> **This is a redacted sample.** A real audit of a real site,
> published with the owner's permission, with the organisation's identity
> removed. The site is referred to as **TLN**. Removed: the host name, the
> organisation name, four staff pages, the city, and a named certification.
> Ratings, observations and reasoning are exactly as signed. Content hashes
> are of the original unredacted pages and cannot be recomputed from this
> document.

You are applying an evidence-backed AI discoverability audit to a website.

Target:        https://tln.example
Audit:         aud_20260901T114719Z_77cb07cf
Rubric:        v0.2 (scored by adia/0.1.0)
Score:         57.1 of 78 adjusted rubric points, moderate confidence
Fit:           Material discoverability friction

Every item below came from a captured, timestamped observation of the live
site, not from a model's impression of it. The evidence ids in brackets refer
to records in the audit's JSON appendix.

## Rules

1. Work in order. The list is ranked by impact, effort, confidence, and
   dependency; an item that depends on another is listed after it.
2. Keep changes minimal and local. Do not redesign anything the audit did not
   flag, and preserve the site's existing voice, structure, and build setup.
3. NEVER INVENT A FACT. Several fixes below could be "completed" by making
   something up. Where an item says so, leave a clearly marked TODO for the
   site owner rather than filling in a plausible value. A fabricated
   testimonial, credential, statistic, or citation is worse than the gap it
   replaces, and this audit exists to remove unsupported claims, not add them.
4. Do not add any claim about ranking, citation, traffic, or revenue.
5. If a fix is ambiguous, say so and stop rather than guessing.

Working through every fix below is roughly **76 hours** of implementation for a competent team — a sum of per-item midpoints, not a quote, and it excludes review and deployment. The ranking exists so you can stop partway and still have taken the most valuable part.


## Fixes, in order

### 1. [CRITICAL] CR-06 Body text is machine-readable prose

**Observed.** 4 of 25 sampled page(s) print template syntax as body text, where a reader and a retrieval system see it: https://tln.example/home/ (90% of visible text, shortcode); https://tln.example/about-us/ (90% of visible text, shortcode); https://tln.example/blog-2/ (68% of visible text, shortcode). 1 further page(s) are affected.

**Why it matters.** A passage made of markup cannot be quoted, summarised, or cited. The page looks populated to the CMS and to a word count, and carries nothing an engine can use.

**Do this.** Render or remove the template syntax so the published text is the text, and re-check the pages that a buyer reads first.

*Evidence: ev_99d14a7a9a1356f6, ev_606797b76fc96939, ev_b445e84670f54e32, ev_d1f573c71d58c005. Impact 5/5, effort 2/5 (1-2h), 2 rubric points at stake.*


### 2. [CRITICAL] CI-01 Direct answer passages

**Observed.** 0 heading(s) across the sample are phrased as questions. 0 of 25 pages open with a definitional sentence in the first 400 characters. No priority questions were supplied at intake, so question-level coverage could not be measured.

**Why it matters.** An answer engine reuses passages that stand on their own. Copy that requires the surrounding page to make sense is rarely quotable.

**Do this.** Give each priority question its own heading and a self-contained two-to-four sentence answer directly beneath it.

*Evidence: ev_d371929ed472fd56, ev_dd7a0953c4cd4d19, ev_6ea47fbdbca89d34, ev_be45aba7a3ecf70d. Impact 4/5, effort 3/5 (half a day), 3.75 rubric points at stake.*

> Depends on: CI-03


### 3. [MAJOR] CI-02 Claim support

**Observed.** 67 distinct material claim candidate(s) were detected across 25 page(s) in 92 occurrence(s): superlative (55), money (6), percentage (3), count (2), compliance (1). 1 of the distinct claims repeat on more than half the sample, which is navigation, footer, or banner text rather than a claim a page makes for itself. 0 appear only on pages carrying no external link that could serve as a source. Links to the organization's own social or code profiles are excluded from this count: they identify the publisher, not the origin of a number.

**Why it matters.** An unsupported number is a liability in an AI answer: it can be repeated without its caveats, and it cannot be defended when a buyer asks where it came from.

**Do this.** For each material claim, state the source, the measurement period, and the scope next to the claim itself, or remove the claim.

*Evidence: ev_d371929ed472fd56, ev_dd7a0953c4cd4d19, ev_6ea47fbdbca89d34, ev_be45aba7a3ecf70d. Impact 4/5, effort 3/5 (half a day), 2.5 rubric points at stake.*

> **Do not fabricate.** Do not invent a source, statistic, or methodology. If a claim has no support, either remove the claim or leave a marked TODO for the site owner to supply the evidence.

> Depends on: CI-04


### 4. [CRITICAL] TA-03 Comparative and decision content

**Observed.** 2 of 25 sampled pages contain comparison language, 1 state a limitation or fit boundary, and 0 use a table. No competitors were supplied at intake.

**Why it matters.** Buyers ask comparison questions, and so do the people prompting answer engines. A site that never states who it is not for cannot be matched to who it is for.

**Do this.** Publish one honest comparison or decision page with a criteria table, and state explicitly where the offering is not the right fit.

*Evidence: ev_d371929ed472fd56, ev_dd7a0953c4cd4d19, ev_6ea47fbdbca89d34, ev_be45aba7a3ecf70d. Impact 4/5, effort 3/5 (half a day), 3 rubric points at stake.*


### 5. [CRITICAL] TE-03 Customer and outcome evidence

**Observed.** 6 of 25 sampled pages reference customers, clients, or engagements in their own text rather than in a block repeated across the sample; 3 state an outcome-shaped result; 6 long quotation(s) that could be testimonials were found; and 0 page(s) mention a baseline or measurement period.

**Why it matters.** Outcome claims without a baseline and a scope cannot be checked, and an engine that repeats them exposes the business to a claim it cannot defend.

**Do this.** Publish at least one case example with the situation, the intervention, the measurement window, the baseline, and the result, with the customer's permission recorded.

*Evidence: ev_d371929ed472fd56, ev_dd7a0953c4cd4d19, ev_6ea47fbdbca89d34, ev_be45aba7a3ecf70d. Impact 4/5, effort 4/5 (1-2 days), 3 rubric points at stake.*

> **Do not fabricate.** Do not invent customers, testimonials, names, logos, or outcomes. Leave a marked TODO for the site owner. A fabricated testimonial is fraud, not a fix.

> Depends on: TE-01


### 6. [MINOR] CI-03 Stable passage structure

**Observed.** Across 25 sampled pages the median page carries 3 heading(s). 15 declare exactly one H1, 0 give at least one heading an id that can be linked directly, and 25 use a table or list. 10 page(s) declare more than one H1. 7 page(s) skip a heading level: https://tln.example/: h1 followed by h3; https://tln.example/home/: h1 followed by h3; https://tln.example/about-us/: h1 followed by h3. No sampled heading carries an id, so no passage on the site can be linked or cited directly.

**Why it matters.** Headings and anchors are the unit an engine extracts. A page with no internal structure can only be cited whole, which usually means it is not cited at all.

**Do this.** Give every page one H1, a descriptive H2 for each distinct answer, and a stable id on each H2 so a passage can be linked directly.

*Evidence: ev_99d14a7a9a1356f6, ev_606797b76fc96939, ev_b445e84670f54e32, ev_d1f573c71d58c005. Impact 3/5, effort 2/5 (1-2h), 2 rubric points at stake.*


### 7. [MINOR] EN-01 Organization identity consistency

**Observed.** Structured-data identity names: 'TLN' on 4 page(s). Declared og:site_name values: 'TLN'. No email, telephone, or postal address appears in structured data. 24 distinct meta description(s) were served across 25 pages.

**Why it matters.** An engine has to decide which entity a page is about before it can recommend that entity. Conflicting names and missing contact facts force it to guess.

**Do this.** State one organization or person name, one description, and one contact path consistently in visible content and in structured data on every page.

*Evidence: ev_d371929ed472fd56, ev_dd7a0953c4cd4d19, ev_6ea47fbdbca89d34, ev_be45aba7a3ecf70d. Impact 3/5, effort 2/5 (1-2h), 2 rubric points at stake.*

> **Do not fabricate.** Do not invent a legal entity name, address, or contact detail.


### 8. [MAJOR] TA-04 Internal topic relationships

**Observed.** Across 25 sampled pages the median page carries 60 internal link(s). 19 sampled page(s) received no internal link from any other sampled page, and 0 received exactly one. Pages with no inbound internal link in the sample: https://tln.example/about-us, https://tln.example/best-practices, https://tln.example/blog-2, https://tln.example/clients, https://tln.example/home, https://tln.example/hpleads.

**Why it matters.** Internal links are how evidence pages get discovered and how a crawler learns which pages a site treats as important. An orphaned page carries its authority nowhere.

**Do this.** Link every evidence and educational page from the commercial page it supports, and from a topic index.

*Evidence: ev_99d14a7a9a1356f6, ev_606797b76fc96939, ev_b445e84670f54e32, ev_d1f573c71d58c005. Impact 3/5, effort 2/5 (1-2h), 2.25 rubric points at stake.*

> Depends on: TA-01


### 9. [MAJOR] TA-02 First-hand experience

**Observed.** 5 of 25 sampled pages carry 600 or more words. First-hand markers were detected on 2 page(s): first person operator on 0, measurement on 1, artifact on 2. No link to a public repository, dataset, or downloadable artifact was found in the sample.

**Why it matters.** First-hand detail is the part of a page a competitor cannot copy and a summariser cannot generate. It is the strongest reason for an engine to prefer this source.

**Do this.** Publish the method, the numbers, and the artifact behind at least one claim per topic: a repository, a dataset, a template, or a documented before-and-after measurement.

*Evidence: ev_d371929ed472fd56, ev_dd7a0953c4cd4d19, ev_6ea47fbdbca89d34, ev_be45aba7a3ecf70d. Impact 4/5, effort 4/5 (1-2 days), 2 rubric points at stake.*

> **Do not fabricate.** Do not invent research, case evidence, or first-hand experience. Only restructure experience the site already describes.

> Depends on: TA-01


### 10. [UNRESOLVED] TA-01 Priority-question coverage

**Observed.** No priority-question set was supplied at intake, so question-to-page coverage could not be built.

**Why it matters.** Without a bounded question set there is no way to say which buying questions the site answers.

**Do this.** Agree a bounded set of customer questions drawn from real sales conversations, then map each to one owning page.

*Evidence: ev_99d14a7a9a1356f6, ev_606797b76fc96939, ev_b445e84670f54e32, ev_d1f573c71d58c005. Impact 3/5, effort 2/5 (1-2h), 2 rubric points at stake.*

> Depends on: CI-01


### 11. [UNRESOLVED] EN-02 Product and service identity

**Observed.** No offerings were supplied at intake, so offering coverage could not be measured. No Product, Service, or SoftwareApplication entity is declared in structured data on any sampled page.

**Why it matters.** An engine can only recommend an offering it can name, locate at a stable URL, and define. Offerings that live only in prose are hard to recommend specifically.

**Do this.** Give each offering one page with a stable URL, a one-sentence definition, a named audience, and structured data that ties it to the organization.

*Evidence: ev_99d14a7a9a1356f6, ev_606797b76fc96939, ev_b445e84670f54e32, ev_d1f573c71d58c005. Impact 3/5, effort 3/5 (half a day), 2 rubric points at stake.*

> Depends on: EN-01


### 12. [MINOR] EN-03 People and authorship identity

**Observed.** 2 Person entit(ies) and 0 structured author reference(s) were found. 2 Person entit(ies) link to external profiles via sameAs.

**Why it matters.** Expertise content that no identified person stands behind is difficult to weigh, and difficult to defend when a buyer asks who wrote it.

**Do this.** Attribute every substantive page to a named person with a linked biography, role, and credentials, and connect that person to the organization in structured data.

*Evidence: ev_d371929ed472fd56, ev_dd7a0953c4cd4d19, ev_6ea47fbdbca89d34, ev_be45aba7a3ecf70d. Impact 3/5, effort 3/5 (half a day), 2 rubric points at stake.*

> **Do not fabricate.** Do not invent an author, biography, role, or credential.

> Depends on: EN-01


### 13. [UNRESOLVED] AC-01 Intentional crawler treatment

**Observed.** No named AI crawler has an explicit robots.txt group; all are covered by the wildcard group. The wildcard group allows 1 and disallows 13 path pattern(s). No crawler intent was stated at intake, so alignment cannot be judged.

**Why it matters.** Crawler treatment should be a decision, not an accident. Defaults that nobody chose are the common case, and they are the easiest thing to get wrong in either direction.

**Do this.** Decide explicitly which AI crawlers may retrieve the site, record the decision, and express it in robots.txt so a future deploy cannot silently change it.

*Evidence: ev_22600255b991efe6. Impact 2/5, effort 1/5 (under 1h), 1 rubric points at stake.*


### 14. [MINOR] CI-05 Content freshness signals

**Observed.** 3 of 25 sampled pages expose a date in visible text, structured data, or a time element (12.0%). 1 page(s) show structured dates that do not overlap the visible dates: https://tln.example/: structured dates ['2012', '2017'] do not overlap visible dates ['2021', '2022', '2023', '2024', '2025', '2026'].

**Why it matters.** A dated page can be judged current. An undated page has to be treated as possibly stale, which is a reason not to cite it.

**Do this.** Show a visible published and updated date on every article, and emit the same values as datePublished and dateModified in structured data.

*Evidence: ev_99d14a7a9a1356f6, ev_606797b76fc96939, ev_b445e84670f54e32, ev_d1f573c71d58c005. Impact 2/5, effort 2/5 (1-2h), 1.5 rubric points at stake.*


### 15. [MINOR] TE-01 Ownership and accountability

**Observed.** About-style page(s) found: 7. Contact path(s) found: 2. 2 distinct mailto destination(s) and 1 form(s) appear in the sample. 2 organization or person entit(ies) are declared in structured data.

**Why it matters.** Accountability is a precondition for being recommended. An engine asked to suggest a provider has little reason to name one whose operator and contact path are unclear.

**Do this.** Name the operating entity, the responsible people, and a working contact path on a linked page, and mirror those facts in structured data.

*Evidence: ev_d371929ed472fd56, ev_dd7a0953c4cd4d19, ev_6ea47fbdbca89d34, ev_be45aba7a3ecf70d. Impact 2/5, effort 2/5 (1-2h), 1 rubric points at stake.*


### 16. [MINOR] TE-02 Expertise and credential support

**Observed.** 7 distinct credential-shaped claim(s) were detected across 25 page(s) in 55 occurrence(s); 0 appear within 120 characters of a date. 2 of the distinct claims repeat on more than half the sample, so they are stated as site furniture rather than in the page that would explain them. 11 page(s) show six or more images with under 400 words, a pattern consistent with an unexplained logo wall.

**Why it matters.** A credential without a scope, a date, and a responsible entity cannot be verified, and an unverifiable claim is a risk rather than an asset when repeated by a machine.

**Do this.** State the issuing body, the scope, the date, and the entity the credential belongs to next to every credential claim.

*Evidence: ev_d371929ed472fd56, ev_dd7a0953c4cd4d19, ev_6ea47fbdbca89d34, ev_be45aba7a3ecf70d. Impact 2/5, effort 2/5 (1-2h), 2 rubric points at stake.*

> **Do not fabricate.** Do not invent or upgrade a credential, certification, audit, or partnership. These are verifiable facts about a real organization and getting one wrong is a legal problem.


### 17. [MINOR] CI-04 Source attribution

**Observed.** 25 of 25 sampled pages link to an external destination that could serve as a source (100.0%). 445 external link(s) were found in total, but own social and code profiles are excluded from the source count. 0 link(s) point at a standards body, government, academic, or research host. 0 page(s) exceed 600 words while citing no external source.

**Why it matters.** Attribution is how a claim survives being repeated. A page that cites its sources gives an engine a reason to trust the passage it is about to reuse.

**Do this.** Cite the standard, study, or filing next to the claim it supports, with the source name and date visible in the text rather than only in a link.

*Evidence: ev_d371929ed472fd56, ev_dd7a0953c4cd4d19, ev_6ea47fbdbca89d34, ev_be45aba7a3ecf70d. Impact 2/5, effort 2/5 (1-2h), 1.5 rubric points at stake.*

> **Do not fabricate.** Do not invent citations. Link only sources you can verify exist and actually support the nearby claim.


### 18. [MINOR] TE-04 Policy and limitation clarity

**Observed.** Privacy policy reference(s): 1. Terms reference(s): 0. 1 form(s) appear in the sample. 25 page(s) show evidence of third-party analytics being loaded. No terms or service-boundary page was found in the sample.

**Why it matters.** Policies set the boundaries a buyer and an agent both need before acting. Their absence is a friction point at exactly the moment a recommendation would convert.

**Do this.** Publish a dated privacy policy covering analytics and form data, and a terms or engagement-scope page that matches what is actually sold.

*Evidence: ev_d371929ed472fd56, ev_dd7a0953c4cd4d19, ev_6ea47fbdbca89d34, ev_be45aba7a3ecf70d. Impact 2/5, effort 2/5 (1-2h), 1.5 rubric points at stake.*


### 19. [MINOR] CR-03 Sitemap quality

**Observed.** 8 sitemap document(s) were retrieved; 8 parsed cleanly and 0 did not. 573 unique URL(s) are listed. 3 duplicate entr(ies) were listed. 576 of 576 entries carry a lastmod value (100.0%). 25 listed URL(s) were sampled and returned a successful response.

**Why it matters.** A sitemap is the cheapest way to tell every retrieval system which URLs are canonical and when they last changed.

**Do this.** Correct the entries identified in the evidence and ensure lastmod reflects real content changes.

*Evidence: ev_22600255b991efe6, ev_d154ab37597a9b73, ev_fa39beb45d74a750, ev_f353c2379dbf2b58. Impact 2/5, effort 2/5 (1-2h), 0.5 rubric points at stake.*


### 20. [UNRESOLVED] MR-03 Recheck discipline

**Observed.** No recheck schedule for the rubric, crawl sample, question set, competitors, and evidence was recorded at intake.

**Why it matters.** Changing method between reports without disclosure makes movement uninterpretable.

**Do this.** Fix a recheck cadence and record the rubric version, crawl sample, and question set used for each run so score movement can be attributed to the site rather than the method.

*Evidence: none recorded. Impact 2/5, effort 1/5 (under 1h), 0.5 rubric points at stake.*

> Depends on: MR-01


### 21. [UNRESOLVED] EN-04 External identity reconciliation

**Observed.** The site links out to 21 distinct external host(s); 3 sameAs URL(s) are declared in structured data. No owned profiles were declared at intake, so nothing could be reconciled.

**Why it matters.** Owned profiles that the site never links, and profiles carrying an old brand, split one identity into several that an engine must reconcile on its own.

**Do this.** Link every owned profile from the site and declare it in sameAs, and retire or update profiles carrying a former brand.

*Evidence: ev_99d14a7a9a1356f6, ev_606797b76fc96939, ev_b445e84670f54e32, ev_d1f573c71d58c005. Impact 2/5, effort 2/5 (1-2h), 1.5 rubric points at stake.*

> **Do not fabricate.** Do not invent external profiles or links to them.

> Depends on: EN-01


### 22. [UNRESOLVED] MR-01 Baseline question set

**Observed.** No versioned baseline question set was recorded at intake.

**Why it matters.** Without a question set fixed before results are known, any later movement can be explained after the fact.

**Do this.** Agree and version a question set covering buying questions, brand variations, comparison questions, and geography, dated before the first measurement.

*Evidence: none recorded. Impact 2/5, effort 2/5 (1-2h), 1 rubric points at stake.*


### 23. [UNRESOLVED] AC-03 Agent-facing transaction boundaries

**Observed.** 1 page(s) expose a form and 3 mention a purchase surface. The collector never submits a form or performs any state-changing request, so interaction boundaries were not exercised.

**Why it matters.** An agent acting on a customer's behalf needs to know what commits the customer and what does not. Ambiguity here creates unintended commitments.

**Do this.** Document what each form or endpoint commits the visitor to, require authentication for anything that creates an obligation, and state the approval step in public documentation.

*Evidence: ev_99d14a7a9a1356f6, ev_606797b76fc96939, ev_b445e84670f54e32, ev_d1f573c71d58c005. Impact 2/5, effort 2/5 (1-2h), 0.5 rubric points at stake.*


### 24. [UNRESOLVED] CR-05 Rendered content availability

**Observed.** Raw HTML (no JavaScript execution) already carries a median of 191 words and a median of 40 script element(s) per sampled page. 19 of 25 pages carry 150 or more words before rendering, and 25 expose an H1 in raw HTML. 0 page(s) include noscript content.

**Why it matters.** Content that exists only after client-side execution is unavailable to any retrieval path that does not render, and unreliable for those that do.

**Do this.** Serve primary claims, headings, navigation, and supporting evidence in the initial HTML response.

*Evidence: ev_99d14a7a9a1356f6, ev_606797b76fc96939, ev_b445e84670f54e32, ev_d1f573c71d58c005. Impact 2/5, effort 3/5 (half a day), 1.5 rubric points at stake.*


### 25. [UNRESOLVED] MR-02 Referral and conversion attribution

**Observed.** Public markup and response headers indicate 2 analytics or measurement provider(s): google-analytics.com, googletagmanager.com. Whether referrals from AI answer surfaces can be isolated and tied to qualified outcomes cannot be determined from public pages.

**Why it matters.** Reporting mentions or impressions without connecting them to qualified behaviour cannot justify continued spend.

**Do this.** Configure a referral segment for AI answer surfaces and connect it to a defined qualified-outcome event, or record the implementation backlog to do so.

*Evidence: ev_99d14a7a9a1356f6, ev_606797b76fc96939, ev_b445e84670f54e32, ev_d1f573c71d58c005. Impact 2/5, effort 3/5 (half a day), 1 rubric points at stake.*

> Depends on: MR-01


### 26. [MINOR] AC-02 Machine-readable guidance quality

**Observed.** 1 machine-readable guidance file(s) are served: /llms.txt (62861 bytes, 120 link(s)). These conventions are voluntary. Serving them is not a ranking or citation guarantee, and no engine is known to require them.

**Why it matters.** A machine-readable index is cheap to publish and cheap to get wrong: a stale one misleads, and an overstated one invites claims the site cannot support.

**Do this.** Keep the guidance file generated from the same source as the sitemap so it cannot go stale.

*Evidence: ev_62b65440572d4ba4, ev_eaf0dcced0528513, ev_0b7f8ce041452a69, ev_c9beb38676fa86cf. Impact 1/5, effort 1/5 (under 1h), 0.5 rubric points at stake.*


### 27. [MINOR] SD-04 Metadata consistency

**Observed.** Across 25 sampled pages: 25 declare a title, 25 a meta description, 25 an html lang attribute, and 25 Open Graph titles. 1 title(s) are reused across multiple pages: 'Get a healthy strategy with powerful results - TLN' on 2 pages.

**Why it matters.** When a page describes itself differently in its title, its social metadata, and its structured data, an engine has to choose one description and may choose the wrong one.

**Do this.** Generate title, description, canonical, og:url, and structured-data values from one source per page so they cannot drift apart.

*Evidence: ev_99d14a7a9a1356f6, ev_606797b76fc96939, ev_b445e84670f54e32, ev_d1f573c71d58c005. Impact 1/5, effort 2/5 (1-2h), 0.5 rubric points at stake.*


## What this audit could not determine

- This audit sampled 25 page(s) against a configured limit of 25 at a maximum depth of 2. Findings describe the sample, not every page on the site.
- Scored against rubric version 0.2 using tool version adia/0.1.0. Score movement between runs can come from a rubric change as easily as a site change.
- Collection used plain HTTP retrieval with no JavaScript execution. Content that appears only after client-side rendering was not observed.
- Off-domain properties were not retrieved. The run was restricted to the authorized registrable domain.
- No form was submitted, no authenticated content was accessed, and no state-changing request was made.
- 9 check(s) are inconclusive and contribute no points in either direction: CR-05, EN-02, EN-04, TA-01, AC-01, AC-03, MR-01, MR-02, MR-03. Their 22 points were removed from the denominator rather than scored as zeroes.
- Collection stopped at a configured bound (max_pages=25). A larger sample may change coverage-dependent findings.
- This audit measures conditions a site can control against rubric version 0.2. It does not guarantee placement, ranking, citation, recommendation, traffic, leads, or revenue in any search engine or AI answer product, and a low score does not prove a site is absent from AI-generated answers.
- Unresolved checks: CR-05, EN-02, EN-04, TA-01, AC-01, AC-03, MR-01, MR-02, MR-03. These contributed no points in either direction.

Do not "fix" any of the above. They were not measured, so a change there
cannot be verified by a re-audit.

## When you are done

Re-run the audit against the same rubric version and compare:

    python3 -m adia collect --intake intake/tln.example.json --out artifacts/tln-com-recheck
    python3 -m adia report  --out artifacts/tln-com-recheck
    python3 -m adia compare --baseline samples/tln.example/audit-result.json --candidate artifacts/tln-com-recheck/audit-result.json

The delta is the proof the work landed. Fix, re-audit, repeat until the score
stops moving — then stop, rather than optimizing a number past the point where
it means anything.