Knowledge & Evidence

Source identity comes first

A source needs an identity that survives a summary: a title or record label, a version when known, a relevant span and a note about what it can support. A citation alone is not an argument, but it makes the argument inspectable.

When a source changes, the earlier claim should not silently inherit the new wording. Version awareness makes revision an explicit event instead of a hidden rewrite.

Technique: Record each source as (id, locator, version/date, relevant span, support role). A source ledger is a collection of these records. Fieldcraft article FC-012 demonstrates detecting drift when a source revision identifier no longer matches the expected value.

Claims are not evidence

Evidence is an observation, record or source span. A claim is an interpretation about what that material supports. Separating them creates room for a claim to be weak, contested, incomplete or simply premature without damaging the record of what was actually observed.

In a synthesized example, two sources may agree on a measurement while disagreeing about its cause. The shared observation remains; the interpretation stays plural.

Failure mode to avoid: The phrase "the data shows" is almost always a claim conflated with evidence. A more precise formulation: "In study X at date Y, measurement Z was observed under conditions C. This may support claim P, but alternative interpretations include Q and R."

Absence and contradiction are different

A missing source does not negate a claim. A contradictory source does not disappear because it is inconvenient. These are different states with different next actions: retrieve more evidence, resolve a definition, preserve a minority conclusion or hold a decision.

That distinction is especially valuable in AI-assisted synthesis, where fluent prose can make a gap look like a conclusion unless the record gives uncertainty its own place.

Worked example: A literature review of 12 sources finds 10 supporting claim C and 2 that are silent (neither support nor contradict). The reviewer labels support, silence and contradiction separately. A later reader can see that 2 sources were silent and investigate why — perhaps the relevant data was never collected, or the sources address a different population. Fieldcraft article FC-014 demonstrates separating contradiction from absence.

A review is its own record

Review should capture who or what evaluated a candidate, against which edition, under which criteria and with what disposition. It should not overwrite the underlying source or convert confidence into authority.

This makes it possible to revise a claim, re-check a source and retain the history of a prior review without claiming that older work never existed.

Counterexample: A reviewer says "this is good to go" without recording criteria. Later, a different reviewer rejects the same material under different criteria. Without a review record, the first finding cannot be distinguished from the second, and the work cycles without accumulating useful information.

Synthesis preserves dissent

When combining multiple sources into a synthesis, the goal is not to produce the most average claim. It is to show the range of evidence, the points of agreement, the points of disagreement and the open questions.

Fieldcraft article FC-015 demonstrates preserving minority conclusions during synthesis. A synthesis of 5 sources with 4 supporting and 1 dissenting preserves the dissent rather than burying it in a consensus statement.

Projection losses are recorded

Every time knowledge moves from one representation to another — source → summary, summary → synthesis, synthesis → recommendation — information is lost. The projection loss report makes those losses explicit.

Fieldcraft article FC-024 demonstrates a projection loss checker. Given a source record and a projected record, it identifies fields present in the source that are absent in the projection, and fields where precision, granularity or context is reduced.

Counterevidence methodology

Counterevidence is not the opposite of supporting evidence. It is a distinct role in reasoning: a finding that would change the claim if it held, but has not done so yet. A methodology for handling counterevidence includes: (1) stating the claim in falsifiable form, (2) defining what would count as counterevidence, (3) actively searching for it, and (4) reporting what was found or not found.

Technique: falsifier-first research. Before gathering supporting evidence, write down what would disprove the claim. This protects against confirmation bias — the tendency to find what you expect to find. FC-014 demonstrates detecting contradiction versus absence.

Worked example: Claim: 'A structured reasoning prompt improves output quality over an unstructured baseline.' Falsifier: 'The structured method produces outputs rated equal or lower than the baseline on the same rubric with the same model.' The counterevidence search includes running both the method and baseline on the same inputs with a blind evaluator. If no significant difference is found, the claim is not disproven — it remains unsupported and should not be propagated as confirmed.

Boundary: Searching for counterevidence does not guarantee finding it. The absence of counterevidence is not proof. It is a condition that changes confidence but does not eliminate the need for further testing or alternative explanations.

The evidence hierarchy

Not all evidence is equal. A direct observation from a controlled experiment carries more weight than an anecdote from a blog post. An expert review of a published artifact carries more weight than a passing mention in a conference talk. The evidence hierarchy makes these distinctions explicit without turning them into rigid rules.

Hierarchy (synthesized, not exhaustive): (1) Direct, reproducible observation with a named methodology. (2) Published peer-reviewed analysis. (3) Expert review of a specific artifact. (4) Documented field experience from multiple sources. (5) Single-source field report. (6) Opinion or speculation by a domain expert. (7) Unattributed claim or generic assertion.

Practical technique: When citing a source, note its level in the hierarchy. A claim supported by level 4 evidence should not be presented as if it were level 1. The label helps a reader calibrate their trust and understand what would be needed to move the claim up the hierarchy.

Cross-source reconciliation

When two sources produce contradictory evidence, the first step is not to decide which is right — it is to determine whether they are measuring the same thing. Contradiction can arise from different definitions, different measurement instruments, different populations, or different conditions. Each has a different resolution path.

Reconciliation technique (synthesized): (1) Normalize: are both sources using the same definition of key terms? If not, the apparent contradiction may be a definitional artifact. (2) Compare methodology: did they measure differently? If the instruments differ, compare their accuracy and calibration. (3) Compare scope: did they study the same population or conditions? If not, the results may not be contradictory — they may describe different contexts. (4) Report: if the contradiction survives normalization, preserve both findings with the unresolved disagreement labeled. Do not merge them into a false consensus.

Source versioning and drift detection

Sources change. A specification gets a new revision. A research paper is retracted. A dataset is corrected. If you cited the earlier version, your claim should not silently inherit the new content. Source versioning makes revision an explicit event.

Technique: Record each source with its version identifier (date, commit hash, edition number, DOI). When you reference a source, include the version. FC-012 demonstrates detecting drift when a source revision identifier no longer matches the expected value.

Worked example: A specification v2.3 defines an API endpoint. You write a check that validates responses against v2.3. Six months later, the specification advances to v3.0 with breaking changes. Your check still passes against v2.3 — it is not wrong. But it is testing the wrong version. A drift detector would flag the mismatch between the expected version (v2.3) and the current version (v3.0), alerting you that the check needs updating.

Confidence and uncertainty labeling

Every claim carries a confidence, even when the confidence is not stated. Making confidence explicit prevents a claim from being treated as more certain than the evidence supports. The confidence label should be calibrated to the evidence hierarchy level and the number of corroborating sources.

Labeling scheme (synthesized): Confirmed — multiple independent sources agree under controlled conditions. Supported — multiple sources agree but with methodological limitations. Indicative — single source with reasonable methodology. Inconclusive — sources disagree or evidence is insufficient. Speculative — no direct evidence, extrapolation from related domains.

Counterexample: A synthesis that says 'research shows X' without labeling whether X is confirmed, supported, or merely indicative. The reader cannot distinguish a well-established finding from a preliminary result. The confidence label protects both the author (from overclaiming) and the reader (from overtrusting).

The evidence preparation workflow

Turning raw sources into a defensible knowledge artifact follows a reproducible workflow: (1) Assemble sources with version identifiers. (2) Extract relevant spans with source tags. (3) Classify each span as observation, inference, hypothesis, or proposal. (4) Build a claim map: which claims are supported, contradicted, or not addressed. (5) Write the synthesis with explicit confidence labels and preserved dissent. (6) Produce a projection loss report — what was simplified or omitted in the translation from source to synthesis.

Practical note: Step 6 (projection loss) is the most commonly skipped and the most valuable. Every translation loses information. Recording what was lost lets a later reader decide whether the loss matters for their question. FC-024 demonstrates a projection loss checker.