The measurement story

The headline benchmark is a deliberate negative: on the world the distribution ships, FIBER has no measurable advantage over a correctly tuned graph walk or a lexical retriever. Publishing that tie, rather than hiding it, is the product thesis.

1. The reference world: a three-way tie

The reference world (radiogenomic-integrity-demo-v1) contains 761 facts, 750 of them exploratory distractors attached to a single hub. FIBER compiles the query down to 11 facts (1.45% of the world), and the deterministic oracle returns the correct verdict — invalid, with all four leakage witnesses (identity, preprocessing, site, temporal).

Under equal tuning, so do two baselines. A 5-hop walk of the incidence graph and a BM25 retriever at k=11 select exactly the same eleven facts — not the same count, the identical set. On this world the cheapest admissible strategy is the graph walk, not the compiler.

Sound means the strategy's selection, fed to the same deterministic oracle, reproduces the full-context verdict with the same four witnesses. Closure is the fraction of the query's protected facts retained. Admissible is both.

2. Why publish a tie

Without competent comparators, “FIBER compiles a smaller context” is unfalsifiable. The workspace therefore ships an equal-engineering baseline panel (bioprism-baseline: full-context, k-hop incidence, connected component, lexical top-k, query-graph) and reports whatever the panel says. The distribution's own earlier comparison script measured the graph baseline only at depth 7 and unbounded — the two settings where it returns everything — and reported a 69× advantage that vanishes under equal tuning. The project treats that as a strawman and corrects it, per its own blueprint requirements for matched, equal-engineering comparisons.

The tie is information: it says the reference world cannot tell these methods apart, which is a property of the benchmark, not of the methods.

3. A world that discriminates

bioprism-worldgen makes the world's structure a parameter and builds one that can separate the methods: distractors attached near the target instead of at a hub leaf, decisive facts behind a relay chain, and distractor tags camouflaged to tokenise into the protected vocabulary.

Three distinct failure modes appear:

• The graph walk has no usable depth. Depths 5–10 pull in 98% of the world and still miss every decisive witness; depth 11 is the first sound setting, and by then it has taken essentially everything.
• BM25 is right by luck. It reaches the correct verdict from a 91% protected closure — it dropped a protected fact that happened not to matter, and raising k to 50 never recovers it. This is why the harness ranks on admissibility rather than verdict alone: ranking on verdict would have crowned a strategy that violated the mandatory closure and got away with it.
• FIBER is the cheapest admissible strategy — right verdict and full closure at 11 facts. Full-context and graph-11-hop are also admissible, at 762 and 761 facts respectively.

4. Three-implementation parity

Certificates hash canonical bytes, so a certificate produced by one implementation must be replayable by another. Three implementations — the CPython reference runtime, the eager Rust path, and the Rust indexed store — agree byte-for-byte on the reference compilation:

certificate_sha256  c0da17ffc80465258345c8a538171bfd868100cd883e9a20780a0dc5477e7ea4

The numbers on this page come from the comparison harness documented in docs/DISCRIMINATING_COMPARISON.md and docs/FINDINGS.md; the reference-world certificate digest is pinned by the parity tests. Full reports: FINDINGS.md, BASELINE_COMPARISON.md, DISCRIMINATING_COMPARISON.md.