We ran the experiments
because nobody else was.

The chain

One question, asked five ways.
Each answer changed the product.

Nothing here was research for its own sake. Each result arrived as a constraint, and the constraint became code.

1. Density and structure independently raise capability. Dense, curated records outperform sparse ones — and sparse-plus-unstructured is uniquely destructive. → The founding premise: your own structured markdown is the source of truth, not a scraped index.
2. Retrieval-time elicitation is real on a frozen model. Curated context measurably improves a model whose weights never change (p≈0.004). But only a fraction of the available gain was being captured. → Findability was the bottleneck, so retrieval became hybrid vector + keyword, fused by RRF.
3. Curation — not retrieval — is the lever. The same data flips conclusions depending on the metric; a popular richness score turned out to be 40% word count. Length-controlled, the gradient is clean: curation ≫ retrieval ≫ nothing. → Became the standing principle the whole system is organised around.
4. Grounding flips fabrication into abstention. Handed a real record that does not contain the answer, a grounded model abstains instead of inventing. The same test also showed a grounded model relays poisoned memory perfectly — so the substrate has to be tamper-evident, not merely present. → Became the requirements spec LBrain was built to satisfy.
5. We A/B-tested our own ranking layers, in four query regimes. Two features earned their place. Two did not: one gave zero lift, and one displaced the verbatim source it was supposed to summarise. The all-features-on configuration was the worst almost everywhere. → We removed them. The package got smaller and better — 4,575 → 3,941 lines — and reranking became a call-when-needed option instead of an always-on cost.
6. A real, relevant, adjacent record can be worse than no record at all. Authentic material that doesn't answer the question raises misattribution — the model borrows true facts and silently reassigns their provenance. Then a control we added under review partly falsified our own title, and the better idea fell out of it: trustworthiness of a source is not sufficiency of its evidence. → Became admissibility.py — a deterministic gate that labels what actually answers the question, with no second model call.
7. Blurred attribution collapses how much a model can extract. When provenance melts together, usable value drops sharply — worse under strict grading than we first reported. → Serving was rebuilt around attribution-bound records. Measured before shipping: prose returned 5 of 8 reachable answers, structured returned 8 of 8.
8. Across eight model architectures from seven organisations, the record — not the model — controlled the failure rate. Identical failure ordering in all eight. Architecture explained ≈0% of the variance; record structure swung it by 34.4 percentage points. → The empirical basis for bringing any model you like. Your memory shouldn't be locked inside someone's weights.
9. A record that was true when written, and false three weeks later, nearly reached a legal document. It retrieved correctly, was dated honestly, and carried our strongest trust marker. → Became lbrain stale, and honest date labelling that distinguishes a claim date from a file timestamp.

The discipline gets harder to hold as the window fills — which is exactly why it can't live in anyone's attention.
It has to live in the tooling.

This is ongoing. The instrument gets sharper, and the product follows it.

What we retired

The results that came back against us.

A claim we can't defend is worth less than one we killed ourselves. These were retired by our own controls, red teams and re-analyses — before a reviewer got to them.

Every number here has a hash behind it, including the ones we got wrong.