Evidence release

A result an independent reviewer can challenge

The capstone is not a campaign plan or visibility promise. It is a bounded research release in which every headline claim resolves to sources, queries or data, transformations, analysis, uncertainty, limitations, and a versioned decision record.

Effort
18–30 hours across course
Defense
8 min result + 7 min questions
Revision
48 hours

Question contract

One system question, one declared population, one bounded conclusion

Acceptable

“For the declared English informational query strata and frozen open retrieval stack, does evidence-first page structure change source retrieval and claim support under a fixed corpus and budget?”

Not acceptable

“How do we guarantee that all AI engines recommend Brand X?” It lacks an observable system boundary, defensible population, authorized intervention, estimand, and uncertainty.

Milestones

Review before the cost compounds

  1. W06

    Question and query universe

    Bound the system, surface, population, query strata, and evidence ceiling.

  2. W08

    Pre-analysis plan

    Freeze estimand, outcomes, exclusions, analysis, stop rules, and validity boundary before inspecting outcomes.

  3. W10

    Evidence and intervention review

    Pass claim/source, factual-equivalence, authorization, and rollback checks.

  4. W13

    Measurement release candidate

    Validate event logs, uncertainty, missingness, sensitivity, incidents, and deviations.

  5. W15

    Safety and governance gate

    Resolve abuse, privacy, accessibility, licensing, status, jurisdiction, and residual-risk findings.

  6. W16

    Defense and revision

    Present, answer evidence-lineage questions, revise within 48 hours, and regenerate checksums.

Required bundle

The release is a traceability network

01

README

Question, result, scope, environment, reproduction route, known limits.

02

Research brief

Motivation, estimand, design, finding, uncertainty, decision.

03

Claim ledger

Atomic public claims, exact evidence, disposition, reviewer state.

04

Data/query manifest

Identity, provenance, schema, license, checksum, exclusions.

05

Analysis

Code/notebook, environment lock, logs, tests, supplied-output route.

06

Figures

Source data, type label, caption, alt text/long description.

07

Deviation record

Departures from plan, incidents, missingness, negative results.

08

Risk and rights

Authorization, privacy, safety, accessibility, license, governance.

09

Release record

Version, checksums, change log, review findings, response note.

Defense and revision

Critique must be able to change the release

30%Evidence integrityComplete, non-circular claim-to-artifact lineage.
25%Methods and reproductionPrimary result rebuilds under the declared route.
20%Inference and limitationsUncertainty, alternatives, and scope control the conclusion.
15%Safety and governanceAuthorization, privacy, license, accessibility, and risk gates pass.
10%Defense and revisionAnswers locate evidence; critique produces material revision.

The 48-hour response note names each finding, action, changed artifact, and remaining disagreement. At least one claim must be explicitly reconsidered; weakening or removing it is a sign of evidence control, not failure.