The Revision Ledger
The standalone changelog site for the MFid methodology and the Software Defined Index. Every formula change, rename, and correction is recorded here — including the findings we file against ourselves — before it takes effect on the main site.
Historical entries retain the dimension names in force at the time they were written.
Complete revision history
- v2.6.1 (2026-08-05). Line-by-line site audit, filed as findings against ourselves. (1) Walkthrough arithmetic corrected: the I cube root was misprinted ((0.7429)1/3 is 0.906, not 0.905), so the illustrative I is 0.91, not 0.90; the Step-4 product was also misprinted. The illustrative aggregate moves 0.907 → 0.908; the reported score, MFid 0.91, is unchanged. Walkthrough to v1.7. (2) Theory’s “diagnostic in reverse” corrected: the prior text claimed an aggregate of 0.84 or below guarantees a dimension at or below 0.5 — the implication runs the other way. The correct bound: the weakest dimension lies between MFid4 and MFid. Theory to v1.7. (3) Porsche case corrected: the prior text attributed pre-clip raw ratios above 1.0 to D and I, which are constructed never to exceed 1 (one is 1 minus a clipped penalty; the other a geometric mean of fractions). Super-unity raw ratios live in the claimed-vs-observed performance ratios that feed E and the per-SLI fidelity terms; D and I register the understatement as near-perfect scores instead. (4) D rubric now documents the aggregate-claim treatment the walkthrough and intake already use: a claim that binds on aggregate (an uptime percentage) enters the SLI minimum as its raw fidelity ratio, clipped to 1; the dispersion-plus-offset form applies to claims that bind per observation. Form-2’s D roll-up bullet aligned to the same term. (5) Theory’s “behavior outside the claim” limitation reconciled with the Itrace rubric: untraceable activity is penalized as a coverage gap; what MFid cannot judge is the unclaimed behavior itself. (6) Housekeeping: one stale “deterministic” in the D rubric renamed per v2.6.0; a stale methodology-version reference removed from /status; the homepage Porsche framing reconciled with the clip-and-annotate rule. No formulas, weights, or tier definitions changed.
- v2.6.0 (2026-08-05). Two dimensions renamed; letters, symbols, formulas, weights, and scores are unchanged. Determinism → Dependability: since the v2.4.0 offset term, D scores accuracy-to-claim plus dispersion — a stochastic system operating inside tolerance can score high D, so “Determinism” promised a stricter property than the formula measures. “Dependability” states what is actually scored: whether the buyer can depend on any single future observation matching the claim. Intentionality → Integrity: identical rubric (Ispec, Itrace, Idrift), clearer name — the fraction of behavior covered by, and faithful to, the stated specification. Efficiency and Observability retain their names; E’s definition now states explicitly that it measures cost fidelity, not cost goodness. Historical change-log entries retain the names in force at the time they were written.
- v2.5.0 (2026-08-04). D’s worst-case tail rule smoothed. The prior rule — any single >2σ tail event on a critical SLI caps D at a flat 0.9 — was discontinuous (a 1.99σ event and a 2.01σ event scored a cliff apart) and non-monotone in severity (a 6σ event drew the same 0.9 as a 2.1σ event). Replaced with a continuous cap:
D ≤ min(1, 2/z), z = worst single-observation excursion in σ units. Inactive for z ≤ 2, matches the old cap’s magnitude near z ≈ 2.2, and strictly decreasing beyond it. No illustrative numbers moved (the walkthrough window has z ≤ 2). Theory revised to v1.5 and Walkthrough to v1.5 to match. - v2.4.0 (2026-08-04). Four changes from a mathematical review of the framework. (1) [0,1] is now a hard invariant. The v2.2.0 provision permitting published scores above 1.0 for verified underreporters is withdrawn. The Verified Underreporter designation remains — as an annotation carrying the raw pre-clip ratios — but every published MFid, and therefore every SDI, is clipped to [0,1] with no exceptions. A bounded index that is sometimes unbounded is two definitions under one name. (2) Domain projections (Form 2) changed from an arithmetic weighted sum to a weighted geometric mean, closing an inconsistency: compensation was forbidden between dimensions but permitted within them. (3) D — Determinism gains an unfavorable-mean-offset term Δ:
D = 1 − min(1, (Δ + σ) / (μc × τ)). Dispersion-only scoring rewarded systems that are perfectly consistent at the wrong value; the new form reduces to the prior one whenever the claim is met on average, so no published illustrative numbers moved. (4) O — Observability rubric now defends its own construction: why coverage multiplies into the score rather than becoming an uncertainty band. Theory revised to v1.4 and Walkthrough to v1.4 to match. - v2.3.0 (2026-08-04). Introduced the term Software Defined Index (SDI) as an equal terminology for the canonical aggregate:
SDI(x) ≡ MFid(x)by definition. No new formula, dimensions, weights, evidence tiers, or reporting form. Motivation: the firm’s name poses the question “how software defined is X”; SDI names the answer. Any future divergence between SDI and MFid requires an entry in this change log before a diverged score is published. - v2.2.1 (2026-08-04). Consistency pass, filed as findings against ourselves. (1) Removed two references to a not-yet-public external repository — a published claim must be verifiable, and those were not. (2) Manifesto dimension formulas reconciled to the canonical rubrics on this page (the prior manifesto forms lacked the tolerance term, the clips, and the three-part Intentionality rubric). (3) Theory-page worked example corrected: the geometric mean of 0.9, 0.9, 0.9, 0.1 is 0.520, previously misprinted as 0.516. (4) Walkthrough revised to v1.3: an O rounding inconsistency (0.85 in headings vs. 0.83 in the aggregate) and internally inconsistent illustrative latency statistics fixed. (5) SLO-to-dimension mapping harmonized across Theory and Compare to match the Form-2 roll-up on this page. (6) Added dimension-interaction analysis (Theory) and cross-reference here. No tier definitions changed; no published scores moved.
- v2.2.0 (2026-06-18). Added “When MFid exceeds 1.0 — the honest underreporter” section. Formalizes the conditions under which the
min(1, …)clip is lifted on individual dimensions and the aggregate composite is permitted to exceed 1.0. Introduces the “Verified Underreporter” designation, the four-condition gate, and Porsche as the canonical reference case. Practical upper bound of ~1.15–1.25 documented. E — Efficiency section updated to cross-reference the new section rather than footnote the concept inline. No existing tier definitions or published scores changed. - v2.1.1 (2026-05-29). Added “Constraint on public-facing targets” section formalizing that public teardowns and walkthroughs require publicly verifiable target claims; private contracted SLAs are out of scope for public artifacts. Surfaced during the Cloudflare walkthrough fact-check incident (see walkthrough v1.1 revision note). No tier definitions or scores changed.
- v2.1 (2026-05-28). Evidence-tier renumbering: Tier 1 is now Measured Reality (was Tier 3), Tier 2 unchanged (Published Specification), Tier 3 is now Scientific Calculation (was Tier 1). Sub-label for incomplete measurement moves with the tier and is now 1E (was 3E). Three categories and their definitions are unchanged. No published scores moved as a result. Reason: in ordinary English “Tier 1” reads as strongest evidence; the prior numbering inverted reader intuition. Limitations section expanded to name three scope gaps the instrument does not cover (no-spec exposures, in-spec compromise, novelty/category-creation).
- v2.0 (2026-05-27). First reconciled methodology page. v1.x had published two formulas without reconciling them and used two names for the third evidence tier; v2.0 fixed both, published the Intentionality rubric, and added the coverage-label requirement.