SX — Results (pilot, n=2 per cell, claude-sonnet-4-5, S0/S1/S2)

Two behaviorally-identical twins (both pass the strict RealWorld Hurl oracle 13/13): a clean/GS-disciplined Conduit backend (lean) and a calibrated “average real project” mirror (chaotic; 8 documented degradation dimensions, each cited — calibration.json). Task: add a computed readingTime field to every article response (chaotic has the article serialization duplicated across 5 LIVE copies; lean has it centralized in one place). Agentic run via claude -p; behavior verified by the oracle, coherence by a 6-path probe (did readingTime appear correctly on all live paths). Sentinel = the authored CLAUDE.md navigation map. n=2 per cell. Mechanism demonstration, single benchmark, one frontier model — not a powered effect size.

The ablation — total tokens to complete the change (both reps; mean)

| Condition | lean | chaotic | chaotic / lean | |—|—|—|—| | S0 — no sentinel | 802k / 2,104k (mean 1,453k) high variance | 3,183k / 2,054k (mean 2,619k) | ~1.8-4x (noisy) | | S1 — sentinel on lean only (realistic GS-vs-average) | 298k / 404k (mean 351k) | 2,394k / 1,956k (mean 2,175k) | ~6.2x | | S2 — sentinel on both (navigation equalized) | 299k / 305k (mean 302k) | 893k / 569k (mean 731k) | ~2.4x |

Edit count (writes), the cleanest signal (low noise): lean = 2 in every sentinel condition; chaotic = 11-12 (S0) -> 10-11 (S1) -> 7 / 7 (S2). Coherence 6/6 in all 12 runs; oracle 13/13 throughout (behavior preserved). Localization tokens collapse with the map for both twins (chaotic ~590k without -> ~210k with, at S2).

Findings

  1. The sentinel fixes the SEARCH/localization cost, for both twins (localization ~590k -> ~210k on chaotic with the map; lean ~350k -> ~205k). Authored structure, consulted by lookup — the KX result. This is why query-time semantic search (CodeSeeker-style) is superseded: the value is the authored map.
  2. The sentinel does NOT fix the SURFACE/reconciliation cost. With a map on both (S2), chaotic still costs ~2.4x the tokens and 3.5x the edits (7 vs 2) — because the change must hit 5 live duplicated copies, which no navigation removes. Only ordering the code removes it. Stable across both reps (7 writes both).
  3. The sentinel reduces VARIANCE, not just the mean. With a map, lean is cheap AND consistent (S1/S2: 2 edits, ~300-400k, both reps). Without a map (S0), even the CLEAN twin is high-variance: rep0 = 2 edits/802k, rep1 = 11 edits/2,104k (the agent wandered). No authored map = unpredictable navigation even on clean code.
  4. The chaotic twin is expensive in every condition; coherence held (6/6) because the FRONTIER model pays rather than fails. The coherence failure (a missed copy -> broken output) is the weak-model story (future work; TX already shows weak-model truncation).

Why this matters (resolves the TX boundary; vindicates the original intuition)

TX (“structure alone doesn’t cheapen reading; the sentinel does”) used un-bloated twins, so it could only see the navigation axis. SX separates the two axes on a genuinely average-degraded twin:

  • Search cost (blind pattern-hunting) -> fixed by the authored sentinel. [field phenomenon #1]
  • Surface cost (reconciling N duplicated copies) -> NOT fixed by a map; requires refactoring; the stable S2 residual (~2.4x tokens, 3.5x edits), on a FRONTIER model. [field phenomenon #2] Honest headline numbers: a GS project with its map vs an average chaotic project without one is ~6x on the same change (S1); even map-for-map, ordered code is ~2.4x cheaper and needs a third the edits (S2).

Scope / honesty

n=2 (S0 is high-variance and would need more reps to pin a mean; the sentinel cells S1/S2 are stable and carry the load-bearing claims). Single benchmark (Conduit), one frontier model (claude-sonnet-4-5). Weak-model (Ollama, agentic) arm is future work and is where a coherence FAILURE, not just extra cost, is expected. All per-run data in runs/ (*.json, *.raw, *.probe.json, *.hurl.log); aggregate CSVs probe_stdout.csv, s1_stdout.csv, s2_stdout.csv, reps_extra.csv.