/research/acquisition-grounding
Source grounding
A model can propose. It cannot move the source.
This page shows one real gpt-5.6-sol acquisition run against fixed captured source bytes. Extractive proposals survive only when their claimed location recomputes to the captured text. Abstractive proposals are labeled as synthesis. Ungrounded extractive output is refused at the boundary.
Captured source
https://example.com/acq-llm-live-0-g
Highlighted text means the proposal’s span matched these captured bytes. It does not mean the statement is true or admitted.
source identity
Model proposals
- Extractive · span matched
“Reported short interest exceeded the freely traded float.”
characters 0–57 · sha256:547350bdeb7f…
model extraction confidence 1
- Extractive · span matched
“Multiple brokers restricted opening buy orders on January 28, 2021.”
characters 58–125 · sha256:547350bdeb7f…
model extraction confidence 1
Refused at the grounding boundary
This run recorded no dropped extractive proposals. The gate remains active; zero drops is an observation about this run, not a claim that the model cannot fabricate.
Lineage
- Source bytes
- ↓ Capture receipt
- ↓ Model observer execution
- ↓ Grounding gate
- ↓ Candidate proposal
- ↓ ContentProposal
The grounding filter preserves the original observer execution identity. Surviving candidates do not receive a replacement envelope id.
Located in bytes ≠ true.
Proposal ≠ admitted.
Capture receipt ≠ verdict.
The model cannot assign standing, canonical identity, anchoring, or SRS verification here.
Machine details
Raw model output is not committed; only its digest is recorded. This specimen establishes no truth, admission, source authenticity, standing, canonical identity, anchoring, or SRS verification.