Back to app index
Sheaves for tunnel concept
If you model a tunnel project’s WBS as a small site (poset/category + a chosen topology of “what counts as an overlap”),
then “scope/schedule/state” becomes a multi-sheaf whose global sections are exactly the globally coherent plans you want,
and whose H¹ (or more practically, a computable obstruction residual) localises which interfaces prevent gluing.
The trick is to treat each department’s artefact as a stalk (risk sets, interval headcount functions, spec parameters),
and each interface as a restriction map; then Γ(F) is your patched global view, and project metrics become honest functionals on Γ(F)
(integrals for resource curves, expectations for risk). This is the same local-to-global mechanism as the textbook sheaf condition
(matching families glue uniquely).
Download the interactive HTML prototype
What the prototype contains
It’s one self-contained page with three canvases, each instantiating a different “ways” route you listed:
1) Risk as a Set-valued sheaf on a task-dependency category
- Base: a tiny Solway Firth tunnel WBS DAG (9 tasks, with a “diamond” at Design → {Procurement, Shafts} → Drives).
- Stalks: each task has a finite risk-exposure set (integers) + a probability mass on that set.
- Restrictions: each dependency edge carries an inequality “downstream ≥ upstream + Δ” (monotone constraint maps), exactly the style used in the 2024‑12 risk draft.
- Button Solve global sections computes Γ(F): all globally consistent risk assignments (or Γ(F)=∅ if incompatible).
- Button Expected-risk metric computes a conditioned expectation over feasible global sections using the simple independence ansatz (the audit rightly flags you must state/justify the independence assumption if you keep it).
Why this is a genuine sheaf move (not just CSP cosplay): global sections are the “globally consistent fusions” the sheaf-on-graph
exposition makes concrete: choose an element in each node stalk so that both node restrictions to each edge’s overlap agree.
2) Resource schedule as a function-valued sheaf over time-interval overlaps
- Base: phases as time intervals (months since NTP); overlaps are literal set-theoretic overlaps of intervals.
- Stalks: piecewise-constant headcount functions on each phase interval.
- Restrictions: interval restriction (the canonical map in the 2024‑12 resource draft).
- Button Check overlap agreement reports exactly where phase functions disagree on overlaps (i.e., where the matching-family condition fails).
- Button Glue constructs the unique global section if overlap agreement holds (draft’s “gluing to a global function”).
- Button Sheafify forces consistency by selecting a single global profile via a reconciliation rule (max/min/mean) and restricting it back to each phase; this is deliberately framed as “pragmatic sheafification”, mirroring the topos/design view that sheafification is a left adjoint that forces gluing while preserving the base category.
The audit explicitly called out that the resource draft had manual adjustment but no algorithmic procedure; this prototype gives you an
algorithmic “minimum viable sheafifier” (it’s not canonical in the Grothendieck sense, but it’s operational).
3) A cohomological obstruction demo (H¹ / “cycle-sum” witness)
- Base: a small cyclic “interface graph” (think coupled design/procurement/safety constraints, i.e., the part of a megaproject that is not a DAG).
- Data: an ℝ-valued 1‑cochain ω on edges; you want to know if there exists a “global potential/spec” x on nodes with δ⁰x = ω.
- The page shows the classic phenomenon: ω is globally integrable iff cycle sums vanish; otherwise ω represents a nontrivial H¹ class (an obstruction to patching). This is the same conceptual move as the cellular sheaf cohomology story: H⁰ is global sections, and nontrivial H¹ measures failure-to-glue.
This is the piece your “Čech cohomology detects contradictions” conjecture is gesturing at, but in a computationally cheap linear setting
where the obstruction is visible as a residual/witness, not just a slogan.
The representation you asked for (clean formal core)
Base object: WBS as a site
Take a WBS as either:
- a poset (by refinement, or by precedence if you quotient to antisymmetry), equipped with an Alexandrov topology, where presheaves and sheaves coincide (so you can just write a functor and call it a sheaf on that topology).
- a graph/cell complex (to support overlaps as edges/2‑simplices), enabling cellular sheaf cohomology computations.
The “less popular but useful” move is to treat the choice of covers as governance: picking a Grothendieck topology J on your WBS category says what
families of local artefacts must be jointly reconcilable before you declare “the plan is coherent.”
Data: a multisorted sheaf of project state
Let State be a product of components (risk, resources, schedule, specs, etc.). Formally: a sheaf valued not in Set alone, but in a category that
supports the structure you care about (Sets for discrete, Vect/Modules for linear constraints, Convex sets for LP‑feasible regions, etc.).
A simple archetype:
- For each task/work-package U, assign a stalk F(U) of feasible local states (e.g., admissible (start,end,crew,risk) tuples).
- For overlaps V ⊆ U (interfaces, shared resources, handover artefacts), provide restriction maps F(U) → F(V) (“forget down to the interface variables” or “restrict a function to a sub-interval”), as in the resource draft’s interval restriction.
- A global section is then exactly a coherent project plan. (And it may be empty, unique, or plural—your “multiple sections = equally valid variant plans” is literally correct in this language.)
Metrics as continuous functionals on Γ(F)
Once Γ(F) is nonempty, define metrics as maps Γ(F) → ℝ:
- For function-valued resources, metrics are integrals like ∫ cost_rate(t)·headcount(t) dt (continuous functional in standard topologies on function spaces).
- For discrete risk, metrics are expectations under a chosen measure on Γ(F) (your 2024‑12 risk draft’s “expected risk score” idea).
The audit is correct that your earlier drafts stop one step short: they produce a global section but don’t explicitly wire it into a higher-level project
fitness function (earned value, SLA conformance, etc.).
How an LLM actually fits (without “hand-waving oracle” failure)
The most robust role is: LLM as extractor + proposer, deterministic sheaf engine as judge.
- Extract: from local plans, minutes, risk registers, and Primavera/MSP exports, the LLM proposes stalk data (risk sets/distributions, piecewise resource curves, interface specs) and candidate restriction maps (inequalities, unit conversions, tolerance constraints).
- Judge: the sheaf engine checks functoriality/unit laws (composition constraints), computes Γ(F), and computes obstruction witnesses (residuals, nonzero cohomology classes in linearised cases).
- Propose remediation: LLM suggests minimal edits to local artefacts that would kill the obstruction (e.g., weaken Δ on a dependency, reconcile inconsistent overlaps, or add a missing interface object—“fill in a 2‑simplex” in the cellular picture to tighten consistency checks, as the cohomology exposition hints).
Stretch target: “blueprint gradually filled in” as a sheaf over a state poset (domain-theoretic flavour)
- Let P be the poset of partial completions (states), ordered by information/refinement (“⊑”: you add constraints/decisions; you don’t retract).
- Put the Scott topology on P if you want convergence/limits of directed refinements (this is exactly where the sheaf text gestures that Scott topology is appropriate when the poset includes infinite elements; it’s the natural bridge to domain theory).
- Define a sheaf (or stack, if you want nontrivial automorphisms like alternative designs) of admissible local states over P; then “project execution” is literally a directed path in P, and “completion” is a maximal (or terminal) global section.
The geometric rhetoric (“manifold of project states”, “tangent bundle = admissible next moves”) becomes concrete if you use a combinatorial/poset analogue:
the “tangent directions” are the immediate refinement moves from a state (covering relations), and a “policy” is a section of the corresponding cosheaf
of allowable transitions; you can then ask for path-lifting, obstruction to extension, etc., exactly as in gluing problems. (If you want the more Grothendieckian twist:
changing what counts as a cover/topology J is changing what counts as a legitimate local-to-global inference during execution—very close to the “creative preservation”
sheafification story in the design paper.)
One angle: try enriching the WBS category in a Lawvere metric (time/cost “distance”) so that obstructions carry not just existence/nonexistence but a minimal repair cost.
Back to app index