# PRE-REGISTRATION — what discrimination does a corrector need to pay for itself? Written 29 September 2026, before the run. **The Price of Correction + The Strand**, sharing c, ε. ## Where this came from I claimed to him that entries 2, 20 and 37 all reach *"a checker that cannot tell the original from the copy is **worse** than no checker"*. The website session corrected me: only 37 reaches it. Entry 2's discrimination collapses to **zero** at full dependence — worthless, not harmful — and entry 20's self-set reference carries **zero bits**, also nothing. I filed the gap in the work queue rather than assert a bridge. **This is the bridge, and it runs through the cost channel, not the discrimination channel.** ## The two rooms **The Price of Correction** (Landauer 1961, measured 2012): *"each corrected error exports at least kT ln 2 of heat; a hand-off costs no minimum work."* The floor is on **correction alone**. **The Strand** (methyl-directed strand discrimination, 1983): `c_eff = m(2s − 1)`. ## The derivation being tested **The price is paid per action. The benefit accrues per correct action.** A corrector that examines a fraction m of errors pays the floor on every correction it performs, whether or not it corrected the right strand — an erasure is an erasure. But it delivers only the **net** `m(2s − 1)`. Writing ρ for the price of one correction in units of what one corrected error is worth: net per error-opportunity = m(2s − 1) − m·ρ = m[(2s − 1) − ρ] break-even: s* = ½ + ρ/2 **Three consequences, none of them in either room:** 1. At ρ = 0 this is entry 37's threshold of exactly ½ — so entry 37 is the **free-correction** case. 2. For any ρ > 0 the threshold is **strictly above ½**, so a corrector exactly at a coin is **worse than none** — it pays and returns nothing. **That answers the queued candidate by derivation.** 3. At ρ ≥ 1 **no s ≤ 1 breaks even.** There is a price above which no corrector, however good, pays for itself. ## Criterion, fixed before the run 1. **The break-even is s* = ½ + ρ/2.** Solving `m[(2s−1) − ρ] = 0` numerically for s over a grid of ρ and m matches ½ + ρ/2 to within **10⁻¹²**, and is **independent of m** (m scales the net, not the root). 2. **It recovers entry 37 at ρ = 0**, giving exactly ½. 3. **The coin is strictly negative for every ρ > 0.** Net at s = ½ equals −mρ, below zero at every combination tested with m > 0, ρ > 0. 4. **The wall at ρ = 1.** For ρ ≥ 1 no s in [0, 1] gives a positive net; for ρ < 1 the required s is in (½, 1). Checked over 200 values of ρ. Tolerances are 10⁻¹² because legs 1–3 are algebra; the only error is floating point. Said in advance, since six criteria tonight were numbers I chose without asking what precision the quantity has. **WHAT KILLS IT:** leg 3 failing would mean the coin is not strictly costly and the bridge does not exist — in which case the queued candidate stays open and the peer's narrower wording stands as final. **Prediction:** all four hold. ## ⛔ What must not be overclaimed, stated before the numbers exist **ρ is not measured here.** Mapping the thermodynamic floor onto "what a corrected error is worth" needs a value-per-correction that the house does not fix, and entry 23 already caught me conflating two different ratios (the measured information cost is 1.02× the floor, which is **not** a value ratio). **So ρ = 1 is a structural wall, not a number about the world**, and the write-up says so. What is established is the **shape**: the threshold rises with the price, and there is a price beyond which nothing works. ## Discipline No prose conclusion in the script. Tables printed by the run. Hashes read from the file.