Validation Results — Proofpunk v1.0.0

All validation was run against the real plugin source. No mocks, no stubs, no simulated results. Dates of runs: 2026-08-08 (initial pass + policy enforcement pass), 2026-08-09 (skills-ref.zip second-source pass).

1. Manifest Validation — PASS

• marketplace.json and plugin.json parse as valid JSON.
• Required fields present: marketplace name/owner/plugins; plugin name/version/description.
• Names are kebab-case and not on the reserved-name list.
• Marketplace source: ./plugins/proofpunk resolves to an existing directory with a matching plugin manifest.

2. Skill Frontmatter Validation — PASS (10/10)

Every skills/<name>/SKILL.md has YAML frontmatter with name matching its directory and a description. Line counts (all within progressive-disclosure budgets):

Skill	Lines
brainstorm	153
cook	145
end-user-testing	172
full-functional-audit	154
functional-validation	239
plan-hardening	145
prompt-forge	171
ui-experience-audit	194
validation-plan	143
visual-inspection	144

(Line counts reflect the final state after the skills-ref.zip incorporation; all remain within progressive-disclosure budgets, with heavier detail pushed into per-skill references/.)

3. Internal Reference Integrity — PASS (99/99)

Every references/, assets/, and scripts/ path cited from any plugin markdown file was resolved against the filesystem: 99 citations checked, 0 broken. This includes the cross-skill relative paths used inside skills — e.g. skills/functional-validation/SKILL.md reaches the helper at ../skills/end-user-testing/scripts/fresh_evidence.py (expressed there as a dot-dot path) and every skill reaches the shared contract at ../references/evidence-contract.md (expressed there as ../../).

4. Placeholder / Dependency Scan — PASS

• No placeholder tokens ([TODO], [INSERT ...], [PLACEHOLDER], {???}, FIXME) in any shipped file. Four scanner hits were manually reviewed and confirmed to be the placeholder-DETECTION rules themselves (the defect-database row describing lorem-ipsum detection, prompt-forge's no-placeholders quality gate, and this document's own description of the scan) — legitimate content.
• The bundled script imports only Python stdlib (re, sys, pathlib, datetime, __future__). No external or unsupported dependencies anywhere in the plugin.
• No macOS metadata (._*, .DS_Store) and no __pycache__ in the tree.
• python3 -m py_compile on fresh_evidence.py: OK.

5. Functional Tests — fresh_evidence.py (12/12 PASS)

Real invocations against a scratch directory (/tmp/fe-test):

#	Test	Expected	Actual
1	init-run login-flow	exit 0, prints run-<ISO>-login-flow, creates dir + .run-meta + empty inventory	PASS
2	init-run "bad slug!"	exit 2, slug must be kebab-case alnum/dash/underscore	PASS
3	next-step home-loaded	exit 0, prints .../step-01-home-loaded	PASS
4	capture two artifacts (>1KB each)	files land at printed prefixes; step numbers sequential	PASS
5	seal	writes inventory with per-file byte counts + sealed=... count=2 total_bytes=3041	PASS
6	validate (clean run)	exit 0, validate OK: e2e-evidence/run-...	PASS
7	validate with a zero-byte artifact	exit 2, EMPTY: ... (zero bytes) + validate FAIL: 1 issues	PASS
8	validate with a 2020-mtime artifact	exit 2, STALE: ... (mtime 1577808000 < run-start ...)	PASS
9	next-step/seal/validate with no run	all exit 2, no active run	PASS
10	unknown command frobnicate	exit 2 + usage	PASS
11	init-run with no slug	exit 2, init-run requires a slug	PASS
12	validate after removing bad artifacts	exit 0 (recovery path works)	PASS

Malformed and incomplete inputs fail clearly and safely (exit 2 + a specific stderr message); no tracebacks, no partial writes on refusal paths.

6. Capability Smoke Review — PASS

Each of the 10 skills was re-read after authoring to verify its incorporated source capabilities are present and internally consistent (gate patterns match evidence-contract.md, severities match severity-model.md, platform traps match platform-routing.md, end-user mandate cited where actions occur). See consolidation-decisions.md for the full source-to-skill map.

7. Blocked Item (environment limitation)

claude plugin validate . and a live /plugin install smoke test could not be run: the claude CLI is not available in this environment. Mitigation: the structural checks in sections 1–3 implement the documented manifest and layout contract from the archive's own skill-creator and working-with-cc plugin references (required fields, kebab-case names, reserved-name check, skills/<name>/SKILL.md layout, marketplace source resolution). Installing from the packaged archive (README section "Installation") performs the final in-product verification.

8. End-User Validation Policy Enforcement Audit — PASS

Full-plugin audit and rewrite pass enforcing the policy: validation is never faked, skipped, stubbed, or assumed complete — the AI actually executes MCP/automation tools and acts as the end user, for every skill/feature, or reports the outcome UNVERIFIED.

Changes applied and re-verified:

Location	Change
../references/evidence-contract.md	Added "The Non-Negotiable Validation Rule" after the Iron Rule; added two refusal rules (no completing validation without the AI invoking tools as the end user; no skipping/faking QA steps under any circumstance)
../references/end-user-actor.md	Non-negotiable statement at the head of The Rule
../references/severity-model.md	Added UNVERIFIED verdict; PASS for behavior now requires executed, tool-driven, end-user checks
All 10 skills	Two anti-pattern rows added to every anti-pattern table (verified placed inside tables with matching column counts)
functional-validation, end-user-testing, full-functional-audit	Verdict templates now record "Driven by: AI as end user via <tools> — <actions performed>" and offer UNVERIFIED for unexecuted criteria
validation-plan	VALIDATION.md handoff requires an explicit record of executed end-user actions
plan-hardening	Consensus checklist requires every behavioral criterion to pair with a driven end-user action; gate block verdict adds UNVERIFIED
cook	End-user verification gate: never skipped/faked; no tool path -> report says UNVERIFIED with reason; finalize cites driven evidence per criterion
brainstorm	Success metrics/validation criteria must be phrased as end-user-executable checks
prompt-forge	Quality gate 8: end-user validation clause for any system-touching prompt
visual-inspection	PASS certifies visual quality only; behavioral claims route to driven validation
ui-experience-audit + report template	UNVERIFIED verdict added; template header records tools + actions actually executed
README.md	Operating Principle 2 restated with the non-negotiable language

Post-change verification: structural validator 0 errors (77 reference paths checked — all new end-user-actor.md citations resolve); every anti-pattern table pipe-count verified; policy-language coverage confirmed in 10/10 skills with 9/10 linking the shared mandate directly and brainstorm linking it from its report criteria; fresh_evidence.py recompiles clean (the script itself was not modified in this pass).

9. Second-Source Pass — skills-ref.zip Incorporation — PASS

Date of runs: 2026-08-09. Source: skills-ref.zip, SHA-256 fe1209f9bdf10fc2249552dbcd8b4b48b9f59a69aeef9fe0db7d60bc5e009d7d (26,223,521 bytes; 549 top-level entries, 334 with SKILL.md). Full source-to-destination map in consolidation-decisions.md ("Second Source Set").

Changes applied in this pass:

Location	Change
../skills/functional-validation/references/{ios,web,api,cli}-validation.md	4 platform runbooks carried verbatim from the ref functional-validation (85/88/75/120 lines)
../skills/functional-validation/SKILL.md	Step 1 runbook-loading table; Failure Diagnosis Table; Multi-Platform Order; Evidence Quality Standards; Mock Detection Red Flags; NEVER list (158 -> 239 lines)
../references/preflight-checks.md	NEW shared reference from ref preflight (61 lines); linked from end-user-testing Step 3 and platform-routing.md
../references/ci-gates.md	NEW shared reference from ref build-quality-gates (34 lines); linked from end-user-testing Related Skills
../skills/end-user-testing/assets/verdict-template.md	NEW asset adapted from ref e2e-validate/templates/verdict.md with Driven-by / Actions-executed / UNVERIFIED (40 lines)
../skills/validation-plan/references/task-file-format.md	NEW from ref code-task-generator with End-User Actor policy applied (60 lines); linked from validation-plan Step 3
../skills/full-functional-audit/SKILL.md	Per-platform team templates (iOS / Web-Full / API-only) with EXCLUSIVE/PARALLEL mutex annotations (134 -> 154 lines)
../skills/prompt-forge/SKILL.md	Input-type handling table, capability-inventory injection, model tone guidance (136 -> 171 lines)

Post-change verification (all real runs against the packaged source):

• Structural validator: 0 errors, 99/99 reference paths resolve (up from 77 — every new citation from end-user-testing, validation-plan, platform-routing.md, and these docs to the new references/assets resolves). 4 warnings, all manually confirmed as the placeholder-DETECTION rules themselves.
• New-file spot checks: all 8 new carried/authored files non-trivial (1,204–6,680 bytes); policy language (UNVERIFIED, end-user/driven-action clauses) confirmed present in both new policy-bearing files (task-file-format.md, verdict-template.md).
• fresh_evidence.py functional re-run: 12/12 PASS against a fresh scratch directory (init-run, bad-slug refusal, next-step sequencing, artifact capture, seal with count=2 total_bytes=3548 in the inventory, clean validate, empty-artifact refusal, stale-artifact refusal, no-run refusals, unknown-command refusal, missing-slug refusal, recovery validate). The script was not modified in this pass; the re-run confirms no collateral damage.
• Exclusions honored: testing-anti-patterns / tdd-workflow / testing-strategy (mock-first TDD conflicts with the Iron Rule), website-bug-hunt (duplicates bug-hunnter + full-functional-audit), worktree-merge-validate (Auto-Claude ecosystem dependency), and the empty ak-brainstorm / ak-cook stubs were NOT incorporated — verified absent from the plugin tree.

10. Official skill-creator Format + Packaging Pass — PASS (10/10)

Date of run: 2026-08-09. Tooling: the environment's own skill-creator skill (its quick_validate.py and package_skill.py scripts, external to this plugin).

• First run found one real violation: validation-plan's description used -> arrows; the official contract forbids </> in descriptions (max 1024 chars, kebab-case name ≤64, frontmatter keys restricted to name/description/license/allowed-tools/metadata/compatibility). Fixed (-> → →) and re-validated.
• Final: 10/10 skills "Skill is valid!" under the official validator.
• Packaging: 10/10 .skill files produced by package_skill.py (zip format, <skill-name>/SKILL.md at archive root, integrity-checked). Distribution note: the 10 skills share plugin-root references/ by design (../../references/... links), so the installable unit for the set is the marketplace layout (proofpunk-marketplace.tar.gz) per the plugin marketplace schema; the .skill files prove per-skill format adherence and package correctly, but standalone installs lack the shared references.
• End-to-end dogfood: the full 10-skill series was executed against a real repository (pallets/flask tutorial app, fetched and byte-verified via the GitHub MCP cloud plugin) implementing a real feature ("Mood Ring"): brainstorm → prompt-forge (91/100 rated prompt) → validation-plan → plan-hardening (5 findings dispositioned) → cook (32/32 pytest green) → functional-validation + end-user-testing (19-file sealed evidence run, validate OK) → visual-inspection (found + fixed a real HIGH contrast defect) → ui-experience-audit → full-functional-audit (PASS). Walkthrough report delivered alongside this document.

11. Mission-Fit Expansion Pass (v1.1.0) — PASS (15/15)

Triggered by a post-delivery usefulness audit: the original scan had bucketed 407/445 archive skills as "other" without classification. Full-universe re-scan (664 unique skills, 15 usefulness domains) identified 26 readable, mission-aligned sources in skills-ref.zip (re-downloaded; SHA-256 matched the first pass) and consolidated them into 5 new skills: stack-testing, mobile-validation-runner, root-cause-debugging, production-readiness, red-team-eval.

Results:

• Format: all 15 skills pass the official skill-creator frontmatter validation (one YAML colon-in-flow-scalar error found in root-cause-debugging's description during authoring; fixed with a block scalar before delivery).
• Reference integrity: 149 path citations across the expanded plugin resolved against the filesystem; 10 broken citations found inside incorporated reference bodies (paths relative to their original skill layouts) — all re-pointed to the bundle layout; final count 0 broken.
• Script integrity: bundled Python (with_server.py) compiles; all 6 bundled shell scripts pass bash -n; the Playwright runner parses under node --check.
• Policy: mock-framework sources incorporated only with explicit Iron Rule adaptation notes; mock-first doctrine skills remain excluded (see consolidation-decisions.md, Third Source Pass).
• Honesty gap preserved: Android and accessibility-audit sources exist only in the now-inaccessible (403) skills.zip; the mobile skill documents the Android gap explicitly instead of improvising content.

12. Session-First Reorientation (v1.2.0) — PASS (16/16)

User correction: the plugin must look at the sessions themselves — transcripts carry prompt intent; summaries are claims. Added session-intent with a real stdlib parser (session_intent.py) for ~/.claude/projects/**/*.jsonl.

• Parser verification: end-to-end run against a constructed fixture in the real JSONL event shape — 2 sessions parsed; tool_result events excluded from prompts; intent/steering/commit/file extraction asserted correct. Fixture is constructed data, honestly labeled; first live-corpus run remains the real gate (stated in the skill itself).
• Format: 16/16 skills pass official skill-creator validation.
• References: incorporated claude-code-analyzer body + 4 scripts (bash -n clean); all path citations re-resolved, 0 broken.

13. Doctrine Clarification — Test Runners Are Never Validation (v1.2.1, 2026-08-10)

User correction: the plugin must state explicitly that pytest (and any test runner) is never the validation mechanism — validation is the live system driven as the end user; for a JSON/HTTP Flask backend that means real curl requests to the running server on localhost with assertions on real responses.

Changes:

• references/end-user-actor.md: new "Test Runners Are Never Validation" section + actor/spectator table row. Test-runner output may appear only as REGRESSION evidence, clearly labeled. Framework test clients (Flask test_client) are regression gates, not end users — they bypass the network stack. When a contract asks for "tests pass" as proof of done, the correct reading is: tests green (regression) AND live end-user drive (validation); either alone is incomplete.
• functional-validation NEVER list: explicit ban on citing test-runner output as validation evidence.
• stack-testing rule 5: test output is sealed as REGRESSION evidence only; feature verdicts come solely from live end-user driving.
• examples/mood-ring/.prompts/build-mood-ring/PROMPT.md: output contract amended — the confusing "pytest output (must exit 0)" line is now "REGRESSION gate, labeled, never a validation verdict" plus explicit curl validation requirements; <validation> section names curl-to-localhost as the mechanism for HTTP backends.

Compliance of the existing demo run under the amended contract: the sealed run already validates via curl captures of the live server (steps 01, 03–09, 12, 14, 15) and browser drives (steps 10, 11, 13, 18, 19); step-17 (32 passed) stands as the REGRESSION rail. No re-run required; the prompt amendment aligns the contract with what was actually executed.

14. Installer Script (v1.3.0, 2026-08-10) — PASS (8/8 live tests)

tools/proofpunk-install.sh installs the 16 skills as plain skills (not a plugin) into a chosen target (claude-code / omp / --dir), with collision safety, doctrine injection, and post-install verification. Executed live before shipping (no claimed-but-unrun behavior):

1. --target omp --dry-run — full plan printed, zero writes.
2. GitHub-source install into a sandboxed HOME — 16/16 installed + doctrine bundle + verify green. The verifier caught a real design flaw: repo layout cites doctrine as ../../references/, broken in a flat skills dir; fixed by making installed copies self-contained (citations rewritten, cited shared refs bundled inside each skill) — then verify green.
3. Re-run without --override — 16/16 SKIP on collision, nothing clobbered.
4. --override — replaced with timestamped backups (.bak-YYYYMMDD-HHMMSS).
5. --inject-claude-md twice — rules block appended once; second run "already present — left unchanged" (idempotent).
6. --source-dir local checkout — installs offline path, verify green.
7. --only not-a-skill — exits 3 (caught a second real bug: the EXIT trap's cleanup returned 1 and clobbered the intended code; fixed, re-verified).
8. bash -n syntax clean; exit codes 0/1/2/3 as documented in tools/INSTALL.md.

15. prompt-forge Rework (v1.4.0, 2026-08-11) — PASS

Directive: rating remediations must be applied immediately and always land in a file (new file by default, the input file only with consent); every mode must run sequential thinking, proper XML tags, an authorization engine, and explicit todo tracking.

Changes to skills/prompt-forge/SKILL.md (quick_validate: PASS):

1. New Section 0 — Always-On Workflow, binding on all four modes:
   • 0.1 Sequential thinking — numbered reasoning steps with explicit revision and branching before any authoring/rating/optimizing.
   • 0.2 Todo discipline — one action per todo, one in-progress at a time, immediate completion, pending items finished or reported.
   • 0.3 Authorization engine — decision table: new-file writes need no consent (default action); in-place edits, below-threshold shipping, overwrites, and report-only ratings all require explicit consent.
   • 0.4 File-output contract — per-mode file matrix (NAME.prompt.md / NAME.rating.md + NAME.remediated.md / NAME.optimized.md / .prompts/ stages); chat-only suggestions are a contract violation.
   • 0.5 XML tag standard — canonical skeleton (task, context, skills_to_activate, mcp_tools, sequential_thinking, todos, authorization, constraints, output_contract, validation, example) with a required/optional purpose table; aligned with the mood-ring demo prompt's existing tag usage.
2. RATE §2.1 Remediation application (mandatory) — scorecard alone is unfinished: apply top fixes, write NAME.remediated.md (in-place only with consent), re-score against the same test cases, report before/after; below-threshold results route through 0.3.
3. OPTIMIZE — optimized prompt written to NAME.optimized.md per 0.4; failure classification lands in the sequential-thinking chain.
4. PIPELINE — every stage prompt follows the 0.5 skeleton; every stage maps to a todo.
5. Anti-patterns — six new rows: scorecard-only rating, chat-only fixes, unconsented in-place edits, below-threshold shipping, jump-to-output, ad-hoc XML tags.

Source-gap note (honest gap): the sequential-thinking stage was specified from the established methodology (numbered steps, revision, branching — the ck:sequential-thinking description fragment in the archive inventory). The reference archive's ak-sequential-thinking copy contains zero files and its sequential-thinking copy has no SKILL.md; the primary archive's ck:sequential-thinking was unreachable this session (HTTP 403). No content was copied from a source that could not be read.

Frontmatter description rewritten (1011 chars, under the 1024 limit) to advertise the always-on workflow and file-output remediation contract.

16. prompt-forge Command Surface + implement Skill (v1.5.0, 2026-08-11) — PASS

Two directives: (a) the v1.4.0 prompt-forge rework described remediation abstractly but showed no concrete remediation sample and no CLI arguments; (b) a new implement command should orchestrate the actual implementation process using ALL skills, with session mining and parallel/auto modes.

prompt-forge additions (quick_validate: PASS)

1. Command Surface section — CLI-style flags mapped onto the authorization engine (0.3), so a flag IS the recorded consent: --in-place (in-place edit consent), --report-only (skip file output), --ship-below-threshold (below-threshold sign-off), --out, --depth, --evidence (OPTIMIZE's hard requirement), --dir. Unknown flags rejected; conflicting flags fail fast.
2. references/remediation-sample.md — a full worked RATE pass on a weak prompt: sequential-thinking chain, scorecard (34/100, rewrite), predicted failure modes, four test cases, the remediated file on the canonical XML skeleton, re-score (91/100), and what each flag would have changed.

implement skill (quick_validate: PASS)

New skill skills/implement/SKILL.md — the orchestrator (conductor) to cook's execution engine (player), resolving the overlap question explicitly in both directions:

• Command surface (as of that release; --no-test and --tdd were later removed, and cook was merged into implement — see improvements.md): implement "<goal>" [--parallel] [--auto] [--mine] [--fast] [--no-test] [--tdd] and implement mine [--project] [--since] [--until] [--json], each flag with a why-it-exists column.
• Phase 0 — TRUE success criteria distillation: observable, end-user provable, measurable; approval gate when the goal is not clearly laid out or not understood — the one mandatory stop under --auto.
• Phase 1 — MINE via session-intent: past implementation sessions as an intent matrix (prompts, tools, files, commits) feeding exploration and forging; implement mine alone is a reconnaissance-only pass.
• Phase 2 — EXPLORE with parallel scout agents (structure, patterns, contracts, history); contradictions resolved against the code.
• Phase 3 — FORGE via prompt-forge AUTHOR on the canonical XML skeleton; approved criteria become the prompt's success metrics verbatim.
• Phase 4 — PLAN: validation-plan + plan-hardening; --parallel -> prompt-forge PIPELINE with parallel independent stages.
• Phase 5 — EXECUTE under cook's gates; --parallel lanes with per-lane todo chains; authorization boundaries (destructive ops, out-of-scope edits, below-threshold shipping) bind even --auto.
• Phases 6-8: root-cause-debugging (no retries as fixes), end-user validation (UNVERIFIED = NOT DONE under --auto), criteria-by-criteria proof table + todo ledger.

Propagation

plugin.json 1.5.0 (description + 6 keywords), marketplace.json 1.5.0, README skill table + counts (16 -> 17), installer ALL_SKILLS + INSTALL.md counts/examples updated. Both new/changed skills packaged self-contained (shared refs bundled, citations rewritten).

17. README Command Reference + Architecture Diagrams (v1.5.1, 2026-08-11) — PASS

Directive: the README must lay out every example with every argument — including positional arguments inside skill commands — every meaningful permutation with why / what happens / what you end up with, which skills execute per invocation, and system architecture diagrams.

Delivered in README.md (629 lines, +550):

Historical record. This section describes the surface as shipped in that release. cook and functional-validation were later merged into implement, and the --tdd / --no-test flags were removed. For the current surface see usage-guide.md and improvements.md.

1. Dispatch-at-a-glance table — invocation -> skill fired -> delegation.
2. §1-8 command references for every argumented surface: implement (positionals <goal>/mine, 8 flags, 14 permutations), prompt-forge (positionals per subcommand, 7 flags, 9 permutations, conflict rules), cook (positionals <goal>/<plan-path>, 5 modes, full mode × --tdd 12-permutation table), functional-validation (8 flags composing analyze->plan->execute->fix->report), session-intent (5 args with defaults, filter/output permutations, exit 2), end-user-testing (positional <slug>, lifecycle as the only valid composition, exit codes), stack-testing (--server/--port/trailing positional), installer (13 flags + canonical permutations + conflicts + exit codes). Every row answers why / what happens / what you end up with, and names the skills executed downstream.
3. §9 protocol-skill routing — the 10 flag-less skills: invocation -> what happens -> downstream skills, all edges from the skills' own Related Skills sections (verified by extraction, not written from memory).
4. Five mermaid architecture diagrams — repo layout, layered skill stack with the doctrine foundation, the full delegation graph (real edges only), a sequence trace of implement "add billing webhooks" --parallel --auto --mine, and the evidence lifecycle state diagram.

Mechanical checks: code fences balanced (28/28), 5/5 mermaid blocks closed, all 17 skill names referenced.

18. Related-Skills Closure + Usage Guide (v1.6.0, 2026-08-11) — PASS

Directive: document the plugin and skills properly with /skill-name example --arg usage in Claude Code; ensure every skill needed by the Related Skills graph actually ships; update the installer and docs.

Integrity audit (mechanical, full closure)

1. Related Skills edges: extracted every backtick-quoted reference in all Related Skills sections and resolved each against the skills directory. Found ONE dangling edge: session-intent referenced codebase-truth-audit, which was not in the plugin. Fixed by shipping it: the skill (SKILL.md + references/output-contract.md + scripts/init_audit_workspace.py) integrated as skill #18, with a new Related Skills section closing the edge both directions. Re-run of the audit: 18/18 skills, zero dangling edges.
2. Reference citations: every references/X citation in every skill file (SKILL.md, references, scripts) resolved against the tree — including shared ../../references/ doctrine paths and cross-skill ../skill/scripts/ paths. Zero missing targets.
3. Bundled scripts: every scripts/X citation resolved. init_audit_workspace.py verified LIVE (not by inspection): ran --repo /tmp/repo --label Proofpunk-v151 end to end — produced a real 8-phase audit workspace with captured git evidence (status, remotes, metadata). Exit 0, artifacts non-empty.

Usage guide

New plugins/proofpunk/docs/usage-guide.md: how invocation works in Claude Code (slash command vs natural-language routing vs chaining), then per-skill sections for all 18 skills with positional arguments, flags, and literal /skill-name <args> example invocations, the three bundled CLIs (session_intent.py, fresh_evidence.py, init_audit_workspace.py, with_server.py) with their real argument lists, and seven chaining recipes. INSTALL.md gained an "After install" usage section; README links the guide and shows 18 skills.

Propagation

plugin.json 1.6.0 (description + 3 keywords), marketplace.json 1.6.0, README table + counts (17 -> 18), installer ALL_SKILLS + INSTALL.md counts. Both quick_validates PASS (codebase-truth-audit, session-intent).