Decision archaeology: preserve the assumptions, not just the decision

Most decision records preserve the least interesting part: the sentence at the end. We chose the vendor. We delayed the launch. We built instead of bought. The conclusion survives, while the conditions that made it reasonable disappear.

Months later, the decision looks either sacred or stupid. A successor may obey it because it is written down, or reverse it because the rationale is invisible. Both reactions mistake the residue for the reasoning.

Decision archaeology is the disciplined recovery and preservation of the decision frontier: what the decision-makers were trying to achieve, what they knew at the time, what they did not know, which alternatives were live, which assumptions carried the choice, who had authority, and what future evidence was supposed to trigger review.

The decision tells you what happened. The assumptions tell you whether it should still stand.

What the technique preserves

A useful record separates six things that ordinary minutes tend to collapse.

1. The objective: the outcome the decision was meant to advance, including the success measure and time horizon.
2. The decision frontier: the information, constraints, and uncertainty that were actually available before the outcome was known.
3. The option set: the alternatives seriously considered, including “wait,” “do nothing,” and experiments that could defer commitment.
4. The rationale: why one option fit the objective and constraints better at that moment.
5. The authority: who recommended, advised, objected, approved, and owned execution.
6. The standing: whether the decision is proposed, active, superseded, expired, or under review—and what changed that status.

This is not a demand to record every thought. It is a demand to preserve the load-bearing ones. An assumption is load-bearing when changing it could change the choice. A rejected option matters when it may become viable under different conditions. A dissent matters when it names a failure mode the majority accepted or overlooked.

What existing practice already establishes

Software architecture has a mature lightweight precedent. Michael Nygard’s original architecture decision record format preserves context, decision, status, and consequences. It explicitly keeps reversed decisions and marks them superseded, because the old choice still explains the system even after it stops governing the future. MADR extends the pattern with considered options and an explicit decision outcome.

NASA’s Risk-Informed Decision Making handbook supplies a higher-stakes version of the same logic. It begins with stakeholder objectives and performance measures, compiles feasible alternatives, analyzes them under uncertainty, and ends with selection and documented rationale. The point is not that every ordinary decision needs aerospace analysis. It is that a conclusion cannot be evaluated independently of its objective, alternatives, measures, and uncertainty.

The W3C PROV standard contributes a different layer: provenance can represent the entities used, the activities that transformed them, the agents responsible, and relations such as derivation, attribution, primary source, and revision. It does not prescribe a decision method. It establishes a vocabulary for keeping an output attached to where it came from and who did what.

Experimental hindsight research explains why this preservation must happen before the outcome colonizes the story. Baruch Fischhoff found that outcome knowledge changed judged likelihoods and the perceived relevance of earlier facts; participants were largely unaware of that influence and overestimated what they or others could have known beforehand. A polished retrospective rationale is therefore not a neutral reconstruction.

A 2023 Software Engineering Institute report adds the need for continuous evaluation. Architectural choices interact across goals and can create both positive and negative tradeoffs; analysis should continue as systems and evidence change. Together, these sources support the components of a decision record. They do not prove the generalized protocol below or an AI’s ability to recover missing rationale accurately.

The protocol: capture, challenge, expire

1. Name the decision and its owner

Write the question in a form that could produce more than one answer. Identify the accountable decision-maker, the people supplying evidence, and anyone with approval or veto authority. “The team decided” is not enough when authority is distributed or contested.

2. Freeze the decision frontier

Timestamp what was known, estimated, disputed, missing, and deliberately not investigated. Link the source artifacts actually used. Distinguish contemporaneous evidence from material added after the outcome. If the record is reconstructed later, label it reconstructed.

3. State the objective and constraints

Record the purpose, success measures, time horizon, non-negotiable constraints, and acceptable downside. Separate real constraints from preferences and inherited habits. If two stakeholders optimize different objectives, preserve the conflict instead of manufacturing one shared goal.

4. Preserve the live option set

List the alternatives that received serious consideration and why each remained viable or was rejected. Include waiting, staged trials, reversible pilots, and no action when they were genuine choices. Do not add imaginary alternatives later to make the process look thorough.

5. Build an assumption and uncertainty register

For every load-bearing assumption, record its evidence, confidence, owner, test, and expiry condition. Ask the decisive counterfactual: If this assumption were false, would we choose differently? If yes, it belongs in the record and needs a review trigger.

6. Record the choice, rationale, dissent, and consequences

State the choice in active language. Connect each important reason to the objective or constraint it serves. Preserve known costs, neutral consequences, residual risks, and dissent without turning disagreement into biography. A decision can be reasonable while accepting serious downside.

7. Define standing and revision triggers

Give the decision a status and scope. Set review triggers based on changed evidence: a cost threshold, failed test, altered regulation, new capability, expired dependency, or missed outcome. Calendar reviews help, but a date alone does not explain what would justify reversal.

8. Supersede; do not overwrite

When the context changes, create a linked successor. Preserve the original decision as historically true, mark its present standing, and record which assumption or constraint changed. Keep outcomes and corrections separate from the original rationale so hindsight does not rewrite the frontier.

A synthetic example

A small publication chooses a managed email service instead of building its own delivery system. The surviving note says only, “Use managed email.” Two years later, an engineer sees higher fees and proposes replacing it.

The archaeological record changes the question. The original objective was reliable delivery with almost no administrative burden. The live assumptions were fewer than 500 recipients, irregular sends, no dedicated operations staff, and compliance features more valuable than custom behavior. Self-hosting and a custom provider integration were rejected because their fixed maintenance cost exceeded the expected benefit. The review triggers were sustained volume above 10,000 recipients or the need for delivery behavior the managed service could not support.

If volume is still 400, the context remains. If it is 25,000, a load-bearing assumption has expired. The old decision was not exposed as foolish; its standing became testable.

Where AI helps—and where it must stop

With authorized access, an AI can gather scattered references, compare minutes with the final record, expose missing owners or assumptions, identify options mentioned but omitted from the decision note, and monitor linked evidence for review triggers. It can also ask the counterfactual questions people skip once consensus feels close.

It should not pretend that recovery is capture. A later model can propose candidate assumptions from surrounding records, but those candidates must be labeled inferred, linked to evidence, and confirmed by someone who had standing at the time. It must not reconstruct quotations, attribute motives, diagnose participants, or use private correspondence merely because it is searchable.

The safest AI output has three layers: preserved material directly supported by contemporaneous evidence; recovered material drawn from dated surrounding records; and hypothesized rationale requiring confirmation. Fluency must never erase those labels.

Failure modes

• Hindsight laundering. The observed outcome is smuggled into the account of what was knowable beforehand. Countermeasure: freeze the decision frontier and mark later evidence separately.
• Rationale reconstruction. A plausible explanation is written as the actual explanation. Countermeasure: label recovered and hypothesized rationale; never reconstruct quotations.
• Option erasure. Rejected or politically inconvenient alternatives disappear, making the winner look inevitable. Countermeasure: preserve only the genuinely live option set and why options left it.
• Authority blur. Advice, consent, approval, and execution are collapsed into “we decided.” Countermeasure: record roles and standing explicitly.
• False precision. Confidence scores and dates are invented to make a weak record look analytic. Countermeasure: allow ranges, unknowns, and qualitative confidence tied to evidence.
• Documentation theater. A long template is completed after the choice to justify rather than examine it. Countermeasure: capture the load-bearing record before commitment and keep it short enough to maintain.
• Bureaucratic drag. Every trivial choice receives a full dossier. Countermeasure: use the protocol for decisions that are costly, hard to reverse, contested, precedent-setting, or dependent on volatile assumptions.
• Zombie standing. A decision remains active after its assumptions expire because nobody checks the triggers. Countermeasure: make standing queryable and assign trigger ownership.
• Privacy leakage. Personal messages become permanent rationale evidence without consent. Countermeasure: minimize sources, summarize sensitive support, restrict access, and separate the public decision record from private evidence.

A reusable prompt

Who contributed what

What remains unproven

This article assembles a protocol from established components and one project-level use. It does not establish that decision archaeology improves decision quality, prevents repeated mistakes, or produces better reversal timing. It does not show that an AI can reliably infer missing rationale.

The method earns confidence prospectively: capture a decision before its outcome, later detect a changed assumption, surface the correct review trigger, and help a human revise or retain the decision for reasons visible in the record. Until then, it is a disciplined way to make decisions challengeable—not a machine for declaring what the past meant.