# How the budgeted briefing works

Read the LLMS-TLDR pipeline, selection policy, and token-estimation limits.

> Experimental vendor proposal · opt-in publication · operator-managed client integration

## Purpose

`/llms-tldr.txt` is a bounded, deterministic publication of literal stored WordPress content. It exists for an operator who wants one inspectable site briefing with an explicit approximate token budget.

This CYBERMAPS-specific format complements llms.txt with a compact extract of stored content. It uses deterministic selection rather than AI generation. Operators can supply its URL to a compatible retrieval client or use the output in their own workflow.

## Design constraints

The generator must:

1.  use the same eligible-content inventory as other Cybermaps publications;
2.  execute no shortcode, dynamic block, model, embedding, or remote service;
3.  state its exact stable selection order;
4.  account for every emitted byte, including the header;
5.  include only complete entries;
6.  report selected and omitted counts truthfully; and
7.  stay within the configured approximate budget.

## Input inventory

`Cybermaps\Discovery\PublicationInventory` supplies published, eligible content using the unified publication-eligibility decision. The inventory respects configured AI post types, exclusions, redirects, off-resource canonicals, Genesis/Mai settings, and supported SEO-plugin signals.

The generator does not inspect drafts, private resources, password-protected resources, or attachments.

## Selection order

The stable order is:

1.  configured pinned resource IDs, in the operator’s order;
2.  configured content-type order;
3.  stored modified time descending; and
4.  WordPress object ID ascending as the final tie-breaker.

For each resource the generator renders a complete candidate entry. If adding that entry and the updated coverage header would exceed the budget, that entry is omitted and later candidates are still considered.

There is no quality score, taxonomy cluster, entity extractor, semantic deduplication, information-gain calculation, or inferred priority.

## Extraction

`Cybermaps\Content\VisibleTextExtractor` derives literal visible text from stored post content. It strips markup and ignores non-rendered stored elements; it does not execute shortcodes or server-rendered dynamic blocks.

Each briefing entry contains:

*   title;
*   public URL;
*   WordPress content type;
*   stored last-modified timestamp when valid; and
*   a fixed-length literal extract.

The extract may be incomplete by design, but an entry is never cut midway by the budgeter.

## Budget calculation

The estimator is intentionally simple and disclosed:

```
estimated_tokens = ceil(number_of_UTF-8_bytes / 4)
```

The configured range is 1,000–200,000 estimated tokens; the default is 80,000. This is not a tokenizer for any particular model. It is a repeatable planning estimate.

Every calculation uses the complete candidate output:

*   format metadata;
*   selection and extraction disclosures;
*   configured license assertion;
*   coverage counts; and
*   selected entries.

Therefore `token_estimate <= token_budget` for the emitted publication under the documented estimator.

## Header

The publication states:

```
Profile: Cybermaps Budgeted Site Briefing
Format-Version: 0.2-draft
Status: Experimental vendor proposal
Known-Automatic-Consumers: none documented
Selection-Method: ...
Content-Extraction: ...
Token-Estimate-Method: ...
Token-Budget: ...
Coverage: selected N of E eligible; omitted O; partial 0
```

`partial` is always zero in 0.2 because the generator includes or omits whole entries.

A license value is labeled `License-Assertion` because it is supplied by the site operator. Cybermaps does not verify ownership, legal scope, or enforceability.

## Caching and delivery

The canonical dynamic response uses a one-hour WordPress transient. Content and relevant settings changes invalidate the discovery cache. With full static publication enabled, the same generated body can be materialized at `/llms-tldr.txt`; localized variants can also be materialized when translation output is configured.

Physical delivery bypasses PHP. It therefore cannot appear in PHP-observed crawler analytics and uses server-controlled response headers.

## Limitations

*   The byte-to-token ratio is approximate and model-independent.
*   Stored-text extraction cannot represent runtime output from dynamic blocks or shortcodes.
*   The fixed extract can omit important context.
*   Pinned IDs express operator priority, not objective importance.
*   Stable modified-date ordering favors recent edits, not quality.
*   No semantic duplicate detection occurs.
*   Client integration is configured by the operator.
*   A static file remains stale until the next complete reconciliation.

## Appropriate use

An operator can inspect or deliver the briefing as one bounded snapshot alongside the canonical pages and `/llms.txt`. Any client should treat each URL as the authoritative source for full context.

The draft is opt-in. Future format changes require a version increment, fixtures for exact budget behavior, and migration documentation so integrations can follow a defined upgrade path.