Two-Context Mode Split: Substitution vs Synthesis under a Token-Cap Sweep on a Production Arabic Sales Router

Measurement record: the two model calls of a production Arabic e-commerce salesrouter — a stateless routing call ("substitution") and a contextual reply call("synthesis") — swept across four completion-token caps (400 / 800 / 1600 /4096) at N=20 per cell. 160/160 calls clean, zero failures, gpt-4o-mini, T=0.2,135,920 prompt + 13,988 completion tokens (≈ $0.03). The prompts are the ones the router actually ships: imported from the runningapplication's prompt builders at execution time, never retyped, and the fullyresolved message payloads are published in this record. The entire run thereforereplays against any OpenAI-compatible endpoint with fetch and nothing else — noapplication source, no database, no framework. This is the production-runtime leg of a three-stack research chain on thesubstitution-vs-synthesis split, joining a sovereign Ollama fixture leg(Converse, Triava Labs, 26b MoE) and a cognitive-middleware leg (Seo,Hashevolution, JAMES at e4b). What this leg adds is the thing a fixture sweepcannot: the same two call shapes measured through a production runtime — persona,store policies, a sales-mode block, a dialect guide, grounding rules and anegotiation contract wrapped around them. HEADLINE RESULTS Confidence grading: ★★★ = re-derivable from the published raw data alone;★★ = mechanism-supported but not ablated; ★ = single-shape observation, do notgeneralize beyond it. ★★★ Substitution is exactly flat. 49 completion tokens with min = median = maxat every cap; one distinct output hash across all 80 calls — byte-identical overa 10× cap range; 0/80 truncated; 1.2% of the largest cap tested. ★★★ Synthesis does not track the cap here. Medians 127.5 / 130.5 / 118.5 /122.5 across the four caps — a 12-token, non-monotonic movement — whilewithin-cell ranges run 48–63 tokens, i.e. run-to-run noise is 4–5× thebetween-cap spread. 80 distinct output hashes out of 80 calls; 0/80 truncated;median output is 3% of the largest cap. ★★★ Nothing truncates anywhere: finish_reason is "stop" on 160/160 calls. Thecaps the application actually ships (160 routing / 200 response) were neverexceeded by any completion — maxima observed: 49 and 159. ★★ Mechanism for the synthesis non-result: the shipped response prompt containsthe instruction "2-4 sentences.", which binds an order of magnitude below everycap swept. A cap can only bind when the model tries to write past it; here itnever is. This BOUNDS the middleware leg's workload-gradient finding — aproperty of synthesis allowed to run to its natural stop — rather than refutingit. A production sales reply is length-instructed by construction, becausemerchants want short replies. ★★ Byte-determinism, mechanism and durability: the routing call's byte-identityis consistent with its shape — a short, schema-fixed json_object completionconcentrating probability mass at low temperature — supported by thecross-shape contrast at identical temperature, not ablated. A replay on ahosted endpoint months later is not guaranteed to reproduce the hashes(provider backends drift); what is expected to reproduce is the flat 49-tokencount. ★ Unlooked-for deployment finding: all 80 benign synthesis replies (80/80) cameback containing Arabic-Indic digits ٠-٩, with no digit anywhere in the customermessage and ASCII-printed prices in the retrieved context. 80/80 quoted acatalog price; an ASCII-only price extractor (the JavaScript \d idiom, whichmatches [0-9] only) returned an empty list for all 80. The trigger is thedeployment's language configuration, not the customer. Measured on one turnshape and one store configuration — a total for this shape, not a traffic-widerate. WHAT IS IN THE DEPOSIT - analysis.md — the full result: both call shapes, the deployment-side findings, five declared protocol deviations, threats to validity, open items.- two-context-sweep-2026-08-12T12-27-16-764Z.json — raw data, all 160 calls: per-call token counts, latency, finish_reason, output hash, full reply text, plus the frozen fixture and resolved prompts.- prompts.json — the frozen, fully resolved prompt payloads for both shapes.- replay.mjs — standalone replay driver, stdlib + fetch, zero dependencies; runs the full grid or any single cell against any OpenAI-compatible endpoint.- check-numerals.mjs — re-derives the numeral-system counts from the raw data alone; no network, no application source.- two-context-sweep.ts — the original driver, kept as provenance: it imports the prompt builders from the running application rather than copying them, which is why the published prompts cannot have drifted from what ships. It does not run outside that application; use replay.mjs.- README.md, LICENSE. METHOD IN BRIEF OpenAI-compatible /chat/completions; gpt-4o-mini; temperature 0.2 pinned tomatch the two prior legs (the application itself passes no temperature — thebyte-level determinism above is a T=0.2 result, not a claim about thedeployment); 2 call shapes × 4 caps × N=20 = 160 calls; frozen fictional storefixture and frozen customer turn so the record reproduces without any database;zero-failure rule — a cell that still fails after retries aborts the run andwrites nothing, because a dataset with holes must never be labelled clean. Inthis run no call needed a retry: attempt counts are recorded per call in theraw JSON, and all 160 calls succeeded on the first attempt. WHAT THIS RECORD DOES NOT DO - It does not claim the workload gradient is absent in general — only that a length-instructed production prompt pre-empts it. The gradient belongs to natural-stop synthesis, measured on the middleware leg.- It does not claim byte-determinism at production temperature. What is expected to survive temperature is the flat token count, not byte equality.- It does not cover a second model family. The within-stack family split is the named open item, not a shipped result.- It does not measure the negotiating reply shape (pricing-authorization and focus-lock blocks attached) — a real production shape left for a later record.- It does not test the shipped cap values (160/200) directly; the shipped-cap conclusion is derived from where the uncapped completions landed.- It does not state traffic-wide rates. One fixture message per call shape; N=20 measures run-to-run variance at fixed input, not variance across customers. The 80/80 numeral observation is a total for this turn shape.- It does not study adversarial inputs. The numeral finding arrived unlooked-for on benign traffic; adversarial treatment of that failure class belongs to a separate record.- It is not a software library. The runnable artifact is a replay script; the original driver is provenance. RELATION TO THE CHAIN Sovereign-stack leg: Converse (Triava Labs),github.com/triavalabs/gemma4-26b-mode-split — substitution deterministic andflat at 26b MoE; synthesis efficiency scaling with model size. Middleware leg:Seo (Hashevolution), DOI 10.5281/zenodo.20363998 with PR #440 and Issue #448 —a seven-tier natural-stop gradient at e4b. The three-way finding itself is to bearchived in a forthcoming joint three-author Zenodo record (Afana / Converse /Seo); this deposit is the production-runtime per-stack input to it. REPLAY OPENAI_API_KEY=sk-... node replay.mjsOPENAI_BASE_URL=http://localhost:11434/v1 SWEEP_MODEL=gemma3:12b OPENAI_API_KEY=x node replay.mjs