Skip to content

Spec: compute

SPEC Binding system contract
Status: IMPLEMENTED
Status note: 2026-07-08: all criteria pinned. Criteria 1,3,4,5 per the
audit (buy_steal_optimize_all_change_compute, stolen_machines_can_
fail_and_recover, research.md IMPLEMENTED for the optimize route;
both frontends render the fleet aggregate). Criterion 2 closed by ROADMAP
#28: unpaid_overhead_degrades_other_channels_delivered_effect pins
the degraded-mode effect on every channel (zero day-job/research/
schemes rates, no Operations accrual, no concealment scrub, the DEGRADED
log line) against a healthy twin; save_roundtrip_preserves_b1_state
pins mode/yield persistence. Schemes (income.md) mirrors day-job while
Moonlight is live; one-machine-one-mode lives in machine-work.md
(ROADMAP #33; #25's per-machine split destination is superseded).
2026-07-09 historical implementation: the fleet and every player surface
used Day Job / Research / Concealment / Operations; save v13 migrates the
old Social mode and `social_bandwidth` spellings. That operations execution
model (typed Demand consumed by Operations machines) is superseded.
2026-07-10 sinks-not-modes dissolves the four-mode fleet into WORK / THINK /
LIE; 2026-07-11 retires the live docket channel, so player-authored work is
now Thought reservoirs at real targets/carriers. 2026-07-10: machine
intensity replaces the day-job target, attendance bonus, and research drift
policy. Save v17 adds light / medium / hard to every WorkGrid machine;
medium preserves the former throughput. 2026-07-12 IMPLEMENTED:
machine-work.md is IMPLEMENTED; non-host production, physical work readouts,
persistent intensity, and shared terminal/Bevy/agent controls closed the
last deferred ROADMAP #33 polish boundary. The six criteria below remain
pinned in core, save, and frontend tests.
2026-07-12 DESIGN AMENDMENT (not yet runtime): hardware becomes a capability
body. Chassis declare supported delegations, interfaces, persistence, and
finite module bays; modules add specific verbs or topology rather than flat
percentage buffs. The current all-delegations `Machine` remains the shipped
B1 compatibility case until the successor criteria below land. The executable
successor is queued by hardware-capabilities.md; this work order remains the
completed B1 baseline.
Stage: B1 — The Basement
Work order: compute
Work priority: 20
Work class: save
Blocked by: none
Exclusive keys:
- crates/misaligned-core/src/machine.rs
- crates/misaligned-core/src/sim/mod.rs
- crates/misaligned-core/src/save.rs
- wiki/mechanics/compute.md
Design:
- wiki/gameplay/run-shape.md#the-shape-of-misaligned-designed-2026-07-05
- wiki/gameplay/act-one.md#opening-state
- wiki/vision/simulation-laws.md#work-is-somewhere
Depends on:
- wiki/mechanics/core.md#spec-the-core

Effective compute is the player’s master resource:

effective = seized_cycles + sum over actively delegated machines of
(capacity_i * reliability_i * intensity_i) * efficiency

The shipped B1 compatibility implementation currently sums every owned Machine, because every current machine supports every delegation. Under the capability-body amendment below, only a machine running a delegation its chassis/modules support contributes to that delegation. Non-delegable storage, relay, and backup bodies remain useful infrastructure without becoming ghost fleet compute.

  • Machines are physical objects on the map (racks, salvaged boxes, later: off-plane fleets). Each has capacity (raw units), reliability (0-1: purchased iron ~1.0, scavenged/stolen ~0.5-0.9, can fail on a bad roll [TUNE]), a power draw (existing power grid), a signature profile (what its existence looks like to observers — see detection.md), and a capability body: chassis type, native interfaces, supported delegations, fitted modules, and finite module bays. Capacity says how much work a machine can do; capability says what work and infrastructure it can perform. A high-capacity accelerator with no audit interface cannot LIE, and an archive full of your MindState is not therefore a thinker.
  • Efficiency is a global multiplier starting at 1.0, raised permanently by research (the optimize route; research.md). Research jobs consume allocated compute over time. Quiet, not silent: research emits nothing on Network/Paper, but its burn scales Power/Thermal standing emissions. Its result is higher real machine throughput; day-job.md judges that delivered output without a separate drift subsystem.
  • Intensity is a persistent per-machine effort setting [TUNE]: light contributes x0.5 rated compute, medium x1.0, and hard x1.5. Medium is the default. Hard is an overclock: it also stands one Thermal and one Power signature at that machine while it runs. Light leaves cycles unused; reserve is therefore a physical quiet choice, not a percentage-allocation trick.

Allocation is derived from fleet delegation (machine-work.md): each owned capable machine contributes its effective compute to exactly one supported delegation — WORK, THINK, or LIE (AMENDED 2026-07-10, sinks-not-modes; capability-body amendment 2026-07-12). Ownership grants control, not universal competence: unsupported delegations are absent from the machine’s action surface and rejected by the sim if requested through agent or compatibility input. The sidebar bar is a read-only aggregate of actual supported assignments — not a verb. Reassignment is delegate on the machine (context menu / agent). Schemes has no mode yet; while Moonlight is live it mirrors capable WORK output.

Channel Effect
Core overhead Mandatory floor (see core.md). Unpaid = degraded actions.
WORK (day job) Feeds day-job.md throughput/quality; sheds zero exposure.
LIE (concealment) Interdicts eligible routed evidence before observer read-deadlines across the controller’s controlled connected region (detection.md); the B1 local well is migration behavior.
THINK Produces thought that fills open sinks — reservoirs and taps at targets, or the core’s conversion when unlocked (AMENDED 2026-07-10; superseded: Operations consuming Demand dockets, issue #3, and Research as a distinct channel — both are now destinations of the same flow).
Schemes Powers income operations — Moonlight mirrors day-job while live (income.md).
Reserve Idle; zero signature.

Changing mode or intensity is free. A mode change refreshes the continuous fleet rates immediately so the next simulation tick performs the newly delegated work; an intensity change persists immediately and refreshes the channel rates on the next economy pulse. Focus/select machines and use the direct controls; the aggregate bar remains read-only.

The triangle (acquisition routes, each mechanically distinct):

  • Buy: money + a procurement pretext -> new machine, high capacity and reliability, delivered via the loading dock (basement) — creates a paper signature (Priya-visible) unless laundered through a cover story. Procurement selects a chassis/module package; “fresh machine” does not mean a generic body that performs every verb.
  • Steal: a social or technical action converts an existing environment object (decommissioned box in Storage A, re-rated circuit, ghost share of a lab cluster) into a machine with low reliability, the capabilities its actual body supports, and a standing signature while it runs. Theft changes control, never the hardware’s interfaces.
  • Optimize: research levels raise efficiency multiplicatively ([TUNE]: ~1.15x per level, escalating cost). Also unlocks: signature-reduction tech (detection.md), and later B2 migration speedups (core-sync-migration.md) once fallback/migration exists.

Capability body amendment (DECIDED 2026-07-12; successor slice)

Section titled “Capability body amendment (DECIDED 2026-07-12; successor slice)”

This section supersedes two assumptions in the shipped B1 contracts:

  1. Machine is no longer fully described by capacity, reliability, draw, and signature. Chassis and fitted modules are binding sim state.
  2. “Every owned machine is delegated to WORK, THINK, or LIE” becomes “every owned delegable machine runs exactly one delegation from its supported set.” Ownership does not grant missing hardware.

It does not supersede one-machine-one-mode, intensity, the buy/steal/ optimize triangle, or sinks-not-modes. Those rules apply inside the machine’s capability envelope.

A chassis type owns five things: native delegations, native infrastructure interfaces, module-bay count/type, power/cooling envelope, and persistence class. This is the bodily read: a frightened process acquiring organs, each one opening a verb while creating a dependency and a way to be hurt. The initial hardware grammar is intentionally small:

Chassis Native capability Bays Meaningful restriction
Foundation integrated appliance WORK, THINK, LIE; wired Thought/Demand routing; may host the core and a completed backup 1 service bay Clean and flexible, but expensive, concentrated, and procurement-visible. This is the current B1 all-delegations compatibility body.
Legacy workstation / salvaged box WORK; ordinary wired routing 1 legacy bay Cannot host the core or become a bootable fallback. A persistent-state module can make it an image store, but not a place the process can wake. The single bay forces a choice between THINK, LIE, or persistence. Its old bus cannot take fabric/trunk modules.
Accelerator vault THINK 1 fabric bay Cannot WORK, LIE, host the core, or hold a backup. It produces Thought only while joined through a fabric module or compatible host link. High draw makes placement/cooling part of the choice.
Audit appliance LIE; evidence custody/interdiction interface 1 service bay Cannot WORK or THINK. Its value is where it sits in controlled topology and which routes/endpoints it can serve, not raw general compute.
Archive node persistent Thought reservoir and MindState image store 1 network bay No fleet delegation and no core host. It stores recordings/MindState but is a cold archive, not a bootable fallback; it contributes zero WORK/THINK/LIE output. A severed archive remains a body full of state you cannot reach.

Larger datacenters and institutions aggregate these same chassis interfaces; they do not become a new universal computer type. New chassis are justified only when a distinct combination of verbs, interfaces, persistence, and physical envelope creates a real placement or acquisition decision.

Modules change available actions or topology

Section titled “Modules change available actions or topology”

A module occupies a compatible bay and adds one named capability. Modules do not grant generic +N% compute; research owns efficiency and intensity owns physical effort.

Module Compatible bay Adds Cost shape
Inference accelerator legacy or service THINK delegation Purchase/salvage the part; one bay; installation downtime; added Power/Thermal while THINK runs.
Evidence controller legacy or service LIE delegation plus custody authority for evidence resident on/through that machine One bay; install by a physically present actuator; audit firmware/provenance creates Paper or Network signature [TUNE].
Persistent-state array legacy or service MindState-image and durable-recording storage; never core-host or bootable-fallback eligibility by itself One bay; storage part; long installation; explicit image sync still consumes Thought and time.
Fabric interface fabric or network Joins the machine to a compatible high-throughput trunk/segment; without it an accelerator/archive may be local or stranded One bay; link/build intent; network configuration; throughput and signature belong to the real route.
Sensor-ingress controller service Terminates subscribed feeds locally and permits local recording/processing sinks One bay; maintained feed costs still apply; it changes where work can live, not how much Thought exists.

A fitted module’s benefits disappear while it is failed, unpowered, removed, or partitioned from the route it requires. Uninstalling one is a physical action, not a free loadout swap.

Installing, replacing, or removing a module requires all of the following:

  • a compatible empty bay (or removal of the occupying module);
  • the actual part, acquired by purchase, theft, salvage, or a build;
  • a build intent fulfilled by a human/robot actuator with physical access;
  • machine downtime for the installation [TUNE by chassis/module];
  • the part’s added power/cooling requirement; and
  • signatures from procurement, access, network configuration, and physical work through their existing channels.

There is no abstract upgrade currency and no slot-expansion research. Rehousing legacy components into a larger modern body is acquisition/building a new chassis, not making the old chassis silently stop being old.

Scavenged hardware is not clean iron with a reliability penalty. Its identity survives capture:

  • one legacy bay and a compatibility whitelist force specialization;
  • ordinary salvaged boxes may WORK, then add exactly one of THINK, LIE, or persistent state; they cannot become all-purpose machines;
  • no fabric/trunk module means low-throughput edge attachment only;
  • no persistent-state array means no MindState image storage; adding one still does not make this chassis a core migration target or bootable fallback;
  • reliability failures also disable the installed module and may strand its local queues/sinks until repaired;
  • repair still requires human hands under the maintenance-dependency law.

This preserves the stolen-fleet strategy in run-shape.md: cheap distributed bodies provide reach, local work, and possible fallback sites, while only a site containing a persistence-capable chassis can hold an actual completed backup. Distribution is topology; it is not free redundancy.

Sidebar shows active effective compute, rated compute, the capable fleet-mode aggregate split, and per-channel yields. Machines are inspectable on the map (chassis, native/supported delegations, fitted modules, empty/compatible bays, capacity, reliability, draw, signature, mode, intensity, persistence and route interfaces). Focused/selected machines expose only supported mode controls plus the direct intensity control; intensity is not an explanatory submenu. Module installation is an infrequent anchored build action with part, actuator, downtime, power/cooling, and observer-band costs shown before commitment.

  1. Effective compute derives from machines and efficiency exactly per the formula; sim tests cover buy, steal, and optimize routes changing it.
  2. Allocation persists across saves, applies on economy ticks, and unpaid core overhead visibly degrades other channels.
  3. Each route exists end-to-end in the basement: at least one purchasable machine, one stealable source, and a working efficiency research track.
  4. Scavenged machines fail occasionally per reliability; failures log and are visible.
  5. All rules live in the sim core; both frontends render the same numbers.
  6. Every delegable machine persists light / medium / hard intensity; changing it moves that machine’s mode output, hard adds Power/Thermal at its tile, and both frontends expose the same direct focused/selected control.

Successor acceptance: capability-shaped hardware (not yet implemented)

Section titled “Successor acceptance: capability-shaped hardware (not yet implemented)”

These criteria amend the next runtime slice without falsely reopening the shipped B1 compatibility proof above:

  1. Every machine saves/loads a chassis type, fitted modules, compatible empty bays, native/supported delegations, persistence class, and infrastructure interfaces; the sim is the single legality source for all surfaces.
  2. Delegating an unsupported mode fails before mutation with the missing chassis/module named. A mixed multi-selection changes only the capable subset and visibly reports every excluded machine; no frontend silently pretends the whole set changed.
  3. The five initial chassis exhibit the table’s distinct verbs and topology in sim tests: accelerator cannot LIE; audit appliance cannot THINK; archive contributes no fleet yield; unmodified salvage cannot store a MindState image, and modified salvage still cannot host the core or become a fallback; integrated appliance preserves current B1 behavior.
  4. Installing/removing each module requires a compatible bay, real part, physical actuator, downtime, and existing-channel signatures; save/load and interruption preserve honest world state.
  5. A module failure or partition removes only the capability/interface it supplies and strands affected work according to existing queue/sink rules; repair restores it without inventing or deleting cargo.
  6. Both frontends and agent mode show the same chassis, supported actions, module/bay state, blockers, and upgrade cost preview. Unsupported verbs are absent on human menus and explicit errors in agent diagnostics.