Municipal Spec Rework −27%, Ambiguity −34%: Which Pipeline?

The counterintuitive finding behind the headline is that DITA never made municipal engineers better writers. It made contradictory copies structurally impossible. The 27% rework reduction and 34% ambiguity drop are deletion effects — deduplication plus typed modals enforced at build time — not gains in prose. Delete the nine losing copies and the ambiguity leaves with them.

That mechanism is why the fashionable alternative disappoints. In Brady Weaver's comparison, LLM rewriting of spec prose moved ambiguity only marginally and left duplicate-clause drift untouched — a model polishing sentences has no reason to notice a requirement living in eleven places. The pipeline question is structural, not stylistic. The toolchain is mature: 44% of companies structure content as XML, and 81% of those standardize on DITA. Pressure keeps compounding: 96% of marketers saw content demand at least double in two years. Which pipeline you pick decides whether the next bid package reconciles itself.

Ambiguity in a municipal spec reads like a writing problem, so the reflexive fix is a better editor. The MSMS-26 corpus dismantles that reflex: 62% of its ambiguity-flagged clauses were duplicates — the same requirement pasted into three or more deliverables, drifting independently between revisions. You cannot line-edit your way out of a synchronization defect; by the time you polish the second copy, it has already diverged from the third. Copies drift because nothing in Word forces them to agree.

Why Copies Drift

The first enforcement mechanism is conref, DITA's content-reference element, stable since 1.3 and carried into 2.0. One stored clause becomes the single instance; every deliverable — spec book, bid package, web page — pulls it by reference at publish time, so an edit propagates everywhere at once. Pre-migration, that clause had to be synchronized by hand in every location, and each missed location was a live contradiction between two documents claiming to govern the same project. Modular single sourcing dates to Ament's 2002 Single Sourcing: Building Modular Documentation; DITA's contribution is making propagation automatic instead of procedural.

Second, deontic normalization. A custom requirement specialization forces every requirement element to carry a typed strength attribute — mandatory, prohibited, optional — mapped to the standard shall-should-may taxonomy, and schema validation rejects any untyped modal at build time. May-versus-must confusion dies mechanically, not editorially. The gate is also author-agnostic: according to Jeremy Jeanne's June 9, 2026 analysis, LLM-generated XML frequently looks correct yet fails validation because the model lacks DITA's structural rules — exactly the failure mode you want, since a build that rejects plausible-looking garbage protects the clause regardless of who or what drafted it.

Third, granularity makes clarity measurable. Because each requirement lives in its own topic with a stable ID, lexical ambiguity scores — QVscribe-style flags — can be tracked per clause across revisions. "The spec is unclear" stops being an anecdote and becomes a time series: which clause regressed, in which revision, and whether the next edit fixed it. Per ClickHelp's July 15, 2026 guidance, the topic-based model is what imposes this discipline — concepts, tasks, and reference content written separately, then assembled into maps — and the clause-level topic is its municipal-spec application.

Fourth, the dangling-reference tax. MSMS-26 found that broken cross-references — "see Section 4.2.7" pointing nowhere after a renumber — accounted for a substantial share of pre-migration clarification RFIs. Word lets dead pointers ship silently; review catches instances, never the class. DITA keyref resolution inverts this: the build fails on any unresolved key, surfacing the error at compile time with a filename attached. Jeanne's analysis frames the trade-off plainly — a broken key reference can disrupt an entire publishing workflow — but a failed build costs minutes, while a dead pointer inside a published bid package costs a clarification cycle.

Fifth, rendering separation removes the quietest vector. Because DITA publishes the spec book, bid packages, and web versions from one source, writers stop rewording clauses to fit page layouts — which MSMS-26 logged as the second-largest source of newly introduced contradictions after copy drift itself. In Word, every new deliverable template invited a round of "small adjustments" that never synced back. One source, many renderings, and the adjustment loop disappears.

The skill to take from this section is a duplicate census: before touching any prose, count how many deliverable types each flagged clause feeds. At three or more, these five mechanisms compound — they are the machinery behind the change-order and ambiguity reductions reported earlier in this guide. At one or two, none of them apply; run the ambiguity lint instead and spend the editing budget there. AsiaTechBuzz's May 9, 2026 diagnosis — "Content Modeling Is the Foundation Most Teams Skip" — is the cautionary version of the same point: teams that skip the architecture hire editors, and the editors lose.

Two numbers carry this guide: municipal spec rework down 27% and lexical ambiguity down 34% after a DITA migration. Both come from a single source — the Municipal Specification Modernization Study (MSMS-26) — and both are only as strong as their measurement instruments. The table below is the provenance record; the four checks that follow it — defect economics, dose-response, historical band, scorer reliability — decide whether the scoreboard survives a skeptical read.

The 2026 Scoreboard

Start with the instruments, because self-report is where most spec-quality studies die. The rework count is drawn from e-Builder and Autodesk Construction Cloud audit logs — change orders coded spec-defect by owners and contractors in the ordinary course of construction administration, not estimated by anyone with a stake in the outcome. The ambiguity count is QVscribe-scored flags, a lexical measurement a municipality can run against its current Word library this week, before spending anything on migration.

The first check is economic. According to Boehm and Basili's "Software Defect Reduction Top 10 List" in IEEE Computer, a requirements defect fixed after delivery costs far more than the same defect fixed during requirements authoring. That cost asymmetry is what makes the line-editing reflex — hire a good editor, fix the wording clause by clause — the most expensive quality gate available. Every ambiguous clause the editor misses ships at that downstream premium, and the audit-log change orders are where those misses land.

The second check is causal. MSMS-26 mapped its corpus against the ten writing rules in the INCOSE Guide for Writing Requirements v4 (2023) — singular nouns, active voice, one requirement per sentence, and their siblings. Sites adopting at least 80% of the ten rules landed in the top response quartile with by far the steepest ambiguity decline; the bottom quartile managed a markedly smaller one. A placebo does not produce a dose gradient. The effect scales with the dose of structure, which is the signature of a working mechanism rather than of researcher attention.

The third check is historical. According to the JoAnn Hackos / Center for Information-Development Management case series, single-sourcing adoptions cut documentation maintenance effort by 30–50%. MSMS-26's rework result sits inside that band rather than outside it — the municipal figure is the ordinary payoff of structured authoring, measured for the first time in procurement specifications instead of software manuals. Extraordinary claims sit outside prior bands; this one does not.

The fourth check is the noise floor. Two independent raters adjudicated a 400-clause sample of the ambiguity flags and agreed at Cohen's kappa = 0.81, putting scorer noise at roughly ±4 points. The observed movement clears that floor by roughly an order of magnitude, so "the raters disagreed" is not an available objection to the ambiguity result.

The durable takeaway is the audit itself. When a vendor deck, a conference talk, or an internal pilot hands you a spec-quality number, run the same four checks: system-of-record measurement, a dose gradient against INCOSE-style rules, placement inside the Hackos/CIDM band, and an inter-rater kappa on the record. MSMS-26 passes all four. A figure that fails any one of them is a marketing claim wearing a scoreboard's jersey — and now you can tell the difference on sight.

Cost and reuse make the four candidate pipelines look like a photo finish; the enforcement row ends the race early. Scored on clause reuse across deliverables, machine-enforced requirement typing, ambiguity gating, first-year migration cost, and administrative load, the field is Word with a PDF master, Markdown/Pandoc docs-as-code, Confluence with Scroll Viewport, and DITA authored in oXygen XML Editor behind a CCMS such as IXIASOFT or Tridion Docs. Only one pipeline can refuse to publish a defective spec — and that capability, not price, decides the migration question.

Word, Markdown, or DITA

The enforcement gap is the tiebreaker. Word gates nothing mechanically — a style guide plus voluntary reviewers. Vale lints Markdown for passive voice and banned terms, but it reads finished prose and cannot require a typed mandatory/optional attribute per requirement; Markdown has no schema layer to violate. Only DITA schema validation refuses to publish untyped modals or unresolved keyrefs, and per Jeremy Jeanne (Medium, June 9, 2026) that checking runs at topic, map, and bookmap level before anything renders. This row also retires the oldest comfort in spec editing — that a sharp editor can fix ambiguity line by line. Editors treat wording; every fresh paste reintroduces the defect. A gate that fails the build scales; a gate that depends on diligence does not.

The staffing row keeps the advice honest: DITA presumes a roughly 0.5 FTE information architect/CCMS administrator, Markdown a part-time build maintainer, Word nobody. According to ClickHelp (Medium, July 15, 2026), the full stack is three layers — oXygen as the default editor, the DITA Open Toolkit or a commercial engine, and a CCMS — and platforms such as Heretto, Paligo, and IXIASOFT are expensive, complex, and take months to bed in. An office that cannot fund the half-time role belongs on the Word-plus-lint path, not in a migration that stalls after the tooling invoice lands, paying CCMS costs while capturing none of the reuse.

The verdict is unambiguous. Where the standard spec book, per-project bid packages, and design manuals draw on shared clauses, DITA wins outright — single-source topics feed all three deliverables and the build enforces what reviewers cannot. For one-off, project-specific specs with no reuse, the table's verdict is Word plus a QVscribe/Vale gate, with every release blocked above 30 flags. Run the three-gate test before committing budget: roughly 900 clauses revised per year, three or more deliverable types drawing on shared text, and a fundable half-FTE. Pass all three and migrate; fail any one and stay in Word with the lint cap doing the enforcement.

Start with who is in the sample. The fourteen municipalities in the Municipal Specification Modernization Study, published this year, were volunteers — and public-works offices that volunteer to restructure a spec library are offices that already keep version-controlled clauses and named standards owners. Whatever the migration produced, it was produced inside unusually healthy documentation cultures. A before-and-after design with no control group cannot carry that result to a borough where one part-time engineer owns the spec book, and pretending otherwise is how migration budgets get approved on borrowed evidence.

The deeper limit is that the study cannot separate the architecture from the audit. You cannot convert a clause library to a typed DITA schema without reading every clause, and reading every clause fixes things on its own. Some share of what the scoreboard attributes to single-sourcing may belong to the forced content review, and the published design has no way to split the two. Two more constraints: the endpoints are process metrics — change orders and flagged ambiguity — not field outcomes, and a clause can be lexically crisp and still wrong. The observation window is also short against spec revision cycles that run on multi-year clocks. Drift is a slow variable; the study catches the first revision cycle at best.

What the Data Doesn't Tell You

Then there is the pooling. An average across fourteen dissimilar sites — a county with a dedicated standards office beside a township sharing a consultant — is a blend, not a benchmark. The gains plausibly concentrate where identical clauses genuinely sat in separately maintained files; sites whose reality fell below that drag the mean without telling you which kind of site you are. Before benchmarking yourself against the pooled figure, request the per-site dispersion. If the spread is wide, the average is answering a question nobody asked.

The decision rule survives all of this — but a screening threshold is a screening test, and screening tests fail at the edges in predictable ways. Four of them matter:

None of this flips the rule; it sharpens the intake exam. The audit that matters takes an afternoon: pull ten clauses that appear in your spec book, your bid package, and your inspection checklist, and count maintenance paths — the number of files a human must touch when one requirement changes. If the answer is one, you already have single-sourcing and the study's mechanism has nothing to add. If it is three or more, the rule applies with full force. And if you stay on Word under the threshold, treat the lint cap as a tripwire, not a quality certificate — it catches symptoms at release time, and no amount of hand-polished wording restructures a duplicated corpus.

Nine points. That is the entire rework improvement reported by municipalities with fewer than five engineering FTEs in the Municipal Specification Modernization Study (MSMS-26) — roughly one-third the full-sample effect — and the cause is structural, not noise. The study's protocol required a half-time administrative role to maintain the migrated library, and in lean offices that role consumed most of the hours migration freed up. According to the study's site-level tables, the distribution behind the headline mean is bimodal: adequately staffed offices cluster near the top of the range, small offices near the bottom. Quote the median split beside the headline in any internal memo, and if your shop runs lean, plan for the bottom mode.

The ambiguity gain splits unevenly too. Procedural and materials clauses drove the decline, falling about −41%, because their flagged ambiguity was mostly duplicated boilerplate that single-sourcing deleted outright. Performance clauses carrying numeric acceptance criteria — "shall achieve ≥95% efficiency at rated head" — improved only modestly. Their vagueness is semantic, not lexical: a threshold nobody can test or instrument stays equally opaque through any rewrite, and lexical scorers cannot see it at all. Line-editing fails here by definition; the cure for an untestable criterion is a named test method and an instrumented acceptance procedure, not a better sentence.

What −27% Hides

Two artifacts sit inside the rework figure itself. Nine of the fourteen sites knew the audit schedule in advance, and the "spec-defect" reason codes on change orders were entered by the same engineers whose output was scored. With no blinded control arm, demand effects cannot be excluded — engineers who know they are being measured code their reasons conservatively. Read the rework estimate as an upper bound.

Survivorship cuts the other way. Two additional pilot sites abandoned DITA mid-study after tooling-cost overruns and are excluded from the fourteen; re-including them with zero-gain imputation drags the rework estimate meaningfully lower — weaker, still favorable, but a different pitch to a council.

Nobody measured the readers. Contractors bidding from DITA-generated modular PDFs and the field inspectors using them were never tested for comprehension or bid-error rates. Small-sample eye-tracking work running in parallel to the study — not yet through peer review — suggests modularization raises navigation cost for novice readers even as ambiguity scores fall; the two metrics can move in opposite directions. A specification that lints clean but costs a novice bidder more mental assembly time is not an obvious win on a job site.

Last, the domain ceiling: twelve of the fourteen sites wrote horizontal-construction specifications — roads, water, earthwork — and only two covered IT or software procurement RFPs. Transplanting the headline effect onto API documentation or statements of work is extrapolation, not a finding.

Before the headline reaches a council memo, pull three items from the study's appendices: the median split by staffing level, the clause-type breakdown, and the abandoned-pilot count. If your inventory resembles the twelve construction sites and you clear the small-office floor, the effect is credible for you; otherwise gate releases behind the automated ambiguity lint described earlier and field-test a single modular PDF with a working bidder before migrating anything.

Three deliverable types is the minimum the canonical rule requires before any migration, and the Allegheny County Sanitary Authority — ALCOSAN, the wastewater agency serving the Pittsburgh area — sits exactly on that line. That is what makes its case file the sharpest test in the Municipal Specification Modernization Study: there is no slack above the threshold, so the results cannot be credited to surplus reuse. The standard specifications are extensive, feeding three deliverables — the standard spec book, per-project bid packages, and the consultant design manual. Nor is this an exotic early adopter: Scott Abel's benchmarking survey found 44% of organizations already publishing structured XML, 81% of those on DITA. That benchmark dates to 2013, which is precisely why a fully ledgered 2024–2025 municipal migration adds new evidence.