Results Report

Results Report

Write the complete post-experiment summary report after analysis artifacts are ready.

This skill is for the stage after results-analysis.

Role boundary

results-analysis does

• strict statistics,
• real figures,
• figure interpretation scaffolding,
• stats appendix.

results-report does

• complete experiment wrap-up report,
• decision-oriented narrative,
• figure-by-figure interpretation inside a coherent structure,
• limitations, failure cases, and next actions,
• Obsidian write-back into Results/Reports/.

When the task is to create or redesign paper-ready figures/tables themselves, rely on publication-chart-skill instead of expanding results-report into figure/table production.

Do not replace strict analysis with confident prose. If the analysis bundle is missing, first identify the blocker and request or produce the missing bundle.

Hard gate: do not write a complete report when analysis-report.md, stats-appendix.md, figure-catalog.md, provenance, or the primary comparison question is missing. In that case, write only a blocker summary and route back to results-analysis.

Default output

The default report is an internal research report, not manuscript prose.

It should be named as:

YYYY-MM-DD--{experiment-line}--r{round}--{purpose}.md

Example:

• 2026-03-18--freezing--r03--transfer-summary.md
• 2026-03-18--contrastive-adversarial--r02--ablation-report.md

The note title should be:

{Experiment Line} / Round {N} / {Purpose} / {YYYY-MM-DD}

Read references/report-naming.md before finalizing the filename or note title.

Required frontmatter

---
type: results-report
date: 2026-03-18
experiment_line: freezing
round: 3
purpose: transfer-summary
status: active
source_artifacts:
  - analysis-output/analysis-report.md
  - analysis-output/stats-appendix.md
linked_experiments:
  - Experiments/Freezing-Study.md
linked_results:
  - Results/Freezing-vs-Adapter.md
---

Default report structure

The report must include all sections below.

1. Executive Summary 2. Experiment Identity and Decision Context 3. Setup and Evaluation Protocol 4. Main Findings 5. Statistical Validation 6. Figure-by-Figure Interpretation 7. Failure Cases / Negative Results / Limitations 8. What Changed Our Belief 9. Next Actions 10. Artifact and Reproducibility Index

Read references/report-structure.md before writing.

Workflow

1. Confirm the report object

Lock these fields first:

• date,
• experiment line,
• round,
• purpose,
• linked experiment note,
• linked durable result note if one already exists.

If round is unknown, do not silently invent a semantic round. Use r00 only as a temporary placeholder and state that it should be normalized later.

2. Read the strict analysis bundle

Minimum required inputs:

• analysis-report.md
• stats-appendix.md
• figure-catalog.md
• actual figures, if available

If these are missing, either generate them first with results-analysis or explicitly state which claims cannot be supported. Do not turn loose CSV/Markdown artifacts into a polished report unless the strict analysis bundle has already validated unit of analysis, primary metric, and claim wording.

3. Write the report as a decision object

This report is not a transcript of outputs.

Each section must answer a real question:

• What did we test?
• What changed numerically?
• What is actually supported?
• What failed or remains uncertain?
• What should we do next?

Read references/decision-oriented-analysis.md for the expected reasoning depth.

4. Interpret figures inside the report

Do not only attach figures.

For each main figure:

• introduce why it is included,
• state the key observation,
• explain the supported interpretation,
• explain the decision implication.

Read references/figure-interpretation.md and references/statistical-completeness.md as needed.

5. Choose the write target explicitly

If the current repo is bound to an Obsidian project knowledge base:

• create or update Results/Reports/{report-name}.md,
• link back to the relevant Experiments/ note,
• update the matching canonical Results/ note when a durable conclusion is now supported,
• append a short trace to today's Daily/ note,
• update .claude/project-memory/<project_id>.md.

If the repo is not bound:

• write the report as a local markdown artifact in the requested output location or next to the analysis bundle,
• keep the same filename contract,
• explicitly say that no Obsidian write-back was attempted.

Use obsidian-project-kb-core conventions only for bound repos. Internal experiment reports belong in Results/Reports/, not Writing/.

6. End with explicit next actions

The report must end with operational decisions, for example:

• stop a weak branch,
• schedule one missing ablation,
• promote a stable finding into manuscript-facing writing,
• update the active plan.

Required quality bar

• The report must be dateable, searchable, and attributable to one experiment line and one round.
• The report must cite actual evidence from the analysis bundle.
• The report must include negative results when they matter.
• The report must separate stable conclusion from tentative interpretation.
• The report must say what changed in project belief and what should happen next.
• The report must preserve Claim Candidate wording and must not promote speculative or observed claims into decisive conclusions.

Reference files

Load only what is needed:

• references/report-structure.md
• references/report-naming.md
• references/figure-interpretation.md
• references/statistical-completeness.md
• references/decision-oriented-analysis.md
• references/EVIDENCE-PROPAGATION.md
• examples/example-results-report.md

Freezing / Round 3 / transfer-summary / 2026-03-18

Executive Summary

• Round 3 tested whether a subject adapter recovers the performance lost by freezing most of the encoder.
• Across 5 seeds per condition, the adapter reduced mean WER by 3.8 absolute points relative to the frozen encoder baseline.
• The current evidence supports keeping the adapter branch active, while pure freezing should be deprioritized.

Experiment Identity and Decision Context

• Experiment line: freezing
• Round: 3
• Purpose: resolve whether the freezing gap is best handled by lightweight adaptation or by abandoning the freezing branch.
• Decision pressure: choose the next transfer branch before scheduling the next low-resource run block.

Setup and Evaluation Protocol

• Same subject pool and split as rounds 1-2.
• 5 seeds per condition.
• Primary metric: WER (lower is better).
• Compared methods: Full fine-tuning, Subject Adapter, Frozen Encoder.
• Statistical unit: seed-level final WER.

Main Findings

• Subject Adapter: 27.6 ± 1.0 WER, 95% CI [26.4, 28.8].
• Frozen Encoder: 31.4 ± 1.5 WER, 95% CI [29.6, 33.2].
• Full fine-tuning: 25.9 ± 0.8 WER, 95% CI [24.9, 26.9].
• Adapter beats Frozen Encoder in all 5 paired seed comparisons.

Statistical Validation

• Adapter vs Frozen Encoder: paired Wilcoxon signed-rank test, p = 0.031, Holm-corrected p = 0.047, matched-rank biserial effect size r = 0.90.
• Full fine-tuning vs Adapter: paired t-test, p = 0.11, Cohen's d = 0.64.
• Interpretation: the adapter gain over pure freezing is supported at current n = 5; the gap to full fine-tuning is directionally consistent but still underpowered.
• Unsupported claim boundary: this report does not claim generalization beyond the current subject pool or low-resource regime.

Figure-by-Figure Interpretation

Figure 1 — Main comparison

• Why included: this is the core decision figure.
• Evidence carried in: mean WER, 95% CI, and paired-seed comparisons.
• Supported interpretation: lightweight subject adaptation closes most of the freezing gap.
• Decision implication: future transfer experiments should center on adapter design, not frozen-only variants.

Figure 2 — Training dynamics

• Why included: to explain stability differences.
• Evidence carried in: per-epoch validation traces across seeds.
• Supported interpretation: the frozen baseline oscillates more after epoch 8, matching its wider uncertainty interval.
• Decision implication: branch weakness is not only lower final accuracy but also worse optimization stability.

Failure Cases / Negative Results / Limitations

• Full fine-tuning still leads in absolute WER.
• The evidence is limited to one subject pool and 5 seeds.
• No low-resource stress test or out-of-domain subject split has been run yet.
• Adapter width was fixed in this round, so capacity trade-offs remain unresolved.

What Changed Our Belief

• Before round 3, it was plausible that freezing should be abandoned entirely.
• After round 3, the better hypothesis is that freezing alone is too rigid, but freezing plus lightweight adaptation remains viable.

Next Actions

• Run one low-resource robustness check for the adapter branch.
• Add a width ablation around the current best adapter size.
• Update the canonical result note for adapter-improves-transfer.

Artifact and Reproducibility Index

• analysis-output/analysis-report.md
• analysis-output/stats-appendix.md
• analysis-output/figure-catalog.md
• analysis-output/figures/figure-01-main-comparison.pdf
• analysis-output/figures/figure-02-training-dynamics.pdf

Decision-Oriented Analysis

The purpose of a post-experiment report is not only to record what happened. It should change the project's next decision.

Required final questions

• What should stop?
• What should continue?
• What should be tested next?
• What should be promoted into a durable result note?
• What, if anything, is ready for manuscript-facing writing?

Good closing pattern

• “This round supports X.”
• “It does not yet resolve Y.”
• “The main blocker is Z.”
• “Therefore the next concrete action is A.”

Evidence Propagation

Use this file to keep results-analysis outputs aligned with the final report.

Mapping rule

• analysis-report.md -> main findings and narrative summary
• analysis-report.md#Claim Candidates -> claim wording, uncertainty, and decisions that can be carried into the report
• stats-appendix.md -> test choice, uncertainty, effect size, correction rule
• figure-catalog.md -> figure purpose and per-figure interpretation scaffolding
• figure files -> visual evidence cited in Figure-by-Figure Interpretation

Minimum statistical carry-over

Every strong claim in a results report should preserve:

• sample size or run/seed count,
• metric definition,
• uncertainty summary,
• test name,
• effect size when relevant,
• multiple-comparison handling when relevant.

Unsupported claim rule

If the analysis bundle does not support a claim strongly enough, keep the claim tentative and say why. Do not upgrade a suggestive result into a decisive conclusion during report writing.

Claim candidate carry-over

Every claim carried from results-analysis should preserve:

• the source evidence,
• the allowed wording,
• the forbidden stronger wording,
• uncertainty,
• next check or decision.

Do not convert speculative or observed claims into decisive conclusions. The What Changed Our Belief section should cite either a Claim Candidate or an Evidence Record, not only free-form prose.

Figure Interpretation in Results Reports

A results report should not dump figures.

For each major figure, write four blocks:

• Why this figure exists
• What to notice
• What interpretation is supported
• What this changes in the project decision

Example micro-structure

Figure X

• Purpose: compare adapter vs freezing under the same transfer setting.
• Observation: adapter improves mean WER and reduces variance.
• Interpretation: subject-specific adaptation likely resolves part of the transfer mismatch.
• Decision implication: prioritize adapter ablations before expanding frozen-only variants.

Report Naming Standard

Filename

Use:

YYYY-MM-DD--{experiment-line}--r{round}--{purpose}.md

Rules:

• date must be the report date,
• experiment-line should be short and stable,
• round should be zero-padded only if that is already the project convention; otherwise r3 / r03 are both acceptable if used consistently,
• purpose should describe why the report exists, not a vague label like summary unless that is truly the purpose.

Recommended purpose values:

• transfer-summary
• ablation-report
• failure-analysis
• robustness-check
• round-review

Title

Use:

{Experiment Line} / Round {N} / {Purpose} / {YYYY-MM-DD}

Frontmatter fields

Required:

• type: results-report
• date
• experiment_line
• round
• purpose
• status
• source_artifacts
• linked_experiments
• linked_results

Placement in Obsidian

Internal reports go to:

Results/Reports/{filename}

Do not put internal experiment reports in Writing/ unless they are already manuscript/slides/rebuttal material.

Results Report Structure

1. Executive Summary

Answer:

• what was tested,
• what the highest-confidence conclusion is,
• what decision this changes.

2. Experiment Identity and Decision Context

Answer:

• which experiment line this belongs to,
• why this round was run,
• what prior uncertainty or decision it was meant to resolve.

3. Setup and Evaluation Protocol

Answer:

• datasets / subjects / splits,
• methods compared,
• primary metrics,
• repeated-run structure,
• any deviations from prior protocol.

4. Main Findings

Answer:

• what changed most,
• which comparison matters most,
• where the largest gains or failures appear.

5. Statistical Validation

Answer:

• what evidence supports the major claims,
• what tests were used,
• where the evidence is weak.

6. Figure-by-Figure Interpretation

For each main figure:

• why it is shown,
• what to notice,
• what is supported,
• what remains uncertain.

7. Failure Cases / Negative Results / Limitations

Answer:

• what did not work,
• what instability appeared,
• what limits the current conclusion.

8. What Changed Our Belief

Answer:

• which prior hypothesis is strengthened,
• weakened,
• or still unresolved.

9. Next Actions

Answer:

• stop / continue / ablate / scale / write.

10. Artifact and Reproducibility Index

List:

• source artifacts,
• figure paths,
• scripts/logs,
• linked Obsidian notes.

Statistical Completeness for Results Reports

A report may summarize statistics, but it must not silently weaken them.

Always carry forward

• sample size / seed count
• metric direction
• descriptive statistics
• uncertainty estimate
• test choice
• effect size
• correction rule when relevant
• evidence boundary

Never do this in the report

• upgrade a trend to a conclusion
• omit the sample size
• replace effect size with adjectives like “large” without numbers
• cite a figure without saying what the uncertainty represents