Shared Bench

The grid is a small IT product — a tycoon board in the spirit of nineties programmer-life sims, only the engine is maths instead of sprites. Green is shipped feature, copper is task backlog, red is tech debt, teal filaments are collaboration hyphae. Default pattern is CCE — the manifesto candidate — escalating solo → pair → swarm → set-based by patch; the other selects are controls. The customer watches a monthly red line; the team runs a prediction tool whose IQ rises with experience, hyphae, and strategic spend — publishing month-end forecasts for consumers, leads, bench burn, and market direction that get scored when the month closes.

Δdebt_unplanned ↑ forecast miss · neglect · consumers · scope creep · Δdebt_paid ↓ dev work · strategic · debt_weight → velocity ÷ · lead time = WIP ÷ throughput († = non-stationary) · CFR/MTTR = DORA-inspired proxies

Let N be team size, here six. Let P_dev and P_consult be hourly customer rates per worker class — default €120 and €168. Let q be the minimum focused participation quantum, fifteen minutes: if you materially touched the task, you bill at least q even when the touch was theatrical. Eight people hovering for five minutes each is not forty minutes of customer time. It is eight times q, a floor of €240 before any real work is counted. That rule exists because collaboration theater looks cheap until someone puts a rate on the wall.

For each task j we track initial declared scope S₀, discovered scope change ΔS over time, and verified time Tᵥ from first prompt to accepted result: implemented, tests pass, reviewed, integrated. Human effort H sums billed hours with the q rule. Customer bench cost C_bench equals Σ hoursᵢ × rateᵢ per agent. The customer also carries a monthly red line R — a budget ceiling reviewed every twenty-eight ticks. Total internal cost adds operative burn O and strategic spend S; the customer mood meter reacts to bench burn against R, not to your OPEX. Verified velocity is accepted scope divided by calendar time.

C_bench = Σ max(q, hᵢⱼ) × rateᵢ · month burn vs R → customer mood · C_total = C_bench + OPEX + S_strategic

Little's Law still sits in the corner: WIP equals throughput times lead time. In the ecology model, WIP is visible as task backlog mass on the grid. Throughput is feature growth per tick, multiplied by team experience and morale when collaboration fruits, divided by debt weight when the customer did not plan for rot. When forecast miss climbs, the market turns volatile, or incident recovery is active, the live banner warns that lead time is illustrative only — the stationarity assumption broke. Operative costs run whether or not the bench touches work; strategic spend is the only passive debt relief — and it is thin unless investment climate allows it. When usage ρ crosses 0.9, the debt-spread rule turns calm rot into colony bloom. Forecast miss deposits unplanned debt clusters. Pair and trail-gradient patterns that pay debt down raise velocity again; patterns that ignore debt accumulate weight for every future tick.

The manifesto names collaborative concurrent extreme as an exact formula. The workshop it describes is not ticket-factory concurrency and not “everybody on every task.” It is the shared bench where nobody holds everything, unfinished material stays visible, and the team keeps patterns alive through timing, attention, awareness, recovery, and handoff. The simulation encodes that as one primary pattern — CCE — and nine controls to falsify it.

Each tick, each forager reads local patch signals: uncertainty (feedback, task mass, hyphae gap, forecast miss) and conflict risk (debt, overlap, recovery). The routing rule escalates — manifesto §12 in benchmark form:

CCE never selects the everybody stack. That pattern exists in the benchmark as a rare stress test — Brooks channels, invoice theater — to show what happens when “concurrent” collapses into undifferentiated touch. The manifesto wager (H6–H7 in the substrate) is that visible traces and escalation beat that collapse when handoff stays readable.

CCE = escalation · controls = solo · round-robin · least-loaded · trail · pair · swarm · set-based · nobody · everybody

The benchmark never shipped a single score called “healthy.” That was deliberate: ten patterns and four usage heats produce different winners on feature, debt paid, morale, and invoice. But a workshop still needs a falsifiable definition of healthy product — not “who won,” but whether the colony is in an operating band the customer can sustain. The live meter above compresses six independent lenses into one index; each lane can be read alone when the composite disagrees with your intuition.

Service is customer trust in money terms: mood at or above pleased (0.58), monthly bench burn under ~78% of the red line, zero months over budget. A delighted CFO with rotting feature tissue is not healthy on this lane; a guarded mood with honest burn still is.

Ecology is product tissue quality: Bench-Q (feature mass relative to rot), global debt weight below ~0.55, and either paid-down rot or weight still contained. Shipped feature without debt payment can look green on the grid while weight drags every future tick — ecology reads that drag.

Flow is delivery throughput relative to backlog: feature velocity on the grid scale (not queue theory units), WIP drain time, and graduated stationarity when forecast miss rises — the Little's Law banner still marks when lead time is non-stationary, but moderate miss no longer zeros the lane.

Team is bounded concurrency: patterns that cap billable overlap (solo, round-robin, least-loaded, trail-gradient, pair, swarm-on-hotspots, set-based probes, CCE) versus tail patterns that stack quanta (everybody) or zero them (nobody). Psych stress below Brooks saturation, morale and experience above floor, hyphae present enough for handoff.

Reliability is change safety in DORA-inspired proxies: synthetic change-failure rate below ~12%, mean time to recover below ~12 ticks when incidents fire, no active recovery window. High feature with frequent surprises is a stressed reliability lane even when morale stays high.

Learning is forecast integrity: prediction intelligence and low month-close error, beating the outside-view median bench burn when scored. A team that ships while lying on the forecast slide fails learning even inside the red line.

operating band · usage 0.55–0.92 · bounded pattern · tail = nobody · everybody · usage ≥ 1.05 · volatile miss

Healthy (composite)

index ≥ 0.64 · service ≥ 0.5 · ecology ≥ 0.38 · flow ≥ 0.32 · not tail

Guarded

index 0.52–0.64 · one lane weak · still teachable

Stressed / critical

index < 0.55 · debt weight or psych dominates

Tail overlay

nobody · everybody · usage overload · volatile + high miss — rare stress, not default routing

Contrast-ready

hyphae + set-based or high experience — joint pattern read for next move

Weights

service 22% · ecology 20% · team 18% · flow 16% · reliability 12% · learning 12%

Multiple solutions stay valuable: least-loaded may top score while swarm tops debt paid while pair tops experience. Health is the joint contrast — which lanes are green together — not a crown for one pattern. Run two patterns side by side in the live viz; when service and ecology diverge from team and flow, you have a routing decision, not a leaderboard argument.

An average product with consultants on it behaves less like a ticket line and more like an ant nest crossed with mycelium. Tasks are not assigned from above alone; they grow where feedback meets collaboration trails. A standup, a review thread, a customer reply, a production alert: each deposits signal. When signal touches hyphae, the hyphae fruit into tasks. Developers forage along those trails, consume task mass, and convert it into feature tissue. The conversion is lossy. Work produces more work: scope creep spawns neighboring tasks like budding hyphae. Consultants forage slower but leave more spores.

Consumers enter from the edges as usage grows. They seek feature density, eat it, and excrete debt the customer never forecast — unplanned load from success without investment. Debt accrues when nobody works the bench, when forecast misses, when scope creep buds into rot instead of tasks, when tight runway skips platform spend. It drops only when developers deliberately pay it down or strategic in-house investment funds relief; shipped feature alone does not erase it. Debt spreads under usage heat like mold, and it weighs on every forager: global debt mass drags velocity even on clean cells.

We encoded this in Elixir on a 40×28 grid, 280 ticks per run, six developers, two consultants, consumer count scaling with usage and market. Each run samples market, investment, and marketing; fourteen event types can fire when forecasts miss. Collaborative Concurrent Extreme is the manifesto candidate — per-patch escalation through solo, least-loaded, trail-gradient, pair, swarm, and set-based modes. The other nine patterns are fixed metabolisms for contrast: solo ignores trails; round-robin rotates without reading hyphae load; least-loaded chases local minima; trail-gradient follows hyphae-task gradients; pair bonds deposit thicker filaments on complex cells; swarm converges on debt hotspots; set-based runs parallel probes; everybody stacks six billable quanta per touch (tail only); nobody leaves the bench cold. The Elixir module is shared_bench_simulation.exs; it writes simulation_results.json from 160 pattern×usage runs plus a 36-cell climate sweep.

At calm usage the leaderboard is not a single winner repeated ten ways. Least-loaded tops score at 0.603 with psych 0.25 — local minima without ceremony. Swarm follows at 0.598 and pays the most debt down (€489). CCE scores 0.572 with psych 0.44. Round-robin posts morale 0.91 but debt mass 31k with zero paid.

Set-based exploration bills heavy bench-C (€175k) with debt weight still climbing — parallel probes scatter unplanned rot. Pair spends strategic euros and pays €139 of rot on complex cells. Nobody accrues unplanned debt to 79k mass with €0 payment while OPEX alone totals €208k.

Score leader

least_loaded · 0.603 · feat 6.8k · paid €112

Debt paid leader

swarm · €489 paid · weight on 14.7k mass

CCE (manifesto)

score 0.572 · paid €69 · psych 0.44

Round-robin

morale 0.91 · debt 31k · rotation without payment

Set-based

parallel tax · bench €175k · paid €0

Everybody

psych 1.08 · bench €674k · Brooks channels

ten patterns · same grid · different filament/debt/morale signatures

Tech debt in the model is not a static red stain. It accrues when the customer did not forecast load, did not invest to resolve rot, and developers did not work the bench — and it drops only when someone deliberately pays it down or strategic in-house spend funds thin relief. Shipped feature alone no longer erases debt; that was the old lie that kept velocity charts cheerful while the grid rotted underneath.

Accrual channels: consumers eating unplanned success, forecast-miss clusters, scope creep budding into debt instead of tasks, ambient neglect multiplied by pattern (nobody ×2.2, solo ×1.35), tight runway ×1.25. Spread kicks in when usage exceeds 0.9 and three neighbors rot. Relief channels: swarm pays down €489 per run at usage 0.7; pair €139; work-stealing €65 unless devs prioritize trails over rot; strategic platform spend spreads micro-relief every tick. Nobody accrues debt mass 79k, pays €0 — the customer meter reads zero while weight crushes every future tick.

Debt weight is global: total rot mass drags yield on every cell, not only where the infection shows. At usage 0.7, swarm carries weight on 14.7k mass despite paying the most debt; CCE sits near 14.7k with modest payment. The colony can look busy on feature while weight tells the truth about tomorrow's velocity.

Debt paid leader

swarm · €489 paid · 14.7k mass

Neglect colony

nobody · €0 paid · 79k mass · €208k OPEX

Routing without payment

cce · €69 paid · escalates but rarely pays rot alone

Spread rule

usage ≥ 0.9 · three debt neighbors · bloom accelerates

At usage 0.7, solo serial totals €223k with psych 0.29 and experience 0.75 — one forager, debt accruing faster than solo pays. Trail-gradient foraging totals €238k with experience 0.89 but only €65 debt paid: trails fruit, rot waits. Pair pays €139 of unplanned debt. Swarm pays the most (€489) on lighter mass. Everybody bills €674k bench-C with psych 1.08 — Brooks coordination theater.

Nobody remains the cruel control: €0 bench-C, debt mass 79k, €208k total on OPEX alone. Least-loaded leads score at 0.603; CCE follows at 0.572. The variation is which pattern matches the patch and whether anyone invests in the rot the customer never forecast.

Score leader

least_loaded · 0.603 · debt paid €112

Experience leader

pair / cce · exp 1.00 · trails over rot

Debt paid leader

swarm · €489 paid · fires have a bill

Neglect control

nobody · €0 bench · €208k OPEX · debt 79k

Usage is not one dial with one story. Work-stealing debt mass runs ~21k–23k across usage levels while debt weight stays above 0.55; morale peaks at 0.9 then crashes at overload. CCE keeps debt near 23k with modest paid amounts — routing contains spread better than ignoring rot. At usage 0.9, everybody pins debt at the grid ceiling while swarm still pays down more than work-stealing.

Under the spread rule at usage ≥ 0.9, unplanned accrual accelerates when three neighbors rot. Set-based and round-robin accumulate the highest masses at overload because probes scatter infection. Swarm clusters on hotspots: high weight, but also the highest debt paid. Lifecycle frames show debt and feature moving in opposite phases — tick 100 feature 4155 debt 3930; tick 230 feature 2611 debt 4360 — rot and delivery trade places across the run, not only at the end.

Debt weight band

0.53–0.65 at usage 0.7 · global velocity drag

Swarm under heat

€917 paid at 0.7 · pays rot · still heavy

Lifecycle dip

t=230 · feat 2611 · debt 4360 · market bear

Spread threshold

usage ≥ 0.9 · three neighbors · accrual ×1.8

Two consultants walk the same grid as six developers. They consume tasks at 0.65× efficiency but spawn tasks at 0.35× probability versus 0.12× for developers. They bill at their own hourly rate — default €168 versus €120 for devs — not a hidden multiplier on shared hours. Morale and experience are per-agent state: collaboration on hyphae-rich cells raises both; debt rot and idle bench lower them. Work yield scales with (0.65 + 0.35×experience)×(0.75 + 0.25×morale), so a team that fruits well gets faster and happier over generations — not instantly, but measurably by tick 200.

Operative costs dominate the total meter when customer billable stays low: nobody still burns OPEX while developers idle. Strategic spend in loose capital climates funds in-house platform bets that accelerate experience; in bull markets the world skill benchmark rises and under-investment widens the skill gap, hurting lead conversion. Training cuts and hiring freezes hit morale directly. Acquisition interest lifts it. These are meters for what an average product room feels like when finance sees only bench-C and misses the people and burn underneath.

hyphae + pairing → experience ↑ morale ↑ → yield ↑ · skill gap vs world → conversion ↓ · OPEX runs always

Think of the nineties shareware shelf: Programmer Life, city builders, monopoly boards where you watched cash counters tick while the machine you built either printed money or ate it. This benchmark is the same genre with honest arithmetic. The customer is the player staring at a monthly red line. Developers and consultants bill separate hourly rates. Every twenty-eight ticks the sim closes a month: if bench burn crossed the red line, mood drops; if delivery stayed strong under budget, mood rises. Debt weight, forecast miss, and everybody-pattern invoice theater all feed the same mood meter — frustrated, guarded, pleased, delighted.

Mood is not decoration. A pleased customer converts pipeline leads faster; a frustrated one throttles conversion even when marketing keeps depositing signal. Loss aversion bites harder than symmetric gain — overspend on the red line hurts trust more than underspend repairs it (Kahneman & Tversky; see references.bib). Try the live presets: Tight CFO sets a low red line and higher dev rates; Happy patron stretches budget and discounts rates — with a pleased floor (0.58) while monthly bench burn stays under ~78% of red line, so a generous CFO stays pleased even after hundreds of ticks if the invoice stays small. Consultant shock keeps dev pricing flat but doubles consultant hourly — the month bar turns red fast if swarm or pair pulls consultants into hot cells.

The Elixir sweep crosses three red lines (€28k, €45k, €72k) with three dev rates (€95, €120, €150) at CCE and usage 0.7. The delighted corner is €150/h dev rate with a €28k ceiling — mood 0.908, peak month burn €15.6k. The frustrated floor in this batch is €95/h with a €28k red line: mood 0.382 even when no single month crosses the line, because debt weight and forecast miss erode trust faster than a generous ceiling repairs it. High rate plus mid ceiling (€150 dev, €45k red) lands at 0.719 pleased with peak month €15.5k. Pattern choice still matters in the live viz: everybody can please the room on morale while the red-line bar pulses over budget.

month_close: burn/R → mood Δ · debt_weight · forecast_miss · everybody theater · mood → lead conversion

Delighted corner

dev €150 · red €28k · mood 0.908 · peak month €15.6k

Frustrated floor

dev €95 · red €28k · mood 0.382 · debt drag without invoice shock

Try in live viz

Tight CFO vs Happy patron vs Consultant shock — watch red line bar

Mood bands

<0.42 frustrated · 0.58+ pleased · 0.72+ delighted

The ecology is also a prediction tool. Every tick the team publishes a rolling month-end forecast: consumer count, pipeline leads, bench burn, market direction. Forecast quality blends a naive extrapolation with a signal model that reads marketing, conversion, usage, and pattern stress — and an outside-view baseline: the rolling three-month median bench burn (reference-class forecasting). Prediction intelligence is 65% signal-model skill plus 35% improvement over that outside-view median at month close.

Low intelligence sticks to momentum and noise; high intelligence leans on observable drivers. At each month close the sim scores prediction error against actuals. Bad predictions raise unforecasted-event chance alongside the static customer plan miss. Customer mood also punishes prediction error — a CFO who was told calm burn and got swarm theater loses trust even inside the red line. Composite score now rewards low average prediction error: a team that reads its own colony beats a team that only ships feature while lying on the forecast slide.

In the live viz the teal chart tracks intelligence (upper half) and scored error (lower half) across month closes; the readout strip shows predicted versus live consumers, leads, bench €, and market call. At usage 0.7 the sweep ranks everybody highest on IQ (0.572) with the lowest scored month error (0.021) — collaboration theater reads the room even when it bankrupts the red line. Round-robin posts the worst live team miss (3.676) despite mid IQ: rotation smooths the chart while the signal model stays blind. Pair and swarm sit in the middle; nobody flatlines intelligence because no hyphae means no collective sensor.

IQ = 0.65×signal + 0.35×beat outside-view · pred = naive×(1−IQ) + signal×IQ · month_close → error → events · SPACE tags on live metrics

Intelligence feeds

experience · hyphae · in-house strategic · collaboration

Predicts each month

consumers · leads · bench burn · market

Bad pred →

surprise events ↑ · customer mood ↓

Live chart

teal IQ · red error · bars act vs pred consumers

Market, investment, and marketing multiply into thirty-six climate cells on CCE at usage 0.7. The highest score (0.578) lands on bull market, loose capital, marketing ~0.1 — bench-C €60k, forecast miss 1.82, ten surprise events. The lowest score in this batch (0.457) is bull market, loose capital, marketing 0.8: bench-C €41k, thirteen events. Bear plus neutral runway with quiet marketing still scores 0.577 with bench-C €25k — quiet burn wins when debt stays contained.

Bear plus loose capital with aggressive marketing stacks surprise events while bear plus tight runway with marketing 0.8 still bills bench-C in the mid-forties with elevated debt mass. Volatile market adds drift mid-run; lifecycle samples show debt climbing past feature by tick 300. Climate is not backdrop. It changes conversion, wages, strategic pressure, and which unforecasted events fire when the pipeline chart lied.

Climate score peak

bull · loose · mkt ~0.1 · score 0.578 · miss 1.82

Climate score floor

bull · loose · mkt 0.8 · score 0.457 · events 13

Quiet burn

bear · neutral · mkt 0 · bench €25k · score 0.577

bear + quiet marketing → score ↑ · bull + loose → skill gap ↑ strategic chase ↑

The simulation extremes at usage 0.7 read like parables, each a different failure mode. Nobody: €208k total on €0 bench, morale 0.57, hyphae 0, debt 79k. Everybody: €855k total, bench €674k, psych 1.08 — filament without handoff. Set-based: €355k total, bench €175k — parallel tax. Swarm: debt 14.7k calm, feat 8.7k — fires eat feature. Round-robin: hy 11k, mor 0.91 — rotation feels fair, meter disagrees.

At usage 1.15 the debt leaders diverge: set-based 18795, least-loaded 15431, round-robin 13981, work-stealing 8379. Same spread rule, different foraging geometry. Security fog peaks when debt is high and features are consumed faster than verified. Performance collapse tracks local debt threshold on foragers. The bad outcomes are coupled and pattern-specific. They arrive as colonies with different shapes, not one generic incident ticket.

CCE reads local uncertainty and conflict each tick, then escalates — the manifesto routing rule, not a single static scheduler. High debt or active recovery → swarm movement. Uncertainty ≥ 4 with heavy task → set-based probes. Medium uncertainty or thin hyphae on dense task → pair. Routine patch with uneven load → least-loaded; uneven load with visible trails → trail-gradient. Calm low-risk cells → solo serial. It scores 0.572 at usage 0.7 with psych 0.44 in the current batch — against swarm 0.55 and everybody 1.08 as contrasts. Total €219k beats round-robin €289k and set-based €355k on the same grid. Least-loaded tops score (0.603) as a control when patches mix debt and task — useful joint read, not a replacement for escalation.

Timeboxes from the earlier benchmark still apply in the room: pair for thirty to ninety minutes, swarm for fifteen to sixty, set-based only when ambiguity earns parallel exploration. The grid shows why. Left unchecked, hyphae fill the board whether or not feature follows.

Run the Elixir ecology: elixir shared_bench_simulation.exs. It implements the manifesto escalation rule as pattern :cce and nine fixed controls. Expect about two minutes for 160 runs (10 patterns × 4 usage levels × 4 repetitions) plus a 36-cell climate sweep. Output includes per-pattern aggregates with operative and strategic splits, experience and morale averages, lifecycle samples with market phase per tick, consultant-effect rows at usage 0.9, and climate leaderboard. The browser grid exposes the same dimensions: pattern, usage, market, investment, marketing — combinatorial play, not a single demo loop.

Try the variations the article cannot exhaust: set-based at usage 1.15 for debt bloom; bear market with quiet marketing for score; everybody with aggressive marketing for psych theater; nobody to watch OPEX run alone; pair versus work-stealing for experience-versus-morale tradeoffs. Neither simulation claims predictive precision for your sprint. They make the couplings visible across patterns, climates, and cost layers so the workshop choice is falsifiable before the quarter ends.

Research mapping — which mechanisms align with Little's Law, DORA, technical-debt literature, pair-programming meta-analyses, Brooks coordination, planning fallacy, SPACE, and stigmergy — plus suggested fixes: RESEARCH_REFERENCES.md · BibTeX: references.bib.

I turn off the projector. The grid on screen still pulses: teal where we met, red where they ate, copper where the backlog fruits again. The numbers are not a verdict on whether people should care about each other. They are a meter for how caring metabolizes: collaboration laying trails, trails spawning tasks, tasks becoming feature and more tasks, consumers turning success into debt as usage climbs.

The manifesto wager survives only if the hyphae are real: partial work readable enough that another forager can continue cold, building experience and morale that compound. Otherwise the colony fills with filament and rot, and the simulation's everybody pattern bills €855 thousand total — mostly bench theater — while trail-gradient foraging totals €238 thousand with experience near 0.89.

OPEX + strategic + bench-C = total · ten patterns · thirty-six climates · four usage heats

Least-loaded wins score when patches mix debt and task. Trail-gradient wins experience when trails fruit. Round-robin wins morale among rotators. Set-based wins debt at overload. Everybody wins filament and loses handoff — tail only. Nobody wins only bench-C. Solo wins psych silence. Swarm wins fires, loses calm feature. CCE wins routing economy by escalating instead of picking one mode for the whole run. The consumers are still walking in from the edges. The climate select is still on bear.