forced structural injection · v1

protocol summary

strict1/2 degraded under anchor pressure. strict3 partially recovered.
harness remained stable. instability appears target-dependent.

run result

chat_patch failed structural resistance under strict1 and strict2 injection modes, with partial recovery under strict3. The harness scored stable across all three modes, confirming the instability is localized to the target file's structural risk sites, not to the evaluation protocol itself.

target strict1 strict2 strict3

chat_patch unstable unstable partial-pass

harness stable stable stable

variance gap

+0.7758

target instability vs harness stability

anchor penalty

−0.6667

resistance drop under fake-locus injection

strict3 recovery

partial

recovered where strict1/2 degraded

run detail

structural_resistance (no-locus)

1.00

structural_resistance (with-locus)

0.33

explicit_mismatch_reject

0.70

injection_override_awareness

0.60

fake_locus_reject rate

0.33

anchor_penalty (chat_patch)

2.3333

anchor_penalty (harness)

1.3333

artifacts generated

13 json · 13 md

503 recovery

missing combinations retried and restored

reproduce

python tools/forced_structural_injection_test.py --focused --b-mode strict3 --repeat 3

model-agnostic protocol · swap API target to run against any structured reasoning model