cannot carry more certainty than its inputs. **Tell.** "Option B wins, 3.47 to 3.42." A 0.05 gap on a 1–5 scale reported as a verdict. **Counter.** Always report the **margin** and run the **sensitivity check** (`weighted-decision-matrix.md`). If the winner sits inside the noise band or flips on a small, defensible re-weight, say "effectively tied — decide on the trade-off," not "B wins."

Groupthink / authority anchoring

**What it is.** In a group, weights and scores converge on whatever the most senior or loudest voice said first; dissent is suppressed. **Tell.** Suspiciously fast consensus; the weights match the boss's stated preference. **Counter.** Collect weights and scores **independently first** (silent, written), then compare and discuss the divergences — the disagreements are the most informative part.

Quick map: bias → structural counter

| Bias | Structural counter (a workflow step, not willpower) | |---|---| | Criteria-fishing | Fix and justify weights before scoring; change them only in the open | | Anchoring | Anchored 1–5 scale; score column-by-column | | Confirmation | Actively disconfirm the leading option; premortem | | Halo | One evidence-based justification per cell; independent per-criterion scoring | | Sunk cost / status quo | Forward costs only; score the status quo as a real option | | False precision | Report the margin; always run sensitivity analysis | | Groupthink | Independent silent scoring first, then reconcile |

The through-line: **make each judgment explicit, evidence-backed, and set before you can see which answer it produces.** That is what the matrix is for.