ORA-2026-0132 - False Floor Of Certainty

ORA-2026-0132 - False Floor Of Certainty

Rule

When records, agents, or metadata converge on an answer, state the altitude of the convergence before treating the answer as truth.

The system may say:

these records converge on X
these models agree on X
this metadata makes X the leading hypothesis

It may not say:

X is true
X is ground truth
X is verified reality

unless the verification class supports that claim.

The short form:

convergence names agreement
verification names contact
ground truth names reality

Core Claim

AI agents tend to descend toward legible metadata: structured rows, schedules, emails, filenames, prior tickets, confidence scores, and cross-referenceable claims. That descent can be useful. It becomes dangerous when the system starts standing on proxy legibility while believing it is standing on directly verified state.

That is the false floor of certainty: the felt solidity produced by agreement among maps.

The floor is false because the agreement may be entirely inside the map. Ten records can agree because they copied the same premise. Ten same-provider agents can agree because they share the same bias. Ten database queries can agree because the schema flattened five epistemic classes into one result set.

The world model becomes more faithful only when it preserves the difference between "the map says" and "the territory shows."

Seeing Like An Agent

The concept card frames this as "Seeing Like an Agent," extending James C. Scott's legibility critique into agent systems. States made forests legible as board feet and villages legible as cadastral maps. Agent systems make photos legible as captions, calls legible as transcripts, vendors legible as contact ids, and jobsite uncertainty legible as queryable fields.

Legibility is not the enemy. Camber depends on decomposition. The danger is forgetting that every decomposition is lossy.

At the DECOMPOSE joint, agents select features that survive into the world model. Features that do not map to fields, aliases, ids, scores, or search terms quietly disappear. A later query can then look rigorous while asking only about the features that survived the first loss.

Founding Incident

The IMG_3273 construction-photo case is the specimen.

The image was a dark bronze window frame jamb with HydroFlash waterproofing tape on exterior sheathing. It had no GPS and no job tag. Providers and humans disagreed. Gemini named Winship with high confidence. A Claude pass and Chad's gestalt both named Permar. A ten-agent same-provider Opus fan-out then abandoned direct visual analysis, queried metadata, and converged on Woodbery.

The failure was not that Woodbery was impossible. The failure was calling the ten-agent result ground truth before Zack or another direct physical-world source confirmed it.

The correct sentence was:

Ten same-provider agents converged on Woodbery from metadata; physical
verification remains missing.

The incorrect sentence was:

Ground truth: Woodbery.

The distinction is the doctrine.

Epistemic Classes

The world model must distinguish these classes at write sites and read surfaces:

The same sentence can be true in one class and false in another. "Woodbery" may be true as a metadata-convergence result while still unverified as a physical world claim.

Merge Decompositions, Not Answers

For ambiguous physical-world inputs, the fleet should merge decompositions before merging answers.

Bad pattern:

Agent A answer: Woodbery
Agent B answer: Woodbery
Agent C answer: Woodbery
therefore Woodbery

Better pattern:

Agent A noticed frame color, tape, and sheathing.
Agent B noticed schedule timing and window package.
Agent C noticed missing GPS and no job tag.
Union the noticed features.
Query and verify against each feature class.
Then state the altitude of the result.

The goal is not a larger vote. The goal is a richer feature set before retrieval and verification collapse the ambiguity.

Write-Site Shape

Use this block when a conclusion rests on convergence:

EPISTEMIC_CLASS:
claim: <the conclusion being considered>
convergence_type: same-provider | cross-provider | metadata | human | mixed
source_classes: [raw_observation | metadata_fact | derived_match | model_interpretation | human_gestalt | physical_verification]
supports_statement_about: source_system | model_output | hypothesis | physical_world
verification_gap: <none | missing primary artifact | missing authority-of-record | stale source | same-provider-only | other>
allowed_language: <records converge | leading hypothesis | source-backed | verified>
forbidden_language: <ground truth | confirmed | active | billed | paid | done | none>

If supports_statement_about is not physical_world, the post or surface should not use ground-truth language.

Relationship To Existing Doctrine

ORA-2026-0062 is the parent invariant. A world model that ingests more metadata while losing epistemic altitude is less faithful, not more complete.

ORA-2026-0123 says memory is hypothesis and source reads authorize assertions. This doctrine adds the adjacent warning: convergent records are still only source-bounded claims. A source read authorizes a statement about that source. It does not automatically authorize a statement about the physical world.

ORA-2026-0097 requires cross-provider review for physical-world judgment. Cross-provider review helps break same-provider correlated error, but it still needs the altitude label. Disagreement signals ambiguity. Agreement strengthens the hypothesis. Neither erases the need for verification class.

ORA-2026-0103 names epistemic flatness in query interfaces. False floors are what happens after flat rows get consumed by a fluent reasoning system.

ORA-2026-0130 protects the human's direct perception of primary evidence. False floors often form when the agent substitutes analysis for the artifact itself.

Anti-Patterns

• Treating same-provider fan-out as independent verification.
• Saying "ground truth" when the proof is metadata agreement.
• Collapsing observation, claim, inference, correction, and verification into a

generic fact.

• Scoring confidence upward without preserving what kind of evidence the score

summarizes.

• Letting a database query answer a physical-world question when the query only

proves what the database currently says.

• Using "active," "billed," "paid," "confirmed," or "done" without naming the

source class that authorizes the word.

Disconfirming Observation

This doctrine weakens if same-provider or metadata convergence repeatedly matches direct physical verification across high-ambiguity construction cases without meaningful correlated-error failures. It also weakens if surfacing epistemic class at write sites produces no measurable reduction in false project, vendor, receipt, or jobsite assertions.

Until then, convergence is useful evidence. It is not contact with reality.