Administrator
Published on 2026-11-12 / 15 Visits
0
0

AI Discovery Needs a Novelty Gate: The Maya-Site Correction

Forty-eight AI agents screened lidar images and flagged a real archaeological structure. The detection was good enough to survive a raw point-cloud check and resemble a site mapped by professional archaeologists. The discovery claim still failed.

On November 11, 2026, Pietro Miotti corrected his account: the Maya site his agents had described as unmapped already appeared in a 2025 paper by Francisco Estrada-Belli and colleagues. His agents had missed it during literature review. The episode is useful because it separates two tasks that research-agent systems often collapse:

  1. Detection: does this dataset contain a credible candidate?
  2. Novelty: is the candidate absent from prior scientific and institutional records?

The first task passed. The second did not. A scientific agent should never be allowed to convert one result into the other without a separate novelty gate.

What the correction establishes

Miotti's account says the workflow converted roughly 1,200 km² of lidar strips into images and asked 48 general-purpose agents to look for pyramids, platforms, plazas, and other regular forms. The agents flagged multiple candidates. For the selected hilltop, they also inspected the raw point cloud to distinguish ground geometry from canopy artifacts.

That is a meaningful detection pipeline. It produced a candidate that was not merely a visual illusion.

The workflow then asked which candidates had not already been discovered and searched literature and archives. Miotti sent a report with confidential coordinates to Mexico's National Institute of Anthropology and History, or INAH. The reply shown in his post says that no registered archaeological site was identified at that location.

This is where the evidence changed scope. A negative result in one institutional register is not a negative result across the scientific literature. The later correction says the same site had been mapped in Estrada-Belli et al. 2025, a regional study that aggregated public and commissioned lidar, including reprocessed environmental lidar from southern Campeche and southern Quintana Roo.

The exact coordinates remain confidential. Public readers therefore cannot independently reproduce the site-to-map match from coordinates. The identity of the prior site should be attributed to Miotti's correction. That limit matters: responsible correction should narrow a claim without creating a second unverifiable certainty.

Detection and novelty produce different artifacts

A research pipeline should make the difference visible in its data model.

Gate Question Minimum artifact Valid output
Detection Is there a candidate in the observed data? Frozen input, candidate coordinates, image and point-cloud evidence Candidate detected
Classification What might the candidate be? Features, alternatives, domain criteria Provisional interpretation
Registration Is it present in a named registry? Registry identity, query, date, response Registered or not found in that registry
Literature novelty Has equivalent evidence been published? Search log, full-text and figure review, citation graph Prior match, unresolved, or no match found
Field confirmation Does independent domain work confirm the object and interpretation? Expert review, survey, excavation, or comparable evidence Confirmed within a stated scope

The phrase not found in registry X is auditable. Unmapped, unknown, and discovered for the first time are much broader claims. They require broader evidence.

Why ordinary literature search missed the answer

The 2025 paper is titled New regional-scale Classic Maya population density estimates and settlement distribution models through airborne lidar scanning. Its abstract describes a 95,000 km² regional model. A specific plaza or hilltop can be present in a map without appearing in the title, abstract, keywords, or searchable prose.

This creates a retrieval mismatch:

  • the candidate is represented as coordinates, terrain geometry, and a local visual pattern;
  • the literature is indexed mainly as text, titles, abstracts, and citations;
  • the decisive prior evidence may live in a figure, map legend, table, appendix, or supplementary file.

A text-only agent can read every abstract returned by a query and still fail the actual novelty task. The missing operation is spatial correspondence: does this coordinate and uncertainty region overlap a previously mapped feature?

The failure also exposes a general systems risk. When one overall workflow generates candidates, selects a promising result, searches for prior work, and drafts the discovery narrative, later stages can inherit the detector's preferred framing. The public record does not reveal whether these steps shared one agent context in this case. Independent novelty review is valuable because it starts with the opposite objective: find the earlier record that would falsify the claim.

An eight-step novelty gate

1. Freeze the candidate identity

Create a candidate package before searching for novelty:

  • coordinates, coordinate reference system, precision, and uncertainty radius;
  • source dataset, acquisition date, processing version, and image hashes;
  • bounding box, scale, orientation, and relevant terrain layers;
  • feature description separated from historical interpretation.

Without this package, different agents may unknowingly search for slightly different objects.

2. Assign separate claim states

Use a state machine instead of a Boolean novel=true field:

candidate_detected → artifact_checked → registry_checked → literature_checked → spatially_matched → domain_validated

Each state records pass, fail, or unresolved. No state inherits the success of the previous one.

3. Search by multiple identities

Search more than the candidate's proposed name. Include region, watershed, survey block, dataset identifier, nearby known sites, expedition names, authors, lidar programs, alternate spellings, and earlier citations. Build a citation graph in both directions from the closest regional papers.

4. Read non-text evidence

Download candidate papers and inspect every map, figure, caption, table, appendix, and supplement relevant to the region. Apply OCR and image extraction when needed, but preserve the original page as the verification artifact. A successful abstract search is not a successful map review.

5. Normalize spatial references

Convert every candidate and prior record into a declared coordinate reference system. Preserve the original coordinates, transformation method, precision, and uncertainty. Compare points, polygons, and map extents with a documented buffer instead of demanding false exactness.

6. Keep registries separate

Record each institutional database independently. A missing registry entry can support an unregistered claim for that system and date. It cannot close literature, field-note, map, heritage, or local-knowledge channels.

7. Run an adversarial countersearch

Give the frozen candidate package to an independent reviewer whose task is to disprove novelty. The reviewer should not receive the detection team's preferred story. Its output must include the strongest candidate prior matches, rejected matches, unresolved sources, and search limits.

8. Bind wording to evidence

Use the narrowest supported statement:

  • candidate detected when the signal survives artifact checks;
  • not found in registry X as of date Y after a named registry query;
  • no prior match found in the searched corpus after a documented but incomplete review;
  • rediscovered when a credible prior match exists;
  • newly documented only after the required scholarly and institutional checks support it.

What the agents still accomplished

The correction should not be flattened into agents failed. The candidate appears to have been a true positive. General-purpose agents found a structure in rendered lidar that professional archaeologists had also mapped. That is evidence for candidate-screening utility within this case, though the unpublished agent traces and single episode do not establish a general detection rate.

The right engineering response is to retain the detector and replace its authority boundary. Candidate generation can be broad and fast. Novelty claims must be slower, adversarial, and tied to artifacts that another researcher can inspect.

This follows the same logic as a scientific-agent verification stack: different checks expose different failure modes. It also extends Claude-shaped science with a specific discovery interface. A calculation can have an executable oracle; a novelty claim needs a corpus boundary, spatial match, and falsification search.

The operational rule

Do not ask one research agent to find a candidate and then accept its own statement that the candidate is new. Freeze the candidate, move novelty review into an independent context, make maps and coordinates first-class retrieval objects, and publish the evidence state with the result.

AI can make candidate discovery cheap. That increases the value of disciplined novelty verification. The scarce capability is no longer producing an interesting lead. It is knowing exactly what the lead proves.

Sources


Comment