SUPERHERO SUBSTRATE CAPACITY LEDGER — 2026-09-09

Status: NON-AUTHORITY CURRENT-REALITY RESEARCH
Purpose: Quantify currently legitimate $0 / already-entitled execution capacity for making Project Brain superheroes real.
Economic constraint: additional out-of-pocket spend = $0. Do not buy credits, enable auto-reload, or bypass provider quotas.

1. Core distinction

SUPERHERO_COUNT != FRONTIER_MODEL_STREAM_COUNT

A real superhero may combine:

  • goal/controller state;
  • JIT knowledge/skills;
  • deterministic programs/tools;
  • remote CPU execution;
  • occasional frontier-model cognition;
  • independent/deterministic verification.

The current resource landscape does not support one million simultaneous frontier-model streams at $0.

2. Current model-cognition lanes

ChatGPT Work / Codex — already-entitled when available

Current OpenAI documentation:

  • Codex is included across ChatGPT plans.
  • Work and Codex share the plan's agentic allowance.
  • Plus has bounded five-hour and weekly windows.
  • Current estimated Plus local-message ranges per five-hour window include:
    • GPT-6 Astra: 5–45
    • GPT-5.6 Sol: 10–100
    • GPT-5.6 Terra: 25–200
    • GPT-5.6 Luna: 250–2,000
  • These are estimates, not fixed task counts; actual usage varies by task/model/settings.
  • Codex is explicitly designed for parallel agent workflows.
  • This current ChatGPT tool surface does not expose a callable Work/Codex delegation action, so no capacity is invented here.

Sources:

Kilo Gateway free models — empirically callable now

Current official documentation:

  • free-model requests: 200 requests/hour/IP
  • applies to anonymous and authenticated free-model traffic
  • :free and Auto Free are zero-priced on Kilo's side
  • upstream availability/rate limits can still fail

Sources:

Capacity implications for one allowed IP quota:

  • deterministic verifier, one model generation per attempt: at most 200 generation attempts/hour
  • separate model producer + model verifier: at most 100 producer/verifier pairs/hour
  • 200/hour = 0.0556 request starts/second on sustained average
  • at an average 10 s model-call duration, Little's-law average in-flight capacity from that quota is only about 0.56 calls
  • for 1,000,000 logical superheroes sharing only this lane evenly, the budget is 0.0002 model calls/superhero/hour, or roughly one model call per 5,000 superhero-hours

Conclusion: KILO_FREE = QUALIFIED_ZERO_COST_EXPERIMENTAL/BACKUP COGNITION, NOT MILLION-SUPERHERO PRIMARY SUBSTRATE

Multiplying IP identities to evade the quota is inadmissible.

3. Deterministic / tool execution lanes

Val Town Free

Current official limits:

  • $0
  • 100,000 runs/day
  • 1 min wall-clock timeout per normal run
  • 1,000 authenticated API requests/minute
  • 4 GiB worker memory
  • 6 h absolute worker lifetime
  • 10 s worker idle timeout
  • public vals on Free

Sources:

Use for:

  • deterministic preprocessing
  • verification
  • small serverless programs
  • event/scheduled re-entry
  • public/non-sensitive tool execution

Do not treat it as frontier-model cognition or a permanent one-process agent host.

GitHub Actions on public repositories

Current official GitHub documentation:

  • standard GitHub-hosted runners are free for public repositories
  • GitHub Free standard-runner concurrency limit: 20 concurrent jobs

Sources:

Use for:

  • builds/tests
  • verifiers
  • compilers
  • simulations
  • longer deterministic jobs
  • reproducible execution

It is not itself an LLM inference source.

GitHub Codespaces personal Free

Current official allowance:

  • 120 core-hours/month
  • 15 GB-month storage
  • if no valid payment method exists, use is blocked after the quota instead of silently continuing into billable usage
  • if a payment method exists, enforce a stop-usage budget before autonomous use

Sources:

Use for:

  • durable development/test workspace
  • repository-level tooling
  • bounded CPU work

Not a million-worker substrate.

4. Scaling identity

Let:

  • F = number of logical superheroes
  • d = fraction simultaneously requiring frontier-model inference
  • S = available simultaneous frontier-model streams

Necessary condition:

F * d <= S

Therefore a million logical superheroes is only plausible at $0 when d is extremely small, or when a much larger independently granted inference substrate appears.

The architecture must aggressively reduce d by:

  1. solving exact deterministic work deterministically;
  2. reusing verified shared knowledge/skills instead of re-reasoning;
  3. using models only on residual uncertain cognition;
  4. executing ordinary programs/tools outside the model;
  5. verifying cheaply and deterministically where possible;
  6. batching/shared preprocessing when it preserves independence of consequential work.

5. Current falsification

The latest weighted-interval spike sent 20 model calls to tasks whose exact optimum was already computed in the same script by dynamic programming. Result: 0/20 verified model outputs.

This establishes a routing failure, not a need for more inference.

Rule: IF_CHEAP_DETERMINISTIC_EXACT_SOLUTION_EXISTS => MODEL_INFERENCE_ELIGIBILITY = FALSE

6. Current truth about the million target

Not currently supported

  • one million simultaneous frontier-grade LLM agents
  • one million independent high-duty-cycle model streams
  • one million workers each receiving frequent frontier-model calls at $0

Architecturally possible to pursue

  • one million logical work identities/tasks/cells
  • large populations where most work is deterministic/tool-executable
  • sparse frontier cognition shared across a much larger execution population
  • independent consequential execution/verification where cheap mechanisms exist
  • growth as new legitimate donated, entitled, or free compute becomes available

7. Next highest-value empirical quantity

Measure frontier-cognition duty cycle on real project work:

MODEL_DUTY = model-required wall time / total superhero wall time

and:

MODEL_CALLS_PER_VERIFIED_TASK

The million-superhero target becomes quantitatively meaningful only after these are measured on representative work.

The dominant research goal is therefore not "find one million model endpoints." It is:

Minimize model-cognition duty per independently verified unit of useful work, then scale the cheap execution around the residual cognition bottleneck.