Skip to content

Sources and implementation evidence

This ledger records executed work. The feasibility source ledger documents the earlier research and methodology inspection.

Initial executable-slice evidence

The subsequent implementation uses Playwright 1.64.1, CO2.js 0.19.0, Harbor 0.23.0, Codex CLI 0.166.0, and the npm-distributed OpenCode 1.19.5. The local opencode --version reported 2.2.5, but opencode-ai@2.2.5 was unavailable in the registry used by Harbor. Record those as distinct installations rather than assuming interchangeable version numbers. Package versions were checked against the npm registry; runtime dependencies are recorded in package-lock.json and the benchmark configurations.

Evidence Source and interpretation
Controlled observations benchmarks/corpus.json; actual cold/warm captures of original project-owned fixtures, with browser/protocol metadata and hashes
Live targets Guide homepage, image-management instructions, Wagtail homepage, sustainability page; sanitized observations
Provider preflight Sanitized DeepSeek tool-call response; one authenticated request, including provider-attributed energy and cost
Harbor integration Separate verifier documentation, environment-variable boundaries, and inspection of the installed 0.23.0 implementation
Neuralwatt routing OpenCode integration; actual model requests through its OpenAI-compatible adapter
OpenCode failure investigation Harbor issue 4019 describes a matching completed-output/nonzero-exit symptom. This similarity does not establish the underlying cause in our environment.

The source-assisted fixture uses a runner-supplied excerpt and deployment identity. It is a controlled context test, not independent verification of a production deployment. The incomplete case removes observations deliberately; it is distinct from the collector's actual HTTP-error test. The checkpoint distinguishes these development cases from broader methodology and accuracy claims.

The executable slice passed just lint, just build-docs, just test (11 Python tests and 7 JavaScript tests), just check-links, and just check-external-links. The skill metadata validator also passed. Six oracle controls passed, six no-operation controls failed their grader as intended, and both required model arms produced reports for all six revised cases. OpenCode's six execution errors and Harbor's historical Codex redaction corruption are recorded separately from grader acceptance. Published restored artifacts were regraded against their original references, with matching rewards; identical-input hashes were checked across model arms.

Model switch and real-site review

The current checkpoint records Codex CLI routing to Neuralwatt's Responses endpoint with DeepSeek/high, and the Sol/low baseline. The earlier audit draft linked runtime observations, source revisions, arithmetic checks and a guideline coverage register. That draft has been removed; the original evidence remains in the experiment artifacts. Source copies remain read-only.

The source inspection covers tracked public files from Wagtail Guide and Wagtail.org; production revision identity remains unknown. Additional browser probes distinguish initial load from scrolling, JavaScript-disabled behavior, successful journeys, and opt-out state. A separate font browser probe failed with ERR_NETWORK_CHANGED; an HTTP retrieval of the exact stylesheet observed in the successful browser capture corroborated the font rules. The font artifact records this fallback.

The documentation style guide now explicitly relies on git history for editorial dates. Removed datelines were authoring choices, not a requirement of the workspace skill. Source publication dates and recorded measurement timestamps remain evidence.

The revised campaigns completed six fixture trials per model and two real-site review trials per model, without harness errors. The real-site grader corrections and original scores are documented in the checkpoint results; corrected oracle/no-operation controls and new grader regression tests validate the revised contract. They do not independently validate recommendation quality.

Checkpoint verification passed just lint, just test (11 Python and 9 JavaScript tests), just build-docs, just check-links, and just check-external-links (zero errors). Paired model inputs and instruction hashes match within each task. These checks validate the implementation and report provenance, not whole-site sustainability conformance.

Staged collection and contextual discovery

The staged workflow implements owner feedback in Python with uv, retaining the existing CO2.js calculation adapter. Six fresh-context runs cover four pages, including separate accept/reject trajectories on the two Wagtail.org pages. The published artifact preserves initial, post-choice, and post-scroll observations, accounting limits, disclosure retrievals, and technology detections. All six runs reached the bottom. Initial observations remain the headline figures; interaction deltas are not automatically avoidable savings.

Source Use and limitation
PyPI Wappalyzer Pinned 2.0.4 full-browser scan, executed separately from measurement. The artifact records package version, bundled extension digest, raw detections, and architecture-specific review hints. Fingerprints do not prove successful execution or exact deployment versions.
carbon.txt discovery, syntax, and setup DNS/header pointers, root and well-known paths, TOML parsing, and linked disclosures. Parsing a declared version is distinct from schema validation or verification of its claims.
Wagtail carbon.txt Retrieved successfully; declares version 0.4, infrastructure providers, and disclosure URLs. Guide carbon.txt locations returned 404; related project documentation was found using an explicit hint.
Wagtail digital-emissions statement Identifies a site-specific estimate, mobile measurement profile, and carbon assumptions. Keep this attributed statement separate from the sampled desktop measurements and the ecosystem-wide arithmetic finding.
uv script guide and required-version setting Executable script, inline Python/dependency/uv requirements, and committed script lockfile.

Regression tests exercise partial-byte replacement, redirects/cache, failed resources versus incomplete accounting, persistent coverage limits, publication sanitization, private checkpoints, bounded/malformed disclosure retrieval, technology guidance, action failures, and a real browser interaction. A Python specialist reviewed the collector; review findings informed fixes to URL sanitization, action errors, evidence preservation, frame coverage, excerpts, and total request deadlines.

The shared evidence-review skill at this stage was version 0.2.0. Historical model outputs remain tied to their original evidence and instructions. Model-bearing evaluation of the staged evidence followed in the expanded campaign below.

Verification passed just lint, just test (25 Python and 9 JavaScript tests), just build-docs, just check-links, and just check-external-links (zero errors). The script lockfile check and skill metadata validation also passed. These checks establish implementation consistency; they do not independently validate the measurement protocol or recommendation quality.

Per-site reporting, energy, and traffic

The reporting contract adapts the inspected claude-seo audit structure into separate site reports with key measures, top priorities, quantitative evidence, and qualitative findings. Review skill 0.3.0 requires those boundaries. JSON/CSV exports retain raw units, statuses, phase definitions, scenario bases, and exact source digests; human reports use decimal KB/MB. No aggregate WSG score is invented.

The energy experiment ran three valid homepage repetitions per site through the co2-runner source. Its commit and dirty-file hashes identify the actual implementation, including the CPU-package counter selection. Linux RAPL observations cover whole-system package energy during Firefox startup, load, wait, and shutdown. The methodology note cites Firefox documentation and sitespeed.io/Browsertime source. Sitespeed.io's hardware path was researched but not run; its CO2.js sustainability score is a distinct model.

The CrUX adapter retrieved September 2026 monthly rank buckets from the public mirror pinned at 5493dd425c47fc7895d7d92df688d70638fb2aa7. Both exact HTTPS origins matched. The supplied ranking-to-pageviews mapping is pinned at e36613ef9dd2f7cee5557ea794d238af31acbaf7; it models a different population and period. Published estimates remain traffic scenarios, with no annual emissions product. Raw compressed-file and mapping digests appear in the traffic artifact.

Expanded model comparison

The 12-case campaign uses Harbor 0.23.0, Codex CLI 0.166.0, skill 0.3.0, and separate Python verifiers. Both configured models receive matching input, instruction, skill, reference, and grader hashes. Revised oracle controls accept 12/12 and no-operation controls accept 0/12, with no control exceptions. References are coordinator-authored public development labels, not held-out owner judgments.

Sol/low accepts 10/12; Neuralwatt DeepSeek V4.1 Flash/high accepts 10/12 completed cases. DeepSeek required one unchanged-input retry after Codex's platform binary was missing during setup. The original exception remains published. Three rejected reports reflect fail/inconclusive disagreements with correct uncertainty in their prose and metrics; the remaining DeepSeek rejection includes an unsupported traffic scenario after a limited lookup. All original scores are preserved. The published campaign artifact includes reports, diagnostic regrades, available usage, and error types. Inference energy and billing remain unreconciled.

Broader WSG and STAR sample

The scope extension applies the WSG draft and STAR methodology to a larger, explicitly limited development sample. Separate registers preserve the inspected October 2026 WSG version and account for all 71 guidelines without conformance passes. Thirteen content pages and two missing paths produced 32 DOM snapshots. The complete advertised sitemap inventories contain 156 and 525 URLs; these are not complete inventories of states or user tasks.

Read-only source inspection adds exact file hashes, repository revisions, dirty status, and candidate fix locations. The guide language fallback and both sites' initial theme selection are examples where source clarifies runtime evidence. Source deployment identity is still unknown. The policy review distinguishes publication from actual processing and identifies unclear site scope in Wagtail.org's privacy statement.

Python review identified and corrected sitemap-extension URL contamination, malformed-pointer failures, missing total browser deadlines, failed recovery probes counted as evidence, and policy misclassification after redirects. Regression tests cover those boundaries. The final published collection was rerun with the corrected collector; initial private observations remain separate. The backlog preserves the next validation gates and deferred routing, calibration, scoring, and presentation work.

Final verification passed just lint, just test (53 Python and 9 JavaScript tests), just build-docs, just check-links, and just check-external-links (zero errors). All eight new script lockfiles and the review skill's metadata validate. The published scope artifact matches its collector digest, and both per-site registers match the exact formatted scope artifact. These checks establish reproducibility and implementation consistency, not independent audit accuracy.

Scoped scoring and skill-guided review

The scoring research uses ACT Rules Format 1.1 for outcome-mapping distinctions, WSG Impact Ratings for guideline priority points, and STAR for evaluation/reporting context. None is presented as prescribing our website score. ACT's inspected version is the February 2026 Recommendation; WSG and its impact ratings use the October 2026 drafts. All 71 impact-point values were extracted by matching WSG fragment identifiers to the corresponding rating tables; audit/impact-weights.json retains the source version and fetched HTML digest.

The project-owned scoped-wsg-v1 rubric supports 21 bounded guideline groups. Two independent GPT-6.1 Sol/low reviewers applied evidence-review 0.4.0, guideline-review 0.1.0 and claims-review 0.1.0 to the raw site artifacts and supplied source. They were instructed not to read the prior human report conclusions. The coordinator inspected their judgments, including the new eager-logo finding, and retained their outcomes. This tests actual skill execution; it does not isolate a causal skill benefit or provide owner-validated reference labels.

The review artifact includes original judgments, evidence locations, source/rubric/skill hashes and raw-review digests. The deterministic publisher resolves citations to existing JSON nodes and checks their enclosing site; explicit repository associations permit supplied source corroboration. Historical ecosystem arithmetic is associated with Wagtail.org's disclosure; a related disclosure found during guide discovery remains attributed context. Unverified source deployment is not promoted to attestation.

The score artifact preserves all 71 guideline rows per site. Unknowns stay in the attempted-check denominator without becoming failures. Repeated observations do not multiply a guideline's weight. Scores reject unresolved evidence references, missing indexed nodes, changed source artifacts and a changed rubric digest. Source semantics still require review; a resolvable citation alone does not establish that it supports the judgment.

Annual estimate extension

The owner requested annual estimates despite the unknown real page mix. The annual calculator now uses a declared equal mix of the two measured pages per site, averaging repeated branches before assigning page weights. It retains the earlier first-visit-only SWDM v4 assumptions and calls pinned CO2.js 0.19.0 without requesting ratings. The original CrUX/mapping artifacts remain unchanged; the new annual artifact references their exact bytes.

The resulting 5.3 and 65.1 kgCO2e/year estimates are browsing-footprint scenarios, with Wagtail.org's partial accounting visible. Sampled-page sensitivity values are not confidence bounds. These estimates do not alter DCR eligibility, add measured CPU-package energy, or establish a complete organization/site operational inventory. Input, model dependency and calculator hashes are published for reproduction.

Python specialist review identified and fixed fabricated/unresolved citation acceptance, cross-site evidence leakage, stale-rubric reuse and finite-input numeric overflow. Regression tests exercise those failures, score aggregation, unknown/zero-weight denominators, branch averaging and model agreement with the frozen guide trace. The methodology benchmark separately evaluates the interpretation contracts with both configured model arms; the older 12-case suite now pins its historical skill copy.

Verification passed just lint, just test (69 Python and 9 JavaScript tests), just build-docs, local links and external sources (zero errors). The three new executable script lockfiles, four updated script locks, and all four audit-skill metadata files validate. Running just audit-assess regenerated the score, annual scenario, and per-site JSON/CSV artifacts with identical hashes. This confirms deterministic reproduction from the published reviews; it does not make LLM judgments or future live measurements deterministic.

The completed methodology campaign accepted 10/10 Sol reports and 8/10 DeepSeek reports, with no harness exceptions or retries. Both matched all 19 claim labels and 12 numeric fields; DeepSeek's two rejections concern recommendation verbs while its explanations retain the intended publication/investigation advice. The results and interpretation preserve these original scores, identify the author-created action ambiguity, and avoid attributing improved performance to skills without an ablation. All paired task and skill-resource hashes match.

Agent-triggered workflow and expanded WSG interpretation

The repeatable workflow adds arbitrary-origin plans, owner questions/assumptions, explicit journey actions, collector command receipts, frozen skill/expectation packets and validated scoring. It uses the existing byte/DCR collector and new public-browser probes; it does not call an LLM secretly or require a proprietary orchestration API. Source roots remain declared context until a source-assisted evidence adapter records them.

The current WSG auditability classifications describe the evidence needed for all 197 criteria across 71 guidelines in the pinned October 2026 WSG publication. Each criterion has a primary class, supporting review methods, evidence access routes and rationale. The skill reference retains criterion text, source digests and the W3C Software and Document License. Classification is a planning aid, not a count of implemented tests or full-guideline coverage; use the check reference for implementation breadth.

Carbon.txt behavior was checked against its syntax and FAQ, separately from the 1.0-RC1 Internet-Draft. The collector records presence, discovery route and redirects; the interpreter retains disclosures, provider roles, declared periods and certification relationships. Schema validation and independent claim verification remain distinct follow-up work. The carbon evidence note documents how declarations now inform review.

The GreenFrame client adapter follows upstream commands, configuration and scenario documentation, targeting CLI 2.0.1. The bakerydemo-gold-benchmark example supplies useful task sequences but uses older exports and server-container configuration. Preparation/preflight succeeded; the CLI was unavailable, so no GreenFrame measurement was executed. Actual runner validation and image/browser pinning remain required before comparative use.

Two new live sessions collected four pages, six staged branches, twenty supplemental page states and two attempted journeys. The guide search-result click timed out; Wagtail's sustainability selector matched three links. Both model reviewers retained unknown task completion rather than treating visible headings as success or collector failures as site defects. Their exact evidence, reviews and scoped scores remain in the historical agent-session bundles. These Sol/low forward reviews test skill execution, not superiority over DeepSeek or another harness. A future Harbor workflow campaign must retain both required model arms.

Independent Python review found and corrected silently ignored page-level consent actions, missing frozen-scope/weight/skill verification and scorer-version drift. Regression tests cover those boundaries. The older methodology benchmark now uses preserved skill copies so this skill release does not alter its frozen task definitions or reported results.

Verification passed just lint, just test (94 Python and 9 JavaScript tests), just build-docs, local links and external links (930 checked, zero errors). All five new script locks and four audit-skill metadata files validate. Both published pilot scores reproduce byte-for-byte from their artifact bundles, and all publication-manifest digests match. The original methodology benchmark skill digests also match the preserved copies. These checks establish implementation consistency and reproducibility, not independent accuracy of qualitative judgments.

Broader checks and source-assisted reviews

The expanded catalog defines 33 scoped checks across 32 guideline groups. Ten additional browser checks add metadata, fonts/media, keyboard traces, response policy and cold-versus-live preference evidence. Eight optional source checks cover reuse, dependencies, runtime consistency, automation, maintenance, change documentation, budgets and open-source participation. The union of historical and current methods reaches 39/71 groups with partial review expectations; no complete guideline conformance is claimed.

The historical three-page campaign sampled three public pages, seven supplementary states per page, one completed journey per site and selected source evidence. Guide search now submits with Enter and opens the intended article; the Wagtail journey uses a unique footer link. URL, selector and expected-text postconditions prevent the earlier generic-heading false positive. Source collection reads only explicit files, rejects symlink aliases and non-regular files, and preserves unavailable or metadata-only records. It never executes project code or attests deployment.

Independent Codex CLI 0.166.0 reviews requested GPT-6.1 Sol with low reasoning in a read-only tool sandbox. They used frozen skills and packets without prior human reports. The guide score artifact records 15 passing, four failed and thirteen unknown guideline groups; the Wagtail score artifact records thirteen passing, two failed and seventeen unknown groups. These are scoped development judgments requiring owner feedback, not calibrated site rankings. Prompts, output schemas, model settings, evidence and exact hashes remain in the historical bundles. Their report drafts have been removed.

The Markdown renderer verifies the scored inputs, raw review, packet and source hashes before producing reports. Scores reproduce byte-for-byte from the published bundles. Chart JSON/CSV use only the new campaign; previous annual and hardware-energy results are not silently attached to it. Raw logs and copied skill directories remain private, with their digests retained in receipts.

The broader Harbor family freezes sixteen cases across 23 WSG groups. Independent semantic review strengthened three reference facts before freezing: associated useful form errors, actual cache reuse and separate consent contexts. All sixteen reference controls pass and all sixteen no-operation controls fail.

Both GPT-6.1 Sol/low and Neuralwatt DeepSeek V4.1 Flash/high accept 14/16 complete reports through Harbor 0.23.0 and Codex CLI 0.166.0. Only two Sol setup failures were retried, without changing inputs; all graded predictions and original exceptions remain in the published results. Both models preserve correct uncertainty and sensible recommendations in the same two rejected cases, exposing ambiguity about claim-validation labels and untested outcome counts. The case review publishes the facts, reference answers, model judgments and explanations for owner feedback without changing grades.

The publisher regrades artifacts and verifies matching task, skill and collector hashes across campaigns. Separate trajectory inspection corroborates the collector command and returned tool observation in all eight required model handoffs. This establishes a bounded agent-triggered collection experiment, not live-site orchestration, independent recommendation accuracy or a causal skill advantage. Inference energy and billing remain unreconciled.

Final verification passed just lint, just test (110 Python and nine JavaScript tests), just build-docs, local links and external links (1,168 checked, zero errors). The comparison renderer's uv lock validates, and its selection tests reject duplicate graded predictions and incomplete paired campaigns. These checks establish implementation consistency and reproducibility; the two disputed reference contracts still need independent review.

Original agent pilot reproduction

The coordinator executed the full collection workflow once for each known content site, using two pages and one explicit journey per site. Separate GPT-6.1 Sol/low reviewers applied the frozen skills without reading earlier human report conclusions. The original guide review and Wagtail review preserve those judgments after removal of the report drafts. Both preserve uncertain journey results instead of treating collector failures as site defects. These are live development reviews, not a controlled model-comparison campaign or owner-validated ground truth.

The published guide manifest and Wagtail manifest hash exact JSON artifacts. Formatting is intentionally excluded for these frozen files. Raw command logs and skill copies remain private; receipts retain their hashes. The scope and evidence were inspected before publication. Collection ran with uncommitted implementation changes, so the script hashes supplement the recorded base Git revision. Before review, packet construction was strengthened to pin scope, weights, receipt and skill files; this changed packet provenance, not site observations or review expectations.

Recalculate either published score without network or model calls (use a new output path):

uv run --locked --script audit/assess.py   --reviews docs/initial-implementation/assets/agent-sessions/guide/reviews.json   --rubric docs/initial-implementation/assets/agent-sessions/guide/rubric.json   --weights docs/initial-implementation/assets/agent-sessions/guide/impact-weights.json   --assets docs/initial-implementation/assets/agent-sessions/guide   --output .audit-work/guide-agent-recalculated.json

Replace guide with wagtail for the second site. Recollecting the live site or rerunning a reviewer can produce different evidence or judgments and requires a new session. The existing annual scenarios and headline scores retain their original input captures; no mixed-campaign total is calculated.

Readable reports and agent entry point

The sustainability-report skill separates editorial wording from frozen audit judgments. Two Codex CLI 0.166.0 passes used GPT-6.1 Sol with high reasoning to write the guide and Wagtail report explanations from the original review and relevant evidence. They did not rescore either site or run a new model comparison. The renderer verifies the review digest, complete check set and unchanged priority order. The old editorial receipts distinguished published skill-resource hashes from the prompt and raw output execution digests. Those presentation-only artifacts were removed with the earlier report drafts and remain in Git history.

The earlier guide and Wagtail illustrations were captured separately with the bounded screenshot collector and inspected visually. Their manifests recorded URL, viewport, color preference, timestamps, browser version and image hashes. They illustrated page appearance at their recorded capture times without attesting the earlier transfer measurements or changing audit outcomes. Those images and manifests were removed with the report drafts; Git history preserves the reporting experiment.

The orchestration research uses Wagtail CLI as the primary design precedent: one discoverable entry skill, progressively loaded instructions and a Python/Typer interface. The implemented plugin/CLI adds setup, collection and specialist handoffs around the existing file contracts. Manifest validation, temporary-project installation and deterministic handoff tests establish local consistency; a clean end-to-end audit under a second harness is still required before claiming portability. Historical low-reasoning campaigns keep their original configuration; new Sol and DeepSeek work uses high reasoning.

Validation passed with just lint, just test (124 Python and nine JavaScript tests), just build-docs, just check-links and just check-external-links. Browser checks covered both rendered reports at narrow and desktop widths, image loading, table semantics and keyboard scrolling. A regression test preserves access to the final table column on narrow screens. These checks validate the implementation and presentation, not the accuracy of the original audit judgments; no screen reader session was performed.

Fresh audits through the agent workflow

The replacement guide report samples six pages and one completed search-to-article journey; the Wagtail.org report samples seven pages and one completed search-to-sustainability journey. Setup selected 20 and 18 local source files respectively. Collection ran sequentially through the CLI, followed by independent quantitative, guideline and claims subagents, explicit reconciliation, deterministic scoring and separate report-writing subagents. All audit-review and report-writing roles used GPT-6.1 Sol with high reasoning. These are live development audits under Codex native delegation, not a controlled model comparison or a clean-install portability test.

Guide first-load accounting is complete within the recorded protocol on all six visits. Wagtail.org has 14 independent accept/reject visits and partial accounting throughout, including embedded-frame limits on the release article; its ratings remain withheld. Both reports retain their actual observation windows, source/deployment uncertainty and separate screenshot capture times. No historical energy measurements or annual scenarios were attached to the new measurements.

The guide's 22 advisory issues were reconciled without changing the guideline worker's outcomes. The main site's 30 advisory issues led to one outcome change: the demonstrated contradiction in its footprint average changes comprehension from unresolved to failed, while disclosure presence/scope remains a pass. The published handoffs preserve the original specialist results, exact acceptance digests and each resolution. Reporting changes wording and presentation, not those final outcomes.

Earlier report drafts and their presentation-only assets have been removed; historical measurement and benchmark inputs remain frozen. Each replacement report links its current inputs, chart data and a ZIP containing frozen skills and reproduction instructions. The cross-origin disclosure follow-up in B30 records an outstanding collection boundary: the parent policy names the guide, but it was not part of the guide's independently frozen discovery packet.

Validation passed: all 124 Python and nine JavaScript tests, lint/type/prose checks, strict documentation build, and local/external link checks. Both published scores were independently recalculated from their published inputs and matched exactly. Archive contents match the frozen skill/evidence digests. Keyboard and image/heading specialists checked both themed reports, including mobile table scrolling, evidence navigation and all four illustrations; this was a focused browser review, not a screen reader audit. A final renderer adjustment places key measures before the editorial assessment and distinguishes no-banner waiting checkpoints from cookie interactions.