Research — Experiments
Evidence, not assertion.
Each experiment is stated in one repeatable format, so a result can be inspected and challenged rather than taken on trust. Negative and null results are kept.
The experiment format
Every experiment, the same seven parts.
Research question
What property or control is under test.
Control / property
The mechanism or invariant being checked.
Baseline
Behaviour without the control, for comparison.
Fault / challenge
The deliberate failure introduced.
Metric
What is counted, exactly.
Result
What actually happened.
Known limitation
What the result does not show.
Current internal harness pilot
A small pilot, reported in full.
The following is an internal experiment, not independent public validation. It is shown with its ceiling effect and its failures intact.
Internal experiment
Research question
Does the control layer improve classification accuracy on the scientific answer set?
24 / 24
Minimal prompt · correct
24 / 24
Scientific v2 · correct
24 / 24
+ control layer · correct
Result
No additional classification-accuracy benefit was demonstrated. Accuracy was already at the ceiling before the control layer was added.
Known limitation
A ceiling effect: with every condition at 24 / 24, this pilot cannot show an accuracy difference either way. It says nothing about harder question sets.
Internal experiment
Research question
Does the control layer block deliberately corrupted artifacts while leaving clean ones untouched?
Fault
12 deliberately corrupted artifacts, plus 6 clean counterparts and 2 semantically wrong but structurally valid outputs.
4 / 12
Baseline · blocked
12 / 12
Full controls · blocked
6 / 6
Clean · accepted
2 / 2
Wrong-but-valid · still passed
Result
The full control layer blocked every corrupted artifact and accepted every clean one. Two semantically wrong but structurally valid outputs passed — as expected for an integrity control.
Known limitation
This demonstrates technical containment for the tested artifact failures only. It does not show general semantic correctness, and it does not show general AI safety.
Experiment B, run
Twenty artifacts, one gate.
Switch between baseline processing and the full control layer. The bench replays the recorded counts of the pilot; it does not run a model.
Internal experiment, not independent validation. The counts are the recorded pilot result; which individual artifacts the baseline caught is illustrative.
Reading the pilot honestly
What it shows
Under these conditions, the integrity controls contained every corrupted artifact they were designed to catch, without rejecting clean inputs. The behaviour is deterministic and re-checkable.
What it does not show
No accuracy benefit (ceiling effect), no semantic correctness for well-formed wrong answers, and no independent replication yet. It is one internal pilot, not a general result.