Enterprise AI Bootcamp Demo 5

Research backlog and decision log

4 of the 6 closed hypotheses here failed, and one of them had already shipped before it was tested. A backlog showing only successes is not a record of research, it is a record of selective reporting.
Index built 2026-08-10T07:23:47. 1822 passages from 90 arXiv papers and 12 synthetic internal notes. Embeddings: Snowflake/snowflake-arctic-embed-s (ONNX int8, Apache-2.0), loaded from Demo 1. One embedding stack across both demonstrations. Demo 1 harness live model configured
Retrieval, extraction, every statistical test and the backlog are model-free. Only the synthesis paragraph on screen 2 uses a model, and it degrades to extraction.
Supported
1
Null
2
Contradicted
2
Inapplicable
1

Closed hypotheses

IDHypothesisOutcomeEffect 95% CIpnEffort
HYP-2026-014 liveCross-encoder reranking over a fixed candidate pool improves context recall@5
Executed 2026-08-10T07:34:30 through Demo 1's harness (imported), technical-manual question set. This row is written by the run on screen 4, not stored.
not supported — null result+0.147[-0.005, +0.351]0.220101 hour
HYP-2026-009Semantic chunking will raise context recall@5 by at least 3 points
Index build time rose from 11 to 48 minutes for a difference of one point that the interval cannot separate from zero. Rejected on cost. The published negative result agrees, but note it tested a different intervention: semantic boundary detection, not late chunking. Do not cite this as evidence against late chunking.
Evidence: ENG-2503-011 (synthetic internal note), arXiv:2410.13070
null+0.010[-0.040, +0.060]0.6201402 days (re-index and re-score)
HYP-2026-011Carrying the section heading and equipment variant into each chunk will raise recall
The cheapest change in the backlog and the largest effect. It was proposed third because it is unglamorous. Worth saying out loud: the ranking of ideas by expected excitement was inversely related to their measured value.
Evidence: ENG-2509-014 (synthetic internal note)
supported+0.080[+0.030, +0.130]0.0041404 hours (metadata change, no re-embedding of source text)
HYP-2026-006A larger embedding model will raise recall enough to justify the latency
335M and 570M parameter models were within each other's intervals while costing 3x and 5x the latency of the 110M model. The curve flattens; the latency does not. Standardised on the 110M class.
Evidence: ENG-2510-006 (synthetic internal note), arXiv:2210.07316
null+0.010[-0.050, +0.070]0.7101401 day
HYP-2026-002Retrieving 40 passages instead of 5 will raise answer acceptability
Median latency went from 2.1 s to 9.4 s, so it failed the 3 s constraint outright and the accuracy question became moot. Recorded as inapplicable rather than negative: on a GPU serving path the answer could differ, and filing it as a negative result would mislead whoever reads this next.
Evidence: ENG-2602-013 (synthetic internal note), arXiv:2307.03172
inapplicable1401 day
HYP-2025-021An LLM judge can replace human review for the release gate
Judge-human kappa 0.47 against human-human 0.61, and reversing presentation order flipped the judge's preference on 19 of 120 pairs. The judge stayed as a screening filter and the release gate stayed on deterministic checks. This is the item that was already half-shipped when it was tested.
Evidence: ENG-2507-009 (synthetic internal note), arXiv:2305.17926, arXiv:2306.05685
contradicted-0.140[-0.240, -0.040]0.0081203 days including human annotation
HYP-2025-018Our reported context recall of 0.81 is real
Gold passage IDs had been resolved against the improved index and reused to score the baseline, which could not retrieve IDs that do not exist in it. The true figure was 0.68. This is why `leakage_checks` is a mandatory field on the card rather than a suggestion, and why the number was wrong in the direction we hoped.
Evidence: ENG-2511-021 (synthetic internal note), arXiv:2207.07048
contradicted-0.1301402 days of investigation

Every stored row's numbers come from the synthetic internal notes, which are the only invented data in this demonstration and are labelled as such throughout. The row marked live is written by the execution on screen 4 and is absent until that run has happened.

Correcting for running many hypotheses

IDHypothesisp Significant at 0.05Survives BH at FDR 0.05
HYP-2026-009Semantic chunking will raise context recall@5 by at least 3 points0.6200nono
HYP-2026-011Carrying the section heading and equipment variant into each chunk wil0.0040yesyes
HYP-2026-006A larger embedding model will raise recall enough to justify the laten0.7100nono
HYP-2025-021An LLM judge can replace human review for the release gate0.0080yesyes

Benjamini-Hochberg is used rather than Bonferroni because these hypotheses are correlated configurations of the same system, and Bonferroni is over-conservative there. The correction is applied across the whole backlog, not per experiment.

Open, cheapest first

IDHypothesisEffortTest
HYP-2026-015Fixing the candidate pool at 10 rather than 20 costs no recall and buys 150 ms
Directly downstream of whatever HYP-2026-014 returns. If reranking does not help at 20 it certainly will not help at 10, and this item is dropped.
Evidence: arXiv:2411.11767, ENG-2411-017 (synthetic internal note)
1 hourpaired bootstrap; McNemar on hit@5
HYP-2026-016Late chunking raises recall on passages whose meaning depends on earlier context
Note the subset must be fixed in advance. Choosing the subset after seeing the results is how a null becomes a headline.
Evidence: arXiv:2409.04701, arXiv:2312.06648
1 daypaired bootstrap on the subset, with the subset defined before the run
HYP-2026-017Answer quality is unchanged when the judge is calibrated against 50 human labels
The 2026 literature turns LLM-as-judge from a method into a measurement problem. This item is what it would cost us to adopt that properly.
Evidence: arXiv:2601.20913, arXiv:2605.16354, ENG-2507-009 (synthetic)
3 days including annotationcalibrated judge with variance correction, per arXiv:2601.20913

Decision log

Nygard's five sections, with MADR's Confirmation field. A decision without a confirmation clause is an opinion with a date on it.

ADR-014 — Reranking is enabled only at a fixed, small candidate pool
accepted 2026-03-02
Context. The literature disagrees with itself. Cross-encoder reranking is standard practice and reported to cut retrieval failure substantially, while arXiv:2411.11767 finds returns diminish and then go negative as the candidate pool grows, and arXiv:2503.08046 finds the advantage inverts as queries get more conditions. Our own note ENG-2502-004 (synthetic) reports a null on short factoid queries. Reranking latency is also the largest single item in the 3 s budget.
Decision. Reranking is permitted with the candidate pool fixed at 20 and the operating point measured. It is not permitted as a way to widen recall by reranking more candidates. Any claim about reranking must state the candidate count.
Consequences — good
  • The intervention stays inside the latency budget with room for generation.
  • The disagreement in the literature is resolved by naming the condition rather than picking a side.
Consequences — bad
  • Recall lost before the reranker cannot be recovered by it.
  • A future first-stage improvement changes the arithmetic and this must be re-tested.
Confirmation. The hypothesis card HYP-2026-014 executes exactly this comparison through the shared evaluation harness. Re-run it after any index rebuild.
ADR-011 — The release gate stays on deterministic checks, not a judge model
accepted 2025-08-04
Context. We wanted to replace manual review of assistant answers with an LLM judge. The audit in ENG-2507-009 (synthetic) measured judge-human kappa at 0.47 against human-human 0.61, with a systematic preference for longer answers and a position-order sensitivity that flipped 19 of 120 pairs. arXiv:2305.17926 reports the same failure mode.
Decision. The release gate is: the expected value is present, no forbidden value from another equipment variant appears, and every citation resolves to a supplied passage. The judge is retained as a screening filter over the queue only.
Consequences — good
  • The gate is auditable by a maintenance engineer, satisfying the explainability constraint.
  • The gate cannot silently drift when a model version changes.
Consequences — bad
  • Fluency and helpfulness are not gated at all, only correctness proxies.
  • Manual review remains the bottleneck for anything the deterministic checks miss.
Confirmation. Re-run the agreement audit whenever the judge model changes. A kappa below the human-human figure keeps the gate deterministic.
ADR-009 — GraphRAG is out of scope until indexing capacity changes
accepted 2025-06-10
Context. GraphRAG (arXiv:2404.16130) answers global questions our current system cannot. Its index construction runs a language model over every chunk and summarises every community. Costed against our corpus in ENG-2506-002 (synthetic) as a multi-day CPU job, re-run weekly as revisions land, with query-time map-reduce measured at 6 to 11 s against a 3 s budget.
Decision. Rejected under the standing constraint set. Recorded with the costing so it is not re-evaluated from scratch each quarter.
Consequences — good
  • Corpus-level summarisation questions are explicitly out of scope, and users are told so.
  • The rejection has a stated trigger for revisiting rather than being permanent.
Consequences — bad
  • A real capability gap remains open.
  • The costing is specific to our corpus size and goes stale as the corpus grows.
Confirmation. Revisit if GPU capacity becomes available for indexing, or if the corpus shrinks enough that the index job fits in a night.
ADR-007 — Gold passage labels are resolved per retrieval arm
accepted 2025-12-02
Context. Six weeks of reported context recall were inflated because gold passage IDs were resolved against one index's chunk boundaries and reused to score another (ENG-2511-021, synthetic). The error flattered the change being justified.
Decision. Gold passages are resolved per arm from the source document span, stored with the question, and any cross-arm comparison asserts that both arms had a resolvable gold set before reporting a difference.
Consequences — good
  • Cross-arm comparisons became meaningful.
  • The reported baseline dropped from 0.81 to 0.68, which is the true figure.
Consequences — bad
  • Label resolution now runs per arm, so evaluation is slower.
  • Historical numbers before this date are not comparable and are marked as such.
Confirmation. The evaluation refuses to report a cross-arm delta when either arm has an unresolved gold set.