Why your agent needs two kinds of time
Correction probes: the question class that snapshot and RAG systems cannot express — and what happened when we measured it. Baselines score zero at every model size; the bi-temporal system doesn’t care how big the model is, because the model isn’t doing the temporal work.
bi-temporalbenchmarks #2Statically valid, semantically doomed
A 7B model met our operator schemas: perfectly valid plans, zero successful executions. Models invent output paths no input schema constrains. On output contracts — the one-line class of check that took execution success from 0.00 to 1.00 — and why tool manuals must be checkable, not just readable.
small modelstool contracts #3Your verifier believes correct arithmetic on incomplete evidence
A live demo caught our claim verifier approving a count computed over a truncated result page — the value matched, the evidence didn’t. On truncation taint, support as a tuple instead of a boolean, and the ablation where all fifteen incomplete-evidence claims sail through untouched.
verificationwar story #4What happens when you give the same database a model 10× bigger?
We froze everything except the model and scaled 7B→72B. The operator-backed advantage grew 4.6× while baselines stayed flat; belief-state probes hit a perfect 1.000 at 32B. The surprises: AWQ quantization — not scale — is the planning bottleneck, and giving vector-RAG its intended k=20 context made it worse.
scale studyquantizationfair baselines #5We asked four students to write 110 questions. Our system could express 10.
An independent-question study with the operator list withheld: grouped aggregation is the missing capability that matters (30 questions blocked by it alone), writers conflate the two clocks, and an equal-information SQL baseline ties us on flat schemas while losing by +0.124 where computation composes.
coveragefair baselinesstudy