v2.3 — the coherence pass

What this release is

v2.2 hardened the instruments. This release hardens the argument around them. A systematic audit — self-coherence, local coherence, and global coherence of every formula against every other — produced a severity-ranked queue of 125 findings. Every one is now remediated. No claim moved; several claims became true.

The two mathematical errors

The J→kWh dimensional error (Part IV). The published price model read c = e × p_energy + t95 × p_opp, multiplying Joules directly by a USD/kWh price — a 3.6-million-fold overstatement of the energy term, present since the model was written and made load-bearing by v2.2, which pinned cost percentiles to it. The derivation is now per-run and dimensionally sound: c_run = (e_run / 3.6×10⁶) × p_energy + t_run × p_opp, with cost percentiles computed as percentiles of per-run cost, never recomputed from percentile inputs. Appendix A’s bridge paragraph and the r(W) asymmetry ratio now consume the corrected model.

The WNG revaluation artifact, stated backwards (Part V). The wrapper–native growth gap claimed an unadjusted WNG rises when the asset’s price falls — a valuation artifact. The artifact is real but runs the other way: E^wrapper is a stock revalued with every tick while U^native accumulates at transaction prices, so a price rise can lift measured exposure growth on revaluation that no usage produced. The aggregation spec now unit-denominates both sides (revaluation stripped by construction), pins the trailing four-quarter window, and separates this deflator from Red Line 14-B’s workload unit-price deflator — same word, different deflators, both published.

Conventions the thesis now pins once

• Q^RD owns α only. The return-decoupled flow consumes the regression intercept; the residual u is published as its own series and absorbed by nothing. The residual was previously folded into Q^RD at one site and excluded at another — the same paragraph disagreed with itself across two pages.
• Fees: gross vs. retained. The take-rate τ uses gross fees (what buyers pay); FeeCoverage uses retained fees (net of burn, what funds security). Burns are supply arithmetic — never revenue, never buyer evidence. Every “fee+burn” sum in a ratio has been dissolved; the historical label is retired with a footnote at the 30% phase gate.
• WDR is a pair. Stock share and flow ratio, reported side by side, never divided. Red Line 9 now requires both components to rise.
• Two distinct deflators. WNG unit-denominates (strips asset-price revaluation); RL14-B deflates by the workload unit-price index (strips compute-price deflation). The WNG block states the difference explicitly.

Red-line repairs

• RL14 Condition B now has a persistence requirement (two consecutive quarters), a named deflator, and an explicit “fails to grow” definition.
• RL9 specifies the pair reading and reads on the unit-denominated WNG.
• RL11 and RL13 thresholds are instantiated (50% enclosure; 50% curtailment jurisdiction share) instead of “e.g.” placeholders. The audit’s rule: examples cannot guard falsification conditions.
• RL15 clauses are labeled conjunctive, with the reasoning stated.
• Severity labels (“kills” vs. “weakens”) now carry defined meanings, stated once at the head of the red-lines chapter.
• The response-protocol clocks (14-day publication, 90-day remediation) have pinned start triggers and a no-reset rule.

Instrument and schema repairs

• VerifyPower now carries the token price explicitly: q_fee(W*) · P_X / p50,c(W*) — a fixed token fee quantity times the USD price, over the cost percentile. The old form left the numeraire ambiguous.
• AffordabilityRatio guards are specified as three concrete mechanisms (component disclosure, USD unit convention, monotonicity rule) instead of an undefined “floor.”
• r(W) is pinned to a single numeraire (USD via the corrected price model), with the perverse-direction warning: a ratio improved by expensive production is a cost problem wearing an asymmetry credential.
• The canonical VerifyPrice tuple is reconciled across all five sites that define or view it. The starter-set targets vs. constitutional floors are disambiguated; the duplicate heading in Part IV is fixed; SLA latency is scoped prove-side; the blanket 5-second rule defers to registry entries for workload variants.
• Decision rule (Part III): acceptance of a benign-state resilience premium now gates on the worst disruption state’s per-state bound first, with the π-weighted average demoted to tie-breaker — the average can no longer pass what the tail fails.

Also fixed

Definitions repaired across Parts I–VI: the Value Capture Lemma’s failure clause (no monetary premium, not no value at all); the nine-requirements enumeration; the Work-Credit supply identity (lost wallets stay in outstanding, exit effective supply); the θ_t double-count in the price-impact equation; the φ_t “restated lemma” false equivalence; RCR and NUS denominators; CSR’s relationship to dollar-weighted IRR; the collateral level/slope vocabulary; the unlevered-WRR publication rule in the executive memo. Placeholder caps (40% chip-family, 10% reachability) instantiated. Vignette arithmetic annotated as illustrative. The base-rate defense is restated in its honest two-move form: empty precedent rows are evidence against; the change-of-kind argument is why that evidence is not dispositive.

Full details: Appendix J (release history) and the v2.3 CHANGELOG entry.