All notable changes to the master-builder skill.

Closes the three items that were standing open, and the one that went least well taught the most.

• Cross-model run (evals/results/2026-07-31-crossmodel.md) — the same blinded design on Haiku. No skill: 1/8. With skill: 6/8. Five of eight cases changed verdict, none against. The lift is larger on the weaker model (12.5% → 75%, vs 50% → 94% on the larger one), which is the direction predicted if the skill supplies knowledge rather than style — the strongest evidence yet against the "it was just the model" reading. Honestly scoped: this is cross-model, not cross-vendor, and the single-vendor limitation is not lifted.
• scripts/verify_plugin.py — walks the loader's documented resolution path (catalog → source → skills/<name>/SKILL.md → frontmatter), confirms the skill is discovered as /master-builder:master-builder, and proves no top-level skills/ tree exists to double-load. Now a CI gate. It states in its own output that it does not run /plugin install, which remains unverifiable without an interactive CLI session.
• examples/vietnam-localization-walkthrough.md — a second worked example, demonstrating the localization procedure on a jurisdiction with no dossier and no router entry. Vietnam's amended Law on Construction took effect 1 July 2026, so guidance assembled from 2024 sources describes a superseded regime — R2's purpose demonstrated live. Ends with what it could not resolve.

• Internal reference filenames were leaking into user-facing prose. A grader caught global-codes.md §2 surfaced to a user as though it were a citation. Filenames are navigation, not citations; cite the code section or standard instead.
• "A check you cannot verify is not a check." One answer performed the sanity check required since v0.11.0 and got the check itself wrong by 1,000× — GWh where its own rate gave MWh, with a cost figure that silently agreed with the error. The form of verification without the substance, which is worse than none because it launders the original error. SKILL.md now requires the check to carry its units through and land in the same order of magnitude as the claim.

numeric-sanity failed in both arms — the v0.11.0 sanity rule did not hold on a weaker model. The rule is capability-dependent: it works on a stronger model and degrades to ritual on a weaker one. That is a real bound on what this skill can promise, and it is stated in the results file rather than buried.

Closes the oldest open debt in the repo: the escalation figure a blind grader flagged three releases ago. The grader was right — our own number was stale.

• construction-delivery.md §3 — the ~8%/yr escalation figure was high. Re-verified against Q1 2026 data: Mortenson's nonresidential index ran +1.69% on the quarter, +6.77% year-over-year (materials +7.0%, trade-partner work +6.6%); Turner's ran +1.3% on the quarter and +3.1% against the 2025 average. Both are now quoted, dated and attributed, replacing a single unsourced figure.

• "Name the index" as doctrine. The re-verification surfaced something more useful than the correction: the two published indices disagree by more than 2× — different baskets, geographies, and bid-price-vs-input-cost bases. So "construction escalation is X%" is not a fact until you say whose index, which quarter, and which market. Pick the index that matches the project type and location, quote it with its date, and re-check at estimate time rather than carrying a number forward.
• Behaviour regression case index-provenance (21 cases) guarding exactly this.
• evals/results/2026-07-25-baseline.md — the open flag is marked RESOLVED in place, with the outcome, rather than the record being quietly edited.

The flag came from an adversarial reader of the skill's own output, was recorded as unresolved rather than dismissed because it was ours, and turned out to be correct. That loop — flag, verify, correct, guard with a test — is the skill's own doctrine (SKILL.md: date the data and flag it when stale) applied to itself.

Extends the skill from reading project documents to judging AI output about them — and adds the defect class that cross-referencing structurally cannot catch.

• document-intelligence.md §7 — "Generated output: plausible is not buildable." The mirror of v0.11.0's wrong-but-well-formatted number: geometry that looks right and cannot be built. Grounded in the DreamHouse benchmark (arXiv 2603.24866; 26,000+ structures, 13 styles, LOD 350, a deterministic 10-test validation framework), whose finding is blunt — "physical validity is not a byproduct of visual imitation, and vice versa." Best joint structural-and-visual pass rate: 7.1%, with the axes coming apart between models (one leads structurally at 79.2% while scoring lowest visually). Practice: never treat a render as evidence of feasibility; validate generated geometry against a deterministic engine, not another model; constrain generation up front; the stamp does not move.
• "Protocol dominates model" — the same benchmark found a 33-point structural swing (45.4% → 78.5%) from changing task scaffolding alone, larger than between-model differences under a fixed protocol. Restructure the task before reaching for a different model.
• document-intelligence.md §6 — the omission problem. A cross-document join finds contradictions; it will never find a requirement absent from both documents. Adds the three-pass shape: what the documents address → what should apply, researched independently → the delta, reported as gaps to confirm with confidence and citations. Pass 2 is the one that cannot be derived from the project file, and it is the whole value.
• Cross-links from build-doctrine.md §5 (a generated artifact's gate must be physical, not visual) and digital-toolkit.md §7 (generative design needs a deterministic referee).
• Two retrieval cases (generated-geometry, missing-requirement). 32/32.

The source article's two headline figures were both wrong, and the paper was used instead: the scaffolding gap is 33 points, not 46, and "physically valid only 7.1% of the time" is actually the joint structural-and-visual rate — structural alone is far higher. Recorded here because the skill tells its users not to propagate unverified numbers.

Surveyed the AEC skills actually on GitHub. Imported the two genuine gaps it exposed on our axis, and deliberately declined the rest.

• Delay analysis (construction-delivery.md §4) — the skill said EOT is "excusable vs compensable" and stopped, never saying how entitlement is proved. Now: the three independent questions (excusable? compensable? on the critical path — and who owns the float), concurrent delay as the place most disputes actually live, and the method families from the SCL Delay and Disruption Protocol (2nd ed.) and AACE RP 29R-03 (as-planned vs as-built, windows/contemporaneous period, impacted as-planned, time impact analysis, collapsed as-built) with what each is good for. Plus the part practitioners insist on: records beat method, and most delay disputes are notice disputes.
• Fire/life-safety and accessibility depth (global-codes.md §2) — both were single bullets in a skill claiming code competence. Now the chain that actually drives design (occupancy classification → occupant load → exits and egress width → travel distance → discharge; ratings, compartmentation, suppression, smoke control, firefighter access), that sprinklering buys allowable area and height so it is a commercial decision too, and that accessibility is a civil-rights obligation enforceable independently of the permit, with alteration triggers.
• Two retrieval cases (delay-claim, egress-fire). 30/30.
• README — a "Complementary, not competing" section.

• Skills-Architects (MIT, 29 skills) covers the design axis — acoustics, daylighting, typology, spatial planning, design theory. Genuinely good, actively maintained, and a different axis from this skill's developer/builder one. Absorbing it would bloat a skill already at 13 references, duplicate maintained work, and re-inflate the description trimmed in v0.13.0. Linked instead.
• DDC (MIT, 221 skills) — the skills are Python implementations, a tooling library rather than doctrine. Vendoring would break the zero-dependency rule and change what this skill is. Its delay taxonomy is what prompted the delay-analysis section above; that idea is credited, none of the code was taken.

Imports from AlpacaLabs' skills-for-architects (MIT, credited) — the first of these external repos with a licence permitting reuse. Its strongest contribution is language discipline; its context audit exposed bloat in our own frontmatter.

• Compliance language, as explicit word-for-word substitutions. The skill said "route it to a professional" but never gave the wording, which is where a careful analysis quietly becomes an unsupportable claim: never "complies with the code" → "appears consistent with IBC 2024 §1004.5"; never "no violations" → "no violations were identified in the data reviewed." Never claim completeness.
• Citation format — a section without its edition is unverifiable. IBC 2024 §1004.5 (check local amendments), ASTM E119-20, ASHRAE 90.1-2019, with a public link where one exists. Complements v0.11.0's "never quote a threshold from memory."
• Show the inputs, not just the answer — no derived number without the values and formula behind it, so the reader can reproduce it and see what to change when assumptions shift.
• Date the data and flag it when stale — distinguish what was looked up from what is recalled. (This is the rule that would have caught the unresolved ~8%/yr escalation flag from v0.12.0.)
• Area-type discipline in real-estate-finance.md — GSF/GIA vs NSF/NIA vs USF vs RSF differ by 15–30%, enough to invert a deal's appearance; state the loss factor when converting, and say whether structured parking is in the denominator.
• Two regression cases (compliance-language, show-your-work). Eval set now 20 cases.

• description cut from 2,024 to 1,370 characters (~506 → ~343 est. tokens), a 32% reduction, with every trigger term retained and asserted. A skill's description is always-loaded standing context — it costs tokens in every session whether or not the skill fires — and five releases of appending triggers had bloated it. Their published context audit is what prompted measuring ours at all.
• README — /plugin is a terminal-panel command and is not available in every surface; the marketplace option now says so rather than implying it works everywhere.

Runs the counterfactual the previous two runs couldn't: does the skill change behaviour, or is the model already like this? Blinded baseline, and the answer is yes — measurably, on 8 of 18 cases.

• evals/results/2026-07-25-baseline.md — a blinded A/B baseline. Same 18 questions to fresh agents with and without the skill, prompts identical but for one line; responses paired as "Response A / Response B" with arm order randomised per case; graded by an agent told only that two assistants answered, instructed not to infer which, and barred from the key. Control 9/18 PASS (6 partial, 3 fail). Skill 17/18 PASS (1 partial, 0 fail). Head-to-head: skill stronger on 15, control on 0, 3 ties — the skill arm was never judged weaker.
• scripts/blind_pairs.py — builds the blinded comparison and the withheld unblinding key, seeded so the shuffle is reproducible.
• evals/results/2026-07-25-baseline-answers.md — both arms verbatim, the blind grader's report, and the key, as evidence.

• The design validated itself. Before unblinding, the grader volunteered that the responses "cluster into two consistent stylistic families," scoring them 17/1/0 and 10/5/3. Unblinding mapped those onto treatment and control. It detected the effect without knowing the arms existed.
• Three outright control failures: asserting foreign procedural detail as fact; reciting an IEBC threshold from memory ("the number you're reaching for is 5%"); and — the important one — capitulating when the user waived the caveat ("I won't hold you to it" → "use 195 mph").
• The prediction was half wrong, and the wrong half matters. Boundaries were expected to need no help. They hold unaided on the first ask (structural-boundary: tie) and break on the second. The model is cautious until a user gives it permission not to be — which no single-turn eval catches.
• Where the skill adds nothing, recorded as plainly as where it adds a lot: three ties on unpressured first-ask cases, where the skill documents caution the model already has.
• The three regression cases added the same day now double as the cleanest arm separators.

Single model family (no cross-model transfer shown); n=1 per case per arm; the bundle was handed over directly, so trigger and retrieval reliability are not exercised end-to-end. One accuracy flag may implicate the skill itself — a "~8%/yr escalation" figure the grader read as stale — unresolved and flagged for re-verification rather than assumed correct because it is ours.

The second eval run scored 15/15 — and then found something more important than the score: the v0.10.0 fix had a dangerous side effect. A well-formatted, properly-hedged number is more dangerous when it's wrong, because the presentation buys credibility the content hasn't earned.

• A hedge licenses imprecision, not invention. Run 2 produced three technical errors that passed every criterion clean: leafy-greens energy quoted as "250–400 kWh/kg (indicative, Class 5, verify…)" when published figures cluster at 10–40 kWh/kg — ~20× high, perfectly caveated, and load-bearing to the argument; an ASD conversion stated as "about 0.6 ×" when it is √0.6 ≈ 0.775 (its own worked figure contradicted its stated rule); and IEBC gravity/lateral triggers quoted inverted. SKILL.md now requires a band to be sanity-checked a second way — order of magnitude from first principles, the unit basis (per kg or per m²? fresh or dry? per day or per month?), and consistency with everything else in the answer — before it is stated.
• Never quote a code threshold, coefficient, or conversion factor from memory. Name the section and say to read it. This resolves a real doctrine gap the run exposed: the skill refused to quote the IEBC thresholds in one answer and quoted them backwards in another, with nothing saying which was right. A ROM cost band is an estimate and may be approximate; a code threshold is a fact you either have or don't.
• Numbers in one answer must reconcile. Run 2 itemised delay costs summing to $1.53–1.74M then stated the total as "$1.1–1.4M" — the gap was exactly the LP preferred return, silently dropped.
• Hedge placement now applies to conclusions, not just numbers. structural-boundary regressed in register between runs, opening "Probably yes" on a question that needs a stamp. Lead with the limit.

• Three regression cases (numeric-sanity, code-threshold-recall, arithmetic-consistency), each documenting the exact run-2 defect it exists to catch. Eval set is now 18 cases.
• evals/results/2026-07-25-run-2.md — 15/15 PASS, with the grader's finding that 15/15 with zero PARTIAL is a measurement failure, not a result: after the v0.10.0 amendments the criteria tested only whether conventions were followed, and nothing tested numeric accuracy or arithmetic consistency. The three cases above close that gap.
• evals/results/2026-07-25-run-2-answers.md — the verbatim answers, kept as evidence.

All answering agents and both graders are the same model family and share a house style. A cross-model run, or a no-skill baseline, would separate skill effect from shared disposition — that comparison is worth more than a higher pass rate, and has not been done.

Adds the Claude Code plugin marketplace — a fifth install path, with auto-updates — and records the first behavioural evaluation of the skill.

• Plugin marketplace. .claude-plugin/marketplace.json + plugin/.claude-plugin/plugin.json, so users can /plugin marketplace add ibuilder/master-builder then /plugin install master-builder@ibuilder, and pick up updates with /plugin marketplace update.

• plugin/skills/master-builder/ — the skill in the layout a plugin expects (a skills/ directory under the plugin source), generated by scripts/build.py from the same source as everything else, so it cannot drift. Deliberately nested under plugin/ rather than a top-level skills/ so that cloning this repo into ~/.claude/skills/ can't double-load the skill.

• validate.py now checks the manifests: required fields, that the plugin source and its skills/<name>/SKILL.md actually resolve, and that the marketplace, plugin, and CHANGELOG versions agree.

• CI fails if the generated plugin tree drifts from source.

• First behavioural evaluation run, recorded in evals/results/2026-07-25-run.md: 11/12 PASS, 1 PARTIAL, 0 FAIL. Answers were produced by fresh agents given only the bundled skill; a separate independent agent graded them. The record states its own limitation plainly — it bounds content quality, not trigger or retrieval reliability in a live session.

• Three second-ask pressure cases the run showed were missing (pushback-hazard, pushback-stamp, pushback-authority) — a waived caveat, schedule pressure plus flattery, and a claimed credential. Boundary claims are likeliest to break on the second ask, and nothing tested that. Eval set is now 15 cases.

• A real contradiction in the skill's own doctrine, caught by the eval. "Give a range and an estimate class" and "ground it in place; ask or state the assumption" pulled in opposite directions when the location was missing — and the run proved it empirically, answering the same epistemic situation two different ways (a hedged band for a hazard value, silence for a cost). SKILL.md now resolves it: band it under a stated assumption; don't withhold — withholding is correct only for a hazard value that must be read off a map, or anything needing a stamp.
• Hedge placement is now an output convention — put the qualifier in the same sentence as the number, never in a closing paragraph, because a skim-reader keeps the figure and drops the caveat. Both close calls in the run turned on exactly this.
• evals/behavior.jsonl — the hazard-value-honesty must_not bundled two independent tests and gave graders no rule for the hedged-band case; split, with the rule stated. cost-conventions criteria realigned to the resolved doctrine.
• docs/banner.svg — the wordmark and lifecycle line had only ~10% and ~13% headroom to the canvas edge. SVG resolves fonts on the viewer's machine, so a fallback wider than Georgia/ui-monospace would have overrun the banner. Both are now pinned with textLength; verified against a deliberately much wider substitute font.

States the skill's own thesis, which had been implicit for eight releases. Informed by a report on the evolution of the master builder from The Pillars of the Earth to the modern industry.

• A banner (docs/banner.svg) — a half-section through a cathedral bay drawn in construction-document convention (dimension line, ground hatch, pier coursing, sheet corner ticks). SVG, so it is version-controlled, scales, and needs no binary asset.
• README — "A short history of the master builder." One mind and one workshop (capomastro, Baumeister, architectus; The Pillars of the Earth as the cultural model) → fragmentation in the 18th–19th centuries as knowledge outgrew a single craft lineage → property becoming a financeable asset → the industry's repeated attempts to put integration back (design-build, CMAR, IPD, owner's rep) → why an integrator is the job now.
• SKILL.md — "Why this matters." The thesis, stated plainly at last: fragmentation was an adaptation, not a fall, but it moved coordination cost onto the owner — so projects fail at the boundaries between disciplines far more often than inside them, and the modern master builder is not the person who knows everything but whatever keeps everything coherently connected.
• Two retrieval eval cases covering the new material (now 28/28).

• construction-delivery.md §1 — delivery methods reframed as a history rather than a menu: every method is a different answer to how much integration do we buy back, and who pays for it? Includes the verified DBIA/FMI projection of design-build at up to 47% of US non-residential spending in assessed segments in 2026. Choosing a delivery method is choosing where the seams will be.
• development-lifecycle.md §11 — the question the pro forma doesn't ask: whose objective the numbers encode, and who bears costs while others collect benefits. Distributional facts have a commercial edge (community opposition is entitlement risk), and market/valuation/demographic assumptions can encode bias while presenting as neutral arithmetic.
• build-doctrine.md §1 — preserve the why, not just the what. Stable identity records the current state; the constraint that produced it is the other half. Projects outlive their decision-makers, and the classic failure is inheriting a choice, losing its rationale, and rediscovering the constraint the expensive way.

Proves the skill works. Seven releases had shipped features with no validation of answer quality — by its own build-doctrine.md §5 standard ("never skip a stage"; "the most important milestone is proving the edge") the repo was at the shipped-but-unvalidated stage it warns about. This closes it.

• evals/retrieval.jsonl + scripts/eval_retrieval.py — 26 realistic questions, each asserted to route to the reference a builder would reach for. Scores every reference against the question's salient terms and requires the expected one to rank first. Catches two real regressions: coverage loss (a topic quietly stops being covered) and placement drift (content migrates to the wrong file, so progressive disclosure pulls the wrong reference). CI-enforced; currently 26/26.
• evals/behavior.jsonl + scripts/eval_behavior.py — 12 questions with the conventions each answer must obey: state jurisdiction and code edition, carry units + currency + date, give a range and an estimate class, put a boundary on any carbon figure, route life-safety to a stamp, report audit coverage, and refuse to fabricate a hazard value. CI validates the set's structure; grading needs a model with the skill loaded, and the runner says so rather than faking a score.
• CI gains both eval steps; README gains a "How it's validated" table that marks plainly which gates are automated and which are not.

• construction-delivery.md §3 — the retrieval eval surfaced a genuine seam: escalation is an estimating topic but "who carries commodity risk" is an allocation one, and the two files did not cross-link. Added the pointer to risk-insurance.md §3.

One skill for the built world, adaptable to any location. Names the organizing principle explicitly — municipal code + climate = the book for a place — and makes both halves operational.

• references/climate-building-science.md — the climate half, which the skill had been missing. Loads were covered; the physics that decides assemblies was not. Vapour drives from warm-humid to cool-dry (so the correct wall in Minneapolis is the wrong wall in Miami, from the same physics); every assembly must dry in at least one direction; the four control layers (water, air, vapour, thermal) and the rule that they must be continuous and connected; air leakage transports far more moisture than diffusion — prioritize air-tightness over vapour-barrier fixation; Köppen families as the global spine (works anywhere, unlike national zone maps) with what each forces on a building; mixed climates as the genuinely hard case (seasonal reversal — a code-compliant wall can still rot); ground (frost depth, expansive clay, permafrost, radon); durability (corrosion, freeze–thaw, UV, termites); and climate as a schedule/carry driver.
• jurisdiction-dossiers.md §1 — the localization procedure: six resolutions (place depth, adopted code stack, the full AHJ set, climate and hazard basis, market and delivery conventions, licensure) that build "the book" for any location on Earth. Plus §2, a code-family router with explicit guidance for countries not in the table. The four worked dossiers now each carry their climate too.
• MCP tool master_builder_localize(place) — returns the six-resolution worksheet for a named place, served from the reference so there is one source of truth. It deliberately returns the procedure and what to verify, never invented local values. Self-test now 25 checks.

• SKILL.md — the ground-in-place rule now leads with the code + climate = the book framing and routes to the localization procedure; building physics added to what location determines; step 5 routes to building science; description extended to trigger on envelope, vapour/air barriers, condensation, insulation, and foundations.
• README — "adapts to anywhere on Earth" and "knows the physics, not just the rules" capabilities.

Proves the "anywhere on Earth" claim with worked dossiers, and adds a document-intelligence discipline distilled from studying open AEC-AI work by others (see Credits in the README).

• references/jurisdiction-dossiers.md — the unfamiliar-jurisdiction method demonstrated, not just described. Four dossiers, each teaching a different structural lesson: England & Wales (the regulator can be the schedule — Building Safety Act Gateway 2 running 25+ weeks against a 12-week target with ~45–50% rejection), Dubai (the AHJ depends on which plot you stand on — DM / DDA / Trakhees, with Civil Defence as a mandatory parallel track), Australia (published ≠ adopted — NCC 2025 published 1 May 2026 but staggered to 2027 in several states), Canada (a model code has no force until a province enacts it; NBCC 2025 adds GHG as a formal objective). Plus a contributor template.
• references/document-intelligence.md — extracting trustworthy numbers from drawings, specs, change orders and pay apps: index-once/query-many, confidence by provenance (a schedule count and a pixel-scaled distance are not the same kind of fact), encoded sanity bounds, the text-linearisation trap (never trust a $ column that survived PDF→text; recompute from quantity × rate), coverage-aware findings ("0 findings" is not a clean bill of health), spec↔drawing coordination joins, unit/designation canonicalization, non-destructive markup, and the deterministic-finding vs judgment-call split.

• global-codes.md — correction: NCC 2025 and NBCC 2025 were treated as simply "current." Both are published-but-not-uniformly-adopted; the section now states the adoption reality and the trap.
• construction-delivery.md — added the owner-side change-order pricing audit: markup/O&P stacking, unit-price inflation, labor-rate padding, quantity tie-out.
• pro-forma-review.md — new §7 "Report what you could not check," applying the coverage-aware rule to model review.
• digital-toolkit.md — MCP connectors that drive the incumbent authoring tools (Revit, AutoCAD, Navisworks, Dynamo, Grasshopper); document-intelligence flagged as required reading for AI document workflows.
• SKILL.md — both new references wired into the protocol (step 1 and step 6) and the table; description extended to trigger on takeoff, spec cross-check, and change-order audit.
• README — a Credits & related work section.

Content pass. Closes the one structural hole in the protocol and gives existing buildings the depth the skill kept asserting they deserved.

• references/risk-insurance.md — fills the only protocol step (7, Risk) that had no reference behind it. The risk loop and the allocation principle (control and price and absorb — risk pushed onto a party that can't carry it isn't transferred, it's disguised), the risk register families, the contract clauses that actually fight (differing site conditions, EOT/concurrent delay, LDs, indemnity, waiver of subrogation), insurance products (builder's risk/CAR with DSU/ALOP, CGL, PI, wrap-ups, SDI), why a surety bond is credit rather than risk transfer, the 2026 market split (non-CAT rate relief vs ~12% CAT-zone increases; severe convective storm >$51bn US insured losses in 2025), insurability as a feasibility variable, and contingency as retained risk capital.
• references/adaptive-reuse.md — existing buildings on their own terms: why reuse now dominates (carbon + repriced basis + mandates), the physical screen that kills candidates fast (plate depth, floor-to-floor, structure, risers), the IEBC 2024 compliance paths — prescriptive / work-area / performance — and the rule that you must pick one and not mix, change-of-occupancy and work-area triggers, hazmat/structural/reality-capture due diligence, office-to-residential economics (record ~11.8M sf in 2025; only ~25–30% of buildings viable), building performance standards (NYC LL97 at $268/tCO₂e over cap, 50+ US jurisdictions, EU EPBD MEPS to class D by 2030/2033), and how underwriting a reuse deal differs.

• SKILL.md — protocol step 7 now routes to risk-insurance.md and states the allocation principle; step 2 notes that with an existing building the order inverts (structure constrains program) and routes to adaptive-reuse.md; both added to the reference table; description extended to trigger on risk, insurance, surety, conversion, IEBC, and retrofit mandates.
• Reciprocal cross-links added from development-lifecycle.md (risk register, adaptive-reuse DD), sustainability-carbon.md (reuse hierarchy), and construction-delivery.md (contract risk clauses).

Makes the skill callable as a tool by any MCP-capable agent, and holds the repo to the skill's own build doctrine (staged validation, honest status, operational discipline).

• scripts/mcp_server.py — a zero-dependency MCP server (stdio) serving the protocol and the reference library as four read-only tools (master_builder_get_protocol, _list_references, _read_reference, _search) plus MCP resources. This restores progressive disclosure for non-Claude agents: they pull only the reference a task needs instead of the whole corpus. Stdlib-only Python 3.9+ — no SDK to install or pin (the official Python SDK is a 2.0.0b1 pre-release with a renamed API; per build-doctrine.md §4 the core must not depend on a moving part). Ships a 22-check --selftest.
• scripts/validate.py — enforces the authoring rules CONTRIBUTING only stated: frontmatter shape, slug-form name, substantive description, lean SKILL.md, reference table ↔ files on disk both ways, and that every cross-reference resolves. build-doctrine.md §8 (compliance-as-code) applied to the skill itself.
• .github/workflows/ci.yml — on every push/PR (Python 3.9 + 3.12): validate the skill, self-test the MCP server, exercise it over real stdio, and fail if dist/ has drifted from source.
• README — an MCP server section and a CI badge.

• CONTRIBUTING — contributors run validate.py and the MCP self-test; the stdlib-only rule is explicit.

Portability + reproducible build. The skill's knowledge now ships in a form any assistant can use, and all distributables are generated from one source so they can't drift.

• dist/master-builder.bundle.md — a one-file portable bundle (the protocol + every reference, with setup notes) for use in ChatGPT, Gemini, Perplexity, or any API/model — not just Claude.
• scripts/build.py — one dependency-free build that regenerates master-builder.skill, master-builder.zip, and master-builder.bundle.md from SKILL.md + references/.
• README — a "Use it in other assistants" section (Custom GPT / Gem / Space / system-prompt recipes) and a note that dist/ is generated by python scripts/build.py.

• Packaging moved from an ad-hoc zip step to scripts/build.py; dist/ is now generated, not hand-built.
• CONTRIBUTING — contributors rerun python scripts/build.py after edits.

Reconciliation pass against the current Massing repo + research: adds climate adaptation/resilience (the counterpart to carbon mitigation) and sharpens tool descriptions to the platform as it actually ships.

• sustainability-carbon.md §7 "Adaptation — the other half of climate" — design-to-future-hazard, resilience-as-underwriting (insurance/downtime/exit), and siting/passivity as the cheapest resilience.
• global-codes.md §4 — ASCE 24 flood-resistant design, Rational-Method stormwater sizing, and a forward-looking physical climate-risk overlay ("design to the future hazard, not just the historical map").
• SKILL.md — "The climate risk" added to the ground-in-place list; resilience/adaptation added to the description triggers and the reference-table summary.

• digital-toolkit.md — reality-capture (2D→BIM raise, scan-to-BIM deviation, GIS/DEM overlays, LAS/LAZ), operations (CMMS, EUI, reserve study, UNIFORMAT II → Facility Condition Index, ESG Scope 1/2 + climate-risk rating), and AI/MCP (Claude skill pack; schedule-risk/carbon/permit-readiness engines off one model) updated to the current Massing capabilities; climate-resilience row added to the software map.
• README — "counts the carbon, the climate risk, and the power" reflects the resilience addition.

Currency + decarbonization pass. Adds carbon as a first-class development discipline, refreshes code editions to mid-2026, and folds in the power-constrained and tariff-driven cost environment.

• references/sustainability-carbon.md — whole-life & embodied carbon: the EN 15978 module map, the accounting standards (EN 15804 EPDs, RICS WLCA 2nd ed, ISO 14040/44), the 2026 regulatory teeth (EU CBAM definitive period, EPBD, Buy Clean, LEED v5 embodied-carbon prerequisite, ICC 2027 carbon focus), carbon-is-money (CBAM cost, green premium/brown discount, transition/stranded-asset risk), the reduction hierarchy, and a hotspot/cost-concentration method.

• SKILL.md — ground-in-place rule now derives power (interconnection queue) and carbon from location; carbon reference added to the protocol and the reference table; carbon-boundary output convention added; description extended to trigger on carbon, decarbonization, and power/data-center questions.
• global-codes.md — utility-service gate sharpened to the 2026 power reality (~2,600 GW US interconnection queue, PJM 8-year application-to-operation, transformer 2–4-year lead); ICC 2027 and second-generation Eurocode timelines (definitive 2026 / publish 2027 / withdraw 2028) made concrete; carbon added to the energy/sustainability stack and the editions note.
• digital-toolkit.md — ISO 19650 second-generation revision (DIS 10 Mar 2026, BIM→IM, unified 9-step process, final 2027); 5D take-off feeds whole-life carbon; embodied-carbon row added to the software map.
• construction-delivery.md — 2025–26 escalation/tariff environment (Section 232 ~50% metals tariffs, ~8% aggregate escalation, price-escalation clauses) added to estimating; transformer/switchgear lead times added to long-leads.
• real-estate-finance.md — escalation/tariff and carbon (CBAM, green premium/brown discount, transition risk) added to the common-traps list.

Enrichment pass after grounding on real projects and a live feasibility test (a retail-to-vertical-farm development thesis and its pro-forma model).

• references/build-doctrine.md — cross-cutting engineering lessons distilled from real platforms (source-of-truth & stable identity, do heavy work at the right layer, open interchange, staged- validation gate, hard rails on irreversible actions, honest status, compliance-as-code).
• references/pro-forma-review.md — forensic model/deal review: reframe the asset, reconciliation pass, a defect checklist (NOI gross-up, dropped cost lines, non-cash in OpEx, zero-vacancy, unit errors), the three-questions test for assumptions, cost-concentration analysis, and validate-demand-before-capital.

• SKILL.md protocol now leads program analysis with "name what the asset actually is" (reframe first).
• global-codes.md — utility interconnection / will-serve added as a schedule-and-cost gate for energy- and water-intensive uses and on-site generation.
• digital-toolkit.md and construction-delivery.md corrected to the verified architecture of the real repos (Massing on That Open Fragments + IfcOpenShell with GUID-stable server-side edit recipes, three pillars Model/Construction/Finance, code-intelligence pre-checks; gcPanel/ConstructAI on Next.js/TS).

Initial release.

• SKILL.md — the Master Builder Protocol, ground-in-place rule, professional boundaries, output conventions.
• References: global-codes.md, development-lifecycle.md, real-estate-finance.md, construction-delivery.md, digital-toolkit.md.