Evidence quality must survive publication

Publication should expose the decision argument within its evidence authority. Identify public primary guidance, original analysis, illustrative input assumptions and canonical outputs where each claim appears. A primary methodological statement does not validate commercial payoffs; arithmetic parity does not demonstrate field performance. Keeping these distinctions beside the result lets a reader inspect its basis without inheriting an unsupported commercial or empirical conclusion.

The canonical manuscript serves as the common narrative for the PDF and chapter reader. This avoids maintaining separate summaries whose claims drift from the worked study. A quantitative block refers to named ledger tokens, which in turn bind to a specific model fixture or source-supported claim. The renderer resolves those tokens into the displayed values. The method gives reviewers a concrete route from sentence or table cell to its basis. It does not replace semantic review of whether the basis supports the interpretation.

The frozen numerical object consists of an explicit input fixture, its canonical engine output and scalar bindings. The output is retained as full structured data, rather than only the displayed rounded cells. A reader can reproduce the app computation and compare it with the frozen object. Bindings identify the values selected for publication and their pointers. This is especially useful when a chart or table shows only part of a large output. Reproducibility depends on preserving the selected computation and the meaning of its inputs together.

Keep the illustrative label, title, basis, units and limitation attached to each quantitative exhibit. A copied table or independently read chapter may lose the paper-wide note. The local disclosure must therefore prevent the figures being mistaken for observed industry rates, actual client cases or quoted contracts. These examples contain no private or empirical records from which such an interpretation could be supported.

Source cards should be narrow and locator-specific. Their purpose is to show what public material was actually retrieved, what limited claim it can support and what it cannot establish. A source title alone does not prove that a report's interpretation is justified. The record should retain the publisher, edition, exact URL, retrieval object and relevant location. If only a publisher description was readable, the card should say so. An inaccessible full document should not be represented as read merely because another page cites its title.

The shared catalogue holds one public source record for reuse across papers. References retain its scope and limitation; narrow local notes connect it to the decision without repeating long guidance paraphrases. Original mechanism and fixture reasoning must remain independent analysis. Keep a sourced assertion sourced even when its word budget or citation obligation is inconvenient: the claim record follows its substance, not the label that reduces production work.

Rounding should be declared and should not change the stored result. The paper displays selected monetary values at a consistent precision and retains raw outputs in the replication files. A probability stored as a fraction can be displayed as a percentage when the binding format and units make that conversion explicit. This prevents a display convention from silently becoming a new computational input. A reviewer should be able to understand a rounding difference without needing to infer whether the model, table or narrative used a different fixture.

Figures need their own reproduction object. A quantitative chart can have correct-looking axes while displaying no actual series, a clipped point or a value that lacks a reciprocal ledger binding. The plot specification should identify the selected output pointers, drawing ranges and labels. The shared renderer recreates the SVG and compares exact bytes, rejecting empty series and invalid coordinates. This checks the graphical computation. It does not establish that the author chose a useful comparison or wrote an accurate caption; those remain independent review questions.

Reopen the PDF: declared pages and successful rendering can conceal thin, duplicated or clipped content. The audit identifies chapter boundaries, counts qualifying body pages, checks text geometry and fonts, and renders every page for visual review. Its receipt must name the checked properties and remaining semantic or visual judgements. A mechanical pass leaves a review candidate; independent acceptance requires review of the substantive artifact.

The reader should have access to limitations with the result rather than only in an appendix. The app calculates a fixed binary risk-neutral decision with supplied likelihoods and aggregate payoffs. It does not implement causal estimation, utility elicitation, joint signals, adaptive stopping or a dated delay model. A chapter discussing those methods should state their conceptual status. Publishing a diagram of a broader framework does not make the app calculate it. The paper's claim scope should match the actual implemented methods, not the ambition of its contents list.

Publication should preserve failed retrieval and correction evidence privately where appropriate. A source archive can record a blocked URL alongside a successful publisher response, without presenting the failure as verification. A challenge log can retain withdrawn interpretations and their disposition. The public card should describe the current narrow support honestly. The existence of a failure log is not a defect by itself; it can show that the team distinguished absence of evidence from successful access. What matters is that unsupported content is not silently certified.

The final edition needs a stable identity. The manuscript, sources, claim ledger, computation, PDF, chapter fragments and relevant proof receipts should have consistent hashes. A change to a numerical input requires recomputation and a new binding review. A change to the renderer can require a fresh artifact audit even when the prose is unchanged. A change to the source interpretation can require a new semantic review. These distinctions allow an institution to know which properties of an earlier review still apply to the revised edition.

Record mechanical pass, author proof review, independent methodological review and public delivery as separate states of the frozen edition. Model reproduction or a page threshold closes only its checked property. Independent reviewers still need the full artifact and claims; pending dimensions remain pending until their evidence exists. Updating status on that basis prevents a production milestone from being published as certification or substantive endorsement.

An edition drill should follow a hypothetical corrected signal interpretation across every public object. The raw output can remain correct while a chart legend has reversed the branch labels; repair the legend and its explanatory claim, then verify the displayed response against the retained branch account. If the raw likelihood input was wrong instead, replacing only the PDF leaves the chapter reader and exported response policy stale. Trace fixture, output, binding, figure and displayed conclusion to the same edition identity. This drill turns publication integrity into a specific object-consistency test while preserving the separate question of whether the retained primary evidence supports the premise at all.