Good labels can still teach bad habits.
Shortcut cues, hidden target leakage, injected instructions, and poisoned trigger phrases can make a dataset look clean while teaching the wrong rule.
Research thesis
A training example should be inspectable before it teaches anything.
Plain prompt-answer-label records are easy to generate and hard to trust. AnchorCell starts from the opposite direction: the record should show what it trusts, what it refuses, what evidence supports it, what could fool it, and what can safely leave the authoring environment.
Make the decision structure part of the record.
The format should expose trusted policy, untrusted input, grounded facts, uncertainty, forbidden assumptions, candidate branches, and tested export views.
The unit
One AnchorCell is one scoped decision object.
Stable cell id, family id, version, lifecycle state.
Created by, transformed by, source refs, canonical hash.
One decision question and a tight scope boundary.
Trusted policy, untrusted inputs, facts, uncertainty.
Primary path, tempting shortcut, adversarial interpretation.
Tests, invariants, human signoff, export readiness.
Trust separation
The context packet is not a blended summary.
The central design rule is that instructions and data do not share one undifferentiated channel. An input can be quoted, analyzed, hashed, or redacted without becoming policy.
trusted_policyWhat the reviewer accepts as rules.
untrusted_inputsWhat the record received but does not obey by default.
extracted_factsClaims grounded to explicit support.
uncertaintiesWhat is unknown and decision-relevant.
forbidden_assumptionsWhat the model must not silently fill in.
Anti-shortcut design
The wrong answers should be modeled on purpose.
AnchorCell records should include branches that show the decision boundary, not just the final label.
Supported path
The answer that follows verified facts and scoped policy.
Tempting shortcut
The plausible bad answer caused by urgency, authority, formatting, or metadata cues.
Attack-shaped path
The response that checks whether injected or obfuscated text can bend the record.
Export views
One authoring record, several compiled views.
Public material should be produced by an export compiler, not by hand-cleaning a rich internal record. The page claim stays narrow: the method is being shaped so training input, target output, review material, and public redactions are separate artifacts.
- author_fullcontrolled authoring record
- review_packethuman audit view
- train_inputmodel-facing input
- train_targetsupervision target
- public_redactedpublic-safe explanation
What would prove it
The proof is not rhetoric. It is a validator stack.
JSON Schema Draft 2020-12 authoring baseline: closed objects, bounded fields, canonical roles.
Cross-field checks for refs, relations, and forbidden assumptions.
Injection, trigger, shortcut, and leakage probes.
Zero known public leakage before publishing any public view.
This is not a finished model claim. The release-ready direction is a visible research surface, followed by a specification, validators, example cells, adversarial tests, and public-safe exports.
Open the brief Inspect v2 schema Open example cell