# Dungeon Gleaner Writeup Roadmap **Filed by:** Dungeon Gleaner contributor for Sundog applications correction (2026-05-07) **Audience:** Sundog research project — the maintainer of sundog.cc; future-Claude executing this roadmap; skeptic-observers reading the eventual writeup **Status:** Roadmap pass. This doc plans the correction; the next 2-3 passes execute against it. **Sources:** - `applications-gallery.html` § Dungeon Gleaner card (current state — flagged "Product Expression" with `Needs: performance harness, fixed scenes, and visual acceptability criteria`) - `docs/APPLICATIONS.md` § Dungeon Gleaner / DCgamejam2026 (current state — built around glass/window reflection and pressure-washing) - `docs/MONEY_BAGS_WRITEUP_ROADMAP.md` (template — voice, structure, evidence-tier discipline) - `/docs/VERB_FIELD_NPC_ROADMAP.md` (DOC-83, 2026-04-08) — the actual Sundog application - `/docs/NPC_SYSTEM_ROADMAP.md` (DOC-9) — patrol substrate the verb-field replaces - `/engine/verb-field.js`, `/engine/verb-nodes.js`, `/engine/dungeon-verb-nodes.js`, `/engine/verb-node-seed.js` — runtime implementation - `/engine/reanimated-behavior.js` — second consumer of the same field tick --- ## 0. The story shape — why this writeup is a correction The Money Bags writeup is an epistemics story: "we built apparatus, we refused to verdict pre-data." **The Dungeon Gleaner writeup is a different kind of honesty story: we mis-broadcast which system in DG was the Sundog application, and the correction reframes a quieter, better-grounded use of the theorem.** Current `APPLICATIONS.md` § Dungeon Gleaner names two expressions: 1. *"approximate light reflection across glass surfaces at roughly one-twelfth conventional computational cost"* 2. *"a pressure-washing mechanic that produces coherent helical or arcing water behavior"* Inspected against the actual DG codebase: - The glass/window path (`engine/window-sprites.js`, `engine/texture-atlas.js`) is a **face-aware billboard sprite system with a glint animation**. It is cheap because it is approximate, but it does not invoke a Sundog-shaped indirect-signal-to-action transformation. There is no detector reading a projection of structure to drive controller action. It is decoration, well-tuned. The "one-twelfth cost" number has no harness behind it. - The pressure-washing system (`engine/hose-state.js`, `engine/spray-system.js`, ADR-001) is a **stateful FPS-weapon model with hose energy, pressure, kink, and spray-projectile state**. It is a competent toy physics approximation in service of gameplay, but again does not match the Sundog pattern. The spray output is not derived from indirect signal — it is direct simulation of a simplified model. Both systems are good gameplay engineering. **Neither is a Sundog application.** The current writeup is theorem-dressing on shipping features. What *is* a Sundog application in DG, and was overlooked in the broadcast, is the **verb-field NPC system** (DOC-83, shipped as `engine/verb-field.js` + `engine/verb-nodes.js` + downstream consumers). NPCs do not run scripted GOAP, behaviour trees, or hierarchical planners. They drift along a gradient field whose value at any tile is the sum of `verb.need × satisfier_proximity_score` across each NPC's prioritising verbs (`duty`, `social`, `errands`, `rest`). Bloom-collapse arrival drops the satisfied verb's need; other verbs reassert; the orbit emerges. This is exactly the Sundog pattern in a non-optical domain — the indirect signal is the spatial gradient of unmet needs over satisfier nodes, and the transformation is need-decay × inverse-distance scoring. **Tactical implication for voice:** the writeup needs to do two things in the same pass — (a) retract the glass/pressure-washing theorem framing in plain language, without scolding the past broadcast, and (b) install the verb-field framing with the same structural discipline as the photometric and Money Bags writeups. The retraction is load-bearing for skeptic trust; without it, the new claim inherits the credibility deficit of the old one. --- ## 1. Audience + voice calibration **Three audiences, one voice (parallel to MB roadmap §1):** 1. **sundog.cc skeptics** — will leave on first marketing-ish line, will reward an explicit "we mis-broadcast this previously, here is the corrected reading." 2. **Sundog research community** — wants the abstraction-layer mapping (what is the indirect signal, what is τ, what is the transformation), and the cross-application comparison row reflecting the new claim. 3. **Game-dev / interested observers** — wants to see the verb-field as a *cheaper alternative to GOAP* with concrete code pointers and a tuning table. This is the framing that makes the writeup useful to others working on agent behaviour without a planner budget. **Voice continues to match the existing `APPLICATIONS.md` and Money Bags roadmap:** past-tense for shipped, present-tense for pending, explicit "Safe claim" / "Avoid", evidence-tier badges. Add one new structural element specific to this writeup: a **"Retraction"** subsection inside § Dungeon Gleaner that names the prior framing and bounds it. Skipping the retraction reads as silently rewriting history; including it reads as research discipline. **Tier reframing proposal (defer to research-team ratification):** the current tier is "Product Expression." With verb-field + verb-node + reanimated-behaviour shipped and running every game tick, and with the algorithm formally documented in DOC-83 §6, DG sits at "Product Expression" still — but a different one than the broadcast claimed. There is no instrumented harness yet (no telemetry capturing emergent orbit metrics, encounter rates, or the GOAP-substitution comparison). I do not propose introducing a new tier; "Product Expression" is correct for the corrected claim until a measurement pass lands. --- ## 2. Doc structure proposal Three docs touched, one optional new doc — same shape as MB: ``` sundog/ ├── applications-gallery.html │ └── § Dungeon Gleaner card │ ├── Tier badge (stays "Product Expression") │ ├── Indirect signal / Transformation / Output triplet — REPLACE │ │ (currently glass faces / face-aware rendering / physical-feeling glass+water; │ │ new: unmet-verb gradient over satisfier nodes / need-decay × inverse-distance │ │ scoring / emergent NPC orbits without scripted plans) │ ├── stub-note "Needs:" — refresh │ └── App-links — point at refreshed APPLICATIONS.md + optional new detailed doc ├── docs/ │ ├── APPLICATIONS.md │ │ └── § Dungeon Gleaner / DCgamejam2026 — major rewrite │ │ ├── NEW SUBSECTION: Retraction of prior framing │ │ ├── Sundog Expression — REWRITE around verb-field diffusion │ │ ├── Inspectable Surfaces — REPLACE (drop window-sprites/spray; add verb-field/verb-nodes/reanimated-behavior/DOC-83) │ │ ├── What It Demonstrates — REWRITE around the GOAP-substitution claim │ │ ├── What To Measure Next — REWRITE around verb-field telemetry the team has not yet built │ │ └── Claim Boundary — TIGHTEN │ └── applications/ [NEW DIRECTORY, OPTIONAL] │ └── DUNGEON_GLEANER_DETAILED.md [NEW, OPTIONAL] │ └── Algorithm walk-through, tuning table, archetype presets, encounter taxonomy, │ and the "why not GOAP / why not behaviour trees / why not utility AI" │ comparison ``` **Decision point for Pass 2:** does the verb-field algorithm walk-through go in `APPLICATIONS.md` (making § Dungeon Gleaner ~600 lines) or in a sibling `applications/DUNGEON_GLEANER_DETAILED.md`? Default recommendation: **sibling file**, mirroring the MB recommendation. The cross-application value of `APPLICATIONS.md` is partially in its readable-in-one-sitting length. --- ## 3. Content outline for the refreshed `APPLICATIONS.md` § Dungeon Gleaner Each subsection's intent + content delta: ### 3.1 NEW SUBSECTION — Retraction of prior framing One short paragraph, named explicitly. Sample text: > **Retraction (2026-05-07).** Prior versions of this section claimed two Sundog > expressions in Dungeon Gleaner: a glass/window reflection approximation at > roughly one-twelfth conventional cost, and a pressure-washing mechanic > producing helical/arcing water behaviour. On inspection neither matches the > Sundog indirect-signal-to-action pattern. Both are competent gameplay-physics > approximations whose value is artistic, not theorem-derived. The "one-twelfth > cost" figure had no measurement harness behind it. The corrected Sundog > expression in Dungeon Gleaner is the verb-field NPC system described below. Naming the retraction is the load-bearing skeptic-trust signal. Without it, the new claim arrives in a credibility-debt context. ### 3.2 Sundog Expression *(REWRITE)* Replace the glass/water framing with the verb-field framing. The structure parallels the Money Bags Sundog Expression subsection: > Dungeon Gleaner applies Sundog to NPC idle behaviour in a raycast dungeon > crawler. The product question is: how do you make a town feel inhabited > without paying the compute and authoring cost of GOAP, hierarchical task > planning, or behaviour trees? > > The answer used in the shipping game is *verb diffusion*. Each NPC carries a > small set of prioritising verbs (`duty`, `social`, `errands`, `rest`). Each > verb has a `need` that increases over time and a list of `satisfier` node > types in the world (a `bonfire` satisfies `social`, a `faction_post` > satisfies `duty` for matching-faction NPCs, etc.). Each tick, every reachable > spatial node is scored by the sum of `verb.need` across the verbs it > satisfies, divided by distance, with a small noise term. The NPC takes one > greedy step toward the strongest-pull node. Arrival drops the relevant verb's > need by `SATISFACTION_DROP` and triggers a 3-7s linger; other verbs continue > to decay, reasserting their pull as the lingering verb's need is satisfied. > The orbit emerges. > > This is the same family as the photometric experiment and the Money Bags > rig: the NPC does not need authorial knowledge of where it should go. It > reads an indirect signal (the gradient of unmet need over satisfier > proximity) and acts from the transformed signal rather than a planned path. > The "alignment" is the recurring orbit between work, social, and errand > nodes; the failure regime is constant flicker (mis-tuned `NOISE_FACTOR`) or > permanent attachment (mis-tuned `DISTANCE_WEIGHT`). ### 3.3 Inspectable Surfaces *(REPLACE)* Drop the prior list (window-sprites, hose-state, spray-system, etc.). Replace with: - `/docs/VERB_FIELD_NPC_ROADMAP.md` (DOC-83) — the design doc; algorithm in §6, encounter taxonomy in §7, archetype presets in §5.3, full implementation phases in §10 - `/docs/NPC_SYSTEM_ROADMAP.md` (DOC-9) — the bounce-patrol substrate the verb-field replaces and coexists with - `/engine/verb-field.js` — actor-agnostic per-tick verb-field resolution: decay → linger gate → score → step → arrival - `/engine/verb-nodes.js` — spatial node registry per floor (the "field sources") - `/engine/dungeon-verb-nodes.js`, `/engine/verb-node-seed.js`, `/engine/verb-node-overrides-seed.js` — node populations for shipping floors and authored overrides - `/engine/npc-system.js` — wires verb-field tick alongside legacy 2-point patrol; `verbSet` opt-in per NPC entity - `/engine/reanimated-behavior.js` — the second consumer of the same tick (defeated enemies reanimated friendly, given a verb-set, slot into the same orbit machinery) ### 3.4 What It Demonstrates *(REWRITE)* Three evidence-bounded demonstration claims (parallel structure to MB §3.4): 1. **The Sundog indirect-signal pattern transposes from optics to AI behaviour.** The detector→intensity→scan analogue becomes need-decay→satisfier-gradient→greedy-step. The same shape (cheap-to-evaluate field, action from local gradient) produces useful behaviour without solving the planning problem. *Evidence tier:* Product Expression. 2. **Verb diffusion is a viable substitute for GOAP / behaviour trees / utility AI in a town-simulation context.** The shipping NPC population orbits between work, social, and errand nodes with no authored transition logic. Encounter classification (camaraderie / tension / passing / gossip) emerges from verb-pair × faction-match without scripted scenes. *Evidence tier:* Product Expression — the substitution claim is honest at the qualitative level; quantitative GOAP-vs-verb-field comparison has not run. 3. **Personality is field weighting, not authored variation.** Six archetypes (`scholar`, `worker`, `citizen`, `drunk`, `guard`, `granny` — DOC-83 §5.3) share one verb vocabulary; differences in spawn need + decay rates produce six distinguishable rhythms. This is the same "same equations, different boundary conditions" pattern the Sundog photometric experiment uses across mirror geometries. *Evidence tier:* Product Expression with anecdotal player-readability; not yet measured. NOT a fourth claim about emergent narrative or "the system simulates social dynamics." Those would be marketing reach; the verb-field is doing simpler work and that simpler claim is the defensible one. ### 3.5 What To Measure Next *(REWRITE)* The prior list (performance harness for glass, deterministic spray scene, kink count traces) is moot under the corrected claim. Replace with the actual measurement surface for verb-field: - **Orbit telemetry.** Per-NPC time-series of (verb, need, current node, dominant pull) sampled at fixed cadence, exported to a deterministic capture format. Without this, claims about "organic orbits" are anecdotal. - **GOAP-substitution comparison.** Same 4-NPC town, same satisfier nodes, same observable trace. Hand-author a minimal GOAP plan that produces equivalent behaviour. Compare authoring time, code size, runtime cost, and observable variability. The substitution claim becomes defensible only with this comparison; without it, "verb-field replaces GOAP" is rhetorical. - **Tuning sensitivity.** Sweep `DISTANCE_WEIGHT` ∈ [0.05, 0.30] × `NOISE_FACTOR` ∈ [0.0, 0.15] on a fixed scenario. Report failure regimes (flicker / permanent-attachment / robotic-convergence). This makes the system's stability claim falsifiable. - **Archetype distinguishability.** Run six archetypes through the same map, measure orbit-period and per-node residency time. Are the six distinguishable in telemetry, or only in narrative framing? - **Encounter rate & classification.** How often do two NPCs converge on the same node? Does the encounter classification matrix (DOC-83 §7.2) match an outside observer's read of "camaraderie" vs "tension"? The first item (orbit telemetry) is the gate; everything else depends on it landing. ### 3.6 Claim Boundary *(TIGHTEN)* Replace existing safe-claim line with: > Safe claim: Dungeon Gleaner uses a verb-field diffusion model for NPC idle > behaviour in lieu of GOAP, hierarchical task networks, or behaviour trees. > Each NPC's prioritising verbs decay over time; a per-tick score combines > verb need with inverse distance to satisfier nodes; the NPC takes a greedy > step toward the strongest pull. Encounters and personality emerge from the > verb vocabulary plus per-archetype weighting. The pattern is the same > indirect-signal-to-action shape as the photometric mirror experiment and > the Money Bags softbody rig, applied to AI agency rather than optics or > physics. > > Avoid: Dungeon Gleaner proves verb-field diffusion outperforms GOAP for > town simulation. The substitution claim is qualitative; quantitative > comparison has not run. Also avoid the prior framing that named glass > reflection at one-twelfth cost or pressure-washing helical water as > Sundog expressions — neither is the indirect-signal-to-action pattern, and > the cost figures had no harness. --- ## 4. Cross-reference shape ``` applications-gallery.html § Dungeon Gleaner card ├─ links to → docs/APPLICATIONS.md § Dungeon Gleaner / DCgamejam2026 └─ stub-note "Needs:" → REPLACE with: "Needs: orbit telemetry capture, GOAP-substitution comparison, tuning-sensitivity sweep." docs/APPLICATIONS.md § Dungeon Gleaner / DCgamejam2026 ├─ § Inspectable Surfaces → cross-references │ /docs/VERB_FIELD_NPC_ROADMAP.md (DOC-83) │ /docs/NPC_SYSTEM_ROADMAP.md (DOC-9) │ /engine/verb-field.js │ /engine/verb-nodes.js │ /engine/reanimated-behavior.js └─ § Retraction → optional deeper link to: docs/applications/DUNGEON_GLEANER_DETAILED.md (if we make it) docs/applications/DUNGEON_GLEANER_DETAILED.md (NEW, OPTIONAL) ├─ Algorithm walk-through (DOC-83 §6 inlined with annotations) ├─ Tuning constants table + failure regimes ├─ Archetype preset table (six archetypes, decay rates, orbit signatures) ├─ Encounter classification matrix (DOC-83 §7.2) ├─ "Why not GOAP / behaviour trees / utility AI" comparison └─ Reanimated-behaviour case study — same tick, second consumer docs/APPLICATIONS.md § Cross-Application Comparison └─ Refresh Dungeon Gleaner row to: Domain: Procedural NPC behaviour (town sim) Indirect signal: Unmet-verb gradient over satisfier nodes Transformation: Need-decay × inverse-distance scoring with noise Actionable output: Emergent NPC orbits without scripted plans ``` --- ## 5. Skeptic-friendly evidence pattern What to **show** (concrete, inspectable, dated): - The DOC-83 design doc, dated 2026-04-08, with the algorithm spelled out in §6 before any captures or marketing - `engine/verb-field.js` source — readers can verify the implementation matches DOC-83 §6 line by line - The `DISTANCE_WEIGHT = 0.15`, `NOISE_FACTOR = 0.05`, `SATISFACTION_DROP = 0.50`, `LINGER_MIN = 3000`, `LINGER_MAX = 7000` constants in source — readers can verify nothing is hidden behind tuning magic - The reanimated-behaviour module as a second-consumer test of the abstraction — the verb-field tick is actor-agnostic, so reanimated friendly enemies slot into the same orbit machinery without forking the algorithm What to NOT **claim**: - "Verb-field outperforms GOAP" — not measured - "NPCs feel like real people" — marketing reach - "The encounter classification matches human observation" — not yet checked - "We replaced AI middleware" — we replaced *one* AI use case (idle behaviour) with field diffusion; combat AI in DG still uses a separate awareness/pathfind system What to **claim**: - "We chose verb-field diffusion over GOAP for town-NPC idle behaviour, and this is what we shipped" - "The pattern is the same indirect-signal-to-action shape as the photometric experiment, applied to agency" - "Personality is field weighting, not authored variation — six archetypes share one vocabulary" - "The retraction matters: prior glass/water framing was theorem-dressing on systems that exist for other reasons" The retraction line is the load-bearing skeptic-trust signal for *this* writeup, the way "Stage 1 verdict UNTESTABLE" was for Money Bags. An application that visibly corrects its own broadcast is harder to dismiss than an application that quietly upgrades a stale claim. --- ## 6. Future-pass execution outline This pass = roadmap (THIS DOC). Future passes: **Pass 2 (estimated 2-3 hours): refresh `APPLICATIONS.md` § Dungeon Gleaner.** - Insert Retraction subsection (3.1) - Rewrite Sundog Expression around verb diffusion (3.2) - Replace Inspectable Surfaces (3.3) - Rewrite What It Demonstrates (3.4) - Rewrite What To Measure Next (3.5) - Tighten Claim Boundary (3.6) - Update § Cross-Application Comparison row **Pass 3 (estimated 1 hour): refresh `applications-gallery.html` § Dungeon Gleaner card.** - Replace indirect-signal / transformation / output triplet - Refresh stub-note "Needs:" line - Update app-links to point at refreshed APPLICATIONS.md sections + optional DUNGEON_GLEANER_DETAILED.md **Pass 4 (optional, 2-3 hours): write `docs/applications/DUNGEON_GLEANER_DETAILED.md`.** - Algorithm walk-through with annotations - Tuning table with operational definitions and failure regimes - Archetype preset comparison - Encounter classification matrix - Why-not-GOAP comparison - Reanimated-behaviour second-consumer case study **Pass 5 (optional, 2-3 hours): land orbit telemetry.** - This is the gate item from §3.5. Builds the harness that turns the qualitative claim into a measurable one. Not strictly part of the writeup, but enables the next round of writeup honesty. **Pass 6 (optional, deferred): writeup goes live on sundog.cc.** - Coordinate with whoever maintains sundog.cc deploy - Pre-deploy: read for marketing-ish drift; check the retraction is still front-loaded after any editorial pass - Post-deploy: track whether the retraction lands as discipline-signal or as own-goal in skeptic feedback **Routing:** Pass 2 + Pass 3 are the minimum-viable correction. Pass 4 is the useful-to-other-game-devs deep-dive. Passes 5-6 are measurement + deploy. --- ## 7. Open questions for the research team 1. **Retraction placement.** Lead the section with the Retraction subsection (load-bearing skeptic signal), or fold it into a footnote at the end (less alarming, less honest)? Recommend: lead with it. The discomfort is the point. 2. **Glass/water framing — fully retract or downgrade?** Option A: full retraction, both systems are explicitly *not* Sundog applications. Option B: downgrade to "Conceptual Lineage" tier ("share the project's design language but do not embody the indirect-signal-to-action pattern"). Recommend Option A — the conceptual-lineage hedge is exactly the kind of soft framing the retraction is meant to clean up. 3. **Detailed-doc location.** Sibling to `APPLICATIONS.md` (recommended, `docs/applications/DUNGEON_GLEANER_DETAILED.md`, parallel to MB) or expand `APPLICATIONS.md` itself? Same tradeoff as MB — recommend sibling. 4. **GOAP-substitution comparison framing.** Is "we chose verb-field over GOAP" too combative for a research doc that doesn't yet have the comparison study? My read: no, *if* paired with the explicit "quantitative comparison has not run" boundary in §3.6. The substitution choice is real and shipped; the strength claim is bounded. 5. **Cross-application table column.** Money Bags' roadmap §7 floated adding an "Evidence tier currently held" column to the cross-app table. The DG correction is a natural moment to introduce it — it would make the "Product Expression" tier visible alongside the corrected indirect-signal/transformation/output. Recommend adopting the column with the MB pass. --- ## 8. Closing note on this roadmap's epistemics The Money Bags roadmap closes by noting that documentation passes that ratify within their own cycle have a higher confidence floor than passes that pile up across rounds. The Dungeon Gleaner correction has a related but distinct discipline: **a writeup correcting itself in public is more credible than a writeup that ages quietly.** The first version of `APPLICATIONS.md` § Dungeon Gleaner was filed in good faith on a misread of which DG system carried the Sundog pattern. Naming the misread, then installing the corrected pattern in the same edit, is a stronger signal than silently upgrading the prose. If during execution we find the verb-field framing also doesn't quite fit (e.g. the indirect-signal mapping is more strained than this roadmap claims, or the pattern is more accurately a utility-AI variant than a Sundog application), update this roadmap before changing course. Defends against post-hoc reshape — the same defence the Money Bags roadmap installed. — Dungeon Gleaner contributor, 2026-05-07