# One extraction contract, three representations

Status: completed offline preservation/tokenization study, 15 hosted MCP calls and all 60 planned model attempts. Fresh approval October 11 bounded retrieval at 15 attempts/15 credits and models at 60 attempts/US$0.25. Two requests failed before inference; separately authorized continuations retained both failures/reservations and skipped their jobs. The remaining 58 requests produced 28 accepted and 30 incorrect pages. Actual total billing remains unknown; reported usage costs and per-cell correct-output costs are available. Named-author review remains pending.

## Reader task and contract

Return every product in the designated product region, in document order, with exactly `name`, `displayed_price`, `units_sold`, `share`, `sku`, and `buy_url`. Preserve displayed strings, resolve relative links against the supplied source URL, and use null only when that region lacks a field. Do not infer SKU from a URL. No extra products or fields, duplicate rows, coerced numbers, or guessed destinations.

The fixed task identifies the four-column `Product`, `Price`, `Units sold`, `Share` table in the imported catalog, or the product offers under `Desk lamps` / `Travel bottles` in the other cases. Other demonstration tables and documentation links are outside the product region. Every representation still contains the full source conversion; this is task selection, not arm-specific cropping.

The literal oracle is `ground_truth.json`; it is specified separately from the representation generator and never derived from parser output. The fixture author and oracle writer are the same AI execution session, so this is procedural independence, not an independent human labeling exercise. A fresh reviewer must inspect both.

## Frozen inputs

1. `catalog.html`: the first static Requests capture in the previous acquisition study's final live report (`body-007.html`), from examples source `9e5fba96d34a7bb7106a5a0e478a668078c6a312`. Reuse original bytes and record their SHA-256; no new fetch. Three records.
2. `links.html`: an authored, local-only two-product catalog with visible SKUs and identical `Buy` link labels but distinct destinations. Nav, cookie and footer text make whole-page conversion visible.
3. `attributes.html`: an authored, local-only two-product catalog whose SKUs exist only in `data-sku`; names/prices and links remain visible.

The task/schema stays fixed for all three cases. Two deliberately adversarial cases isolate information loss; this small selected set is not representative of the web. Local fixtures stay inside this example, outside the GitHub Pages top-level `fixtures/`. No new public endpoint.

## Comparable format generation (offline)

HTML is the exact saved source, decoded as UTF-8. Markdown runs the inspected converter recipe at extractor source `66d10caea9a90df3c14069bfd910b92f1ad1123a`: BeautifulSoup 4.12.2 `html.parser`, remove script/noscript, html2text 2020.1.16 defaults with `ignore_links=False`. No reader-mode, cropping, summarization, link repair or arm-specific cleanup.

Text uses local Chromium selection (`Range.selectNode(document.documentElement)` then `window.getSelection().toString()`), matching the inspected browser worker's selection recipe at `11cc9f570656298f736979f548c5659c57b64b57`. All network requests are aborted, document scripts are disabled, and every arm starts with the same saved HTML. This is a local reference transformation; it is not the deployed MCP text result or a verification of raw-mode text behavior.

Five local repeats per case/arm check byte determinism, not independent site observations. Count exact representation UTF-8 bytes and content-only tokens with tiktoken 0.11.0 `o200k_base` and `cl100k_base`, including tokenizer asset hashes. Token counts exclude prompts, schema, tool envelopes and billing overhead. They are not observed model charges.

## What is measured now

For each non-null oracle field, test whether its required literal survives in the representation (link URLs may be present relative to the supplied base). The current relative-URL fallback checks pathname only. Every purchase URL in these selected cases has no query or fragment. Before adding query/fragment-sensitive destinations, extend and test that checker; the current score does not establish their preservation. Report all slot results and denominators. Repeated identical values count as separate expected slots; lexical survival does not verify row association, exact-record output or model comprehension.

Counterfactual probes change only link destinations or only `data-sku` values. If two different HTML sources produce identical text/Markdown, a consumer of that representation alone cannot distinguish the changed fields. This proves loss for the selected transform and fixture; it says nothing about model accuracy for retained fields.

## Fixed model experiment — 60 attempts, 58 inference responses

Freeze model `gpt-4.1-mini-2025-04-14`, Chat Completions, temperature=0, top_p=1, max_completion_tokens=1000, no tools/retrieval/history, no retries, identical system/task/schema and source URL. Use five randomized/interleaved repeats for each of 3 cases × 3 arms = 45 calls. Also evaluate each of the 15 saved hosted MCP returns with the same task = at most 60 calls total. Maximum input per request, including prompt/schema overhead: 5000 tokens. Stop before any call exceeding either cap. Repeats describe this model/run only; temperature=0 does not imply determinism.

Record request/task/model/settings hashes, UTC start/end, full response and refusal/truncation/errors, reported model version/system fingerprint if supplied, token usage including cached input, and actual token-rate cost. Exact ordered record equality is the primary accepted-page denominator; correct field slots are secondary. Charge every failed/truncated/incorrect attempt to the numerator. Zero accepted pages => cost per accepted page is undefined. Do not pool unlike tasks or declare a universal winner. Do not drop trials, failures or slow observations.

## Hosted MCP experiment —15 calls executed

Executed using only the existing owned static fixture: `https://scrapingant.github.io/scrapingant-examples/fixtures/html-tables.html`. Initialize one explicitly versioned MCP client, save tools/list schemas, then 3 tools × 5 randomized/interleaved rounds = 15 tools/call requests. Force browser=false, proxy_type=datacenter, proxy_country=US for every arm, no retries, no residential/browser fallback. This is a static acquisition check, not interactive agent planning or browser rendering reliability. Local-only richer cases will not be published to enable live fetching.

Save JSON-RPC requests, response bodies/headers, tool outputs, error strings/isError, elapsed time, source URL, UTC time, and response hashes. The live calls reacquire the fixture, so they cannot establish identical source captures: check against the frozen source and report drift/unknown provenance. Report live acquisition separately from offline same-capture representation results. A successful transport can carry an error string or incomplete content; validate against the oracle.

MCP documentation says these calls use normal retrieval credits. At browser=false/datacenter the documented estimate is 1 credit per successful request: a new ceiling of 15 credits. The inspected tool returns content without per-call credit receipts. A separately authorized existing read-only SQL route verified the expected non-elevated SELECT-only role, primary database, exact test-account scope and one active period. The initial snapshot and 15 post-call snapshots increased by 1 each:15 contemporaneous account credits. No usage API was invoked because source inspection found conditional subscription repair. Account snapshots are not exclusive per-request receipts: concurrent traffic cannot be ruled out. The model phase is reported separately; account deltas are not exclusive retrieval-charge receipts.

## Bounded execution and retained failures

MCP 15/15 attempts completed. Model attempt 1 failed before inference; an explicit owner-authorized continuation preserved its checkpoint/row/response and US$0.0036 reservation, skipped its job and dispatched attempt 2. Attempt 2 also failed before inference without usage/completion. Execution stopped immediately both times. A second separately authorized continuation preserved both checkpoints, prior history and US$0.0072 unknown-charge reservations, then completed attempts 3–60. There were no automatic retries, purchases, account/model fallbacks or additional MCP calls. Operational configuration/provider messages remain private.

Cumulative model count is 60/60; there are no approved attempts left. Both unknown-charge reservations remain: US$0.0072 is allowance accounting, not observed billing. The 58 responses report 48,709 input tokens (14,976 cached) and 4,860 output tokens. At standard published rates the known token-rate cost is US$0.0227668. Cumulative conservative allowance is US$0.0299668; actual total billing is unknown. No extra inference was dispatched to replace missing jobs or balance sample sizes.

Missing inference responses: offline links/Markdown repeat 0 and offline catalog/text repeat 0. All available cell sample counts are n=5 except these two n=4 cells. Matched all-arm comparisons use repeats 1–4 for catalog/links and repeats 0–4 for attributes/hosted catalog. Exact accepted pages: matched catalog HTML4/4, Markdown0/4, text4/4; links HTML4/4, Markdown4/4, text0/4; attributes HTML5/5, Markdown0/5, text0/5; hosted catalog HTML5/5, Markdown0/5, text0/5. Seven unique products remain a small selected corpus, not a website population.

Across all 58 inferences, 28 pages were accepted and 30 incorrect. Twenty-five incorrect responses were empty arrays; five attribute/Markdown responses returned the correct two products/links but null SKUs. Correct-slot denominators include all six fields and expected nulls, unlike non-null preservation slots. The retained report shows all available cell counts and outcomes alongside balanced repeat intersections.

Public hosted captures are in `expected_output/live-mcp-2026-10-11/`; all 60 model attempts are in `expected_output/live-model-2026-10-11/`. Operational provider error messages and account-counter absolute balances/internal period reference were omitted from public copies; HTTP statuses, pre-inference failure markers and minimization hashes remain. Raw private originals were retained outside repositories; public hashes identify the exact downloadable artifacts and private-original hashes document that minimization. HTTP body captures use Requests' decoded response text; credential-bearing response headers are excluded. No receipt or model outcome is fabricated.

HTML and Markdown matched their local catalog references in all 5 observations. Hosted browser=false text was 1083 bytes/436 o200k/440 cl100k tokens, retained the document title and CSS-hidden stock row, and flattened cells to newline-separated strings. Local browser-selection text was 1051 bytes/415 o200k/416 cl100k tokens. Both contained the 12 required table literals; lexical checks remain distinct from the subsequently measured model outcomes. Five hosted HTML outputs matched the frozen source hash; separate Markdown/text calls do not expose underlying source bytes. No hosted browser-mode or richer local-case acquisition was executed.

`mcp_live.py` and `model_run.py` are explicit opt-in runners, independently reviewed before dispatch; default `run.sh` stays offline. MCP ordering uses seed 8, five rounds with all 3 arms interleaved. Model plan uses seed 8,45 same-capture jobs followed by15 hosted-output jobs, common frozen task/schema/system/settings, full serialized request tokenization plus 256 framing tokens before dispatch. All 60 planned inputs preflighted at 751–2375 tokens, below 5000. A request reserves 5000 input + 1000 output cost before transport; every unknown outcome stays reserved and stops, no hidden retries, redirects, model/tier/account fallback or automatic existing-output overwrite. Both explicitly approved rebinds preserved exact stopped checkpoints before appending the next planned job; ordinary fresh-run mode refuses an existing directory. The second rebind also verifies the preceding checkpoint/history and both failed request identities. API-reported model/tier/usage and exact records are checked before another attempt. Guard tests use synthetic transport boundaries only; they are not service-performance evidence.

Approved bounded model ceiling: US$0.25, at most 60 calls. Official model page checked 2026-10-11 lists standard input $0.40/M, cached input $0.10/M, output $1.60/M. Worst permitted allocation without caching: 60 × (5000 × 0.40/M + 1000 × 1.60/M) = $0.216, below $0.25. Reserve the full maximum before transport; unknown usage stays reserved and stops. No purchasing credit, adding payment methods or additional models was authorized.

Costs are token-rate calculations from API-reported usage, not invoice receipts. Incorrect responses remain in the numerator. For zero accepted pages cost per accepted page is undefined; for cells with unknown charges the all-attempt actual cost/ratio is null. Known usage costs are provided separately, so no missing charge is treated as zero. Repeated inputs receive caching discounts and empty arrays use fewer output tokens than correct full records; do not interpret cost differences as isolated effects of representation size.

The analysis separately counts inference responses and completion strings, including refusals/truncation in accuracy denominators when present. Nominal Wilson 95% intervals describe same-input repeats under an independence approximation; correlated repeats and seven selected products cannot support website-population inference. For n=5, 5/5 gives approximately 56.6–100% and 0/5 gives 0–43.4%; these are descriptive, not general reliability claims.

Sources: https://developers.openai.com/api/docs/models/gpt-4.1-mini ; https://docs.scrapingant.com/mcp-server ; https://docs.scrapingant.com/credits-cost ; https://docs.scrapingant.com/llm-markdown (checked 2026-10-11).

## Interpretation and publication gates

Preservation, tokenization, LLM extraction and live MCP behavior are separate outputs. No token reduction/local replay proves production reliability, engineering savings, rankings or paid conversion. Paying-customer success means first-time paid customers; clicks/signups are intermediate events. There is no analytics implementation or customer-attribution result in this packet.

Draft source PRs present completed offline, hosted acquisition and fixed-model evidence, including missing inferences and unknown billing. Publication requires the named author Oleg Kulyk's methodology/conclusions/byline review and explicit further authorization. Keep article `draft: true`; no source merge, deployment, Terraform or issue closure is authorized.