# Lab 012 — Raw lab detail: Buy the harness, not the tier Working notes behind the published entry. This lab's findings died and were replaced several times on the way to publication; the deaths are recorded here because the gate is what caught every one of them. ## The one question Lab 011 ended with a held-out run it could not attribute: a top model inside Claude Code cleared all 22 misses, but that swapped the harness and the model at once. This lab holds Claude Code fixed and moves the model, from a free local Gemma to two generations of Opus, and prices every rung an architect could actually buy. ## The instrument The 22 tasks Lab 011's free local loop first missed, each certified by a reproduce-or-drop gate. Fourteen arms across two apparatuses: the constrained repair loop, and Claude Code driven headless through a LiteLLM translation proxy. Three arms replicated. Cost capture is the method contribution: a dedicated API key isolating one arm's dashboard line, and per-task usage envelopes on the native Anthropic path. No apportioned dollar figure survived to publication. ## The measured spine - Free local 31B loop: 17 and 17 of 22 across two runs. Paid mini in the same loop: 16 and 17. Dead parity; the tier bought no separation in the same apparatus. - gpt-5.4-mini in Claude Code: 22 of 22, twice, $3.14 measured. Opus 4.8: 22 of 22, $13.79. Opus 5: 22 of 22, $12.56. - Token budgets across that fourfold price range: 10.38M, 10.96M, 9.66M. A 13 percent spread against the mini's own 24 percent run-to-run variance. The harness sets the token budget; the tier multiplies the rate card. - Escalating only the five the gate named: $1.12. The residual five are 23 percent of tasks and roughly half the bill; hard tasks are expensive in proportion to being hard. - Instrument literacy gates substitution: Gemma 31B went 17 in the loop to 11 in the harness. The harness multiplies a model that can operate its instruments and taxes one that cannot. ## Findings that died in measurement (kept deliberately) 1. Token efficiency. The draft's original headline had Opus clearing the pool in a fraction of the mini's tokens, built on a hand-recorded 1.28M from Lab 011. Real usage capture: 10.96M, off by 8.6x. The thesis dissolved; the models consume within noise of each other. 2. The latency tail. One 26-minute outlier carried a "the premium buys predictability" claim. The replicate maxed at 4 minutes. Run variance, not model property. 3. The cost direction itself flipped three times before both sides were measured on dedicated meters. Token fixation is named as a trap by the AI Factory Economics framework this lab tests; the first draft committed it, and the correction ships in the writeup. ## The procurement rulings Renting the available GPU class to self-host the escalation rung loses on cost and wall clock at once: agentic sessions park context in VRAM while barely touching compute, so concurrency is bought with memory and the meter scales linearly with the work. Per stream per hour, the owned box holds most of a rented fleet for nothing. Scale out with the smallest instance, never up; for this rung, do not scale at all. Rent the tokens. ## The ladder, and its two directions Spend committed capacity first at every rung it exists: owned idle hardware, unused cloud-spend commitment, unused subscription headroom. One category, three altitudes. Escalate what the gate hands back to the cheapest instrument-literate model. Route overflow back down when the budget binds instead of the clock; the gate makes both directions safe because a verified repair is verified regardless of which rung produced it. ## Operational record (mistakes included) - A first Opus attempt died at task 14 on credit exhaustion: $5.45 of errored sessions, deleted and re-run. Check the resource a metered run spends against before launching. - A g6e.xlarge idled 12 hours 42 minutes (~$24) because a stale note stood in for a check expired credentials could not perform. "Unverified" is the only honest answer when the console is unreachable. - A one-task probe of Opus 5 extrapolated to a $19-20 arm; the arm cost $12.56. Probes establish pricing and nothing else. - The mistakes stay in the published bill. ## Spend About $121 total: $73.75 of rented infrastructure and $47 and change in metered tokens, including the waste itemized above. Both 22-of-22 configurations landed in single digits of dollars; the instrument that told them apart cost roughly one more arm on top. ## What this hands labs 013 and 014 Two questions this lab could not answer with a saturated pool. The self-hosted floor: what is the entry-level open-weight model that clears the instrument-literacy bar Gemma failed. And the tier boundary: mini and Opus tied at 22 of 22 here, so the difficulty line where the 4x rate card starts buying solves has to be found on a harder pool, mined for it. The task difficulty rating built from this lab's fourteen arms is the selection instrument for both.