Track C — Engine Plumbing

Purpose

Tracks A0/A/B produce evidence; Track C builds the machinery that makes producing such evidence routine. Every component here is specified conceptually in Engine Architecture; this page is the v1 scope — what gets built this phase, what is deliberately deferred, and what "working" means for each piece.

C1 — Quest compiler v1

From prototype to pipeline. The novelty engine already does the hard part: LLM proposes a goal plus code that verifies it, scored against an archive. v1 re-points it and hardens it:

1. Input adapter: consumes the gap list (thin atlas cells + weak eval slices) instead of open-ended novelty.
2. Goal + verifier generation: as today, with the verifier emitted in two forms — a client-side check where structured state allows, and a server-side post-hoc check against the recording (VLM-based where needed, per the Klindt POC wall-detection pattern).
3. DSL compilation: output lands in the quests requirements JSON exactly as specified in Quests — Technical Design — the compiler is upstream of, not a replacement for, the quests spine.
4. Triage filter: every candidate scored on the A–E axes; COMMODITY/fake-transfer candidates rejected regardless of gap size — plus The Journey's second test ("does it make the community stronger?"): the compiler scores data value; a human applies the community test using the spatial-coverage proposal's UX principles (guidance without coercion, intrinsic reward, honest confidence) — directed collection must not become "turning play into work." Human review of every quest before publication in v1 — automation of the review is not a v1 goal.
5. Anti-bias emission: ≥3 phrasings per quest generated by default; per-contributor caps per quest family; consent-scope check (quest must not elicit data types outside the contributor's consent version — Phase 0 0.3).
6. Bounty field: priced rareness × leverage (Reward Economy integration); v1 pricing is a lookup table reviewed by a human, not a live market.

Done when a gap identified in the atlas becomes a published, priced, verifier-carrying quest with no hand-written JSON.

C2 — Trigger library + coverage signatures (overlay)

Merge, don't parallel-build (vault revision 2026-07-08): this component re-specifies §7 of the existing realtime-spatial-coverage proposal — which already defines coverage capture-specs with NBV scoring, frontier detection, and the fog-of-war HUD that is the natural quest-delivery surface. C2 v1 = that proposal's items 1–4, with the embedding-space atlas as the cross-title layer above its per-title world-space voxel map, and angle-diversity ("% seen from 3+ viewpoints", axis E) adopted into the coverage-unit definition. First task: run the pending Tier-0 GSI spike — the cheapest existing path to live signatures (rights note: CS2 is a Valve title — Tier-0 is internal calibration only).

The Tesla/Mobileye pattern, minimum version.

• Coverage signatures: the overlay computes compact per-session descriptors (embedding summaries, event statistics) and ships them always — kilobytes, not video. These feed the atlas without waiting for uploads and give the density map its denominator (what the fleet plays, not just what it uploads).
• Trigger library v1: a versioned set of client-side predicates — quest-verifier hooks, atlas-cell membership checks, and simple rarity triggers ("this session entered a thin cell") that flag full-quality upload. Model-disagreement triggers (shadow-mode style) are deferred — they need a deployed model to disagree with.
• Ops loop: trigger → collect → retrain → remeasure, with trigger retirement when a cell saturates (Cruise's anti-bloat rule).

Done when signatures flow for ≥90% of sessions, triggers are remotely versionable, and one full trigger-retirement cycle has been exercised.

C3 — Curation pipeline v1

Dedup (semantic, embedding-based), spec-conformance filters (sync metric, resolution, completeness), quality filters (crash/menu-idle detection), taxonomy balancing across titles and atlas cells — the Cosmos-shaped stack at small scale, expecting single-digit-% survival of raw hours into premium tiers. Every curation decision written to the provenance ledger (0.5).

Done when raw fleet hours flow to tiered, deduplicated, provenance-clean shards without manual triage.

C4 — Gold quests + fraud detection v1

Assume adversaries from day one — pay-per-coverage invites botted play, replayed sessions, and emulator spoofing, and there is no public QC literature for gameplay fraud (we're writing it — Engine Architecture §4).

• Gold quests: tasks with known-verifiable outcomes seeded into the stream — some visible (deterrence), some indistinguishable (detection).
• Input-stream forensics: human-cadence statistics (inter-key timing distributions, mouse micro-jitter, fatigue drift) vs scripted play; replay detection via session fingerprinting and near-duplicate embedding search (shares C3 machinery).
• Reputation: graduated trust per contributor — new accounts earn into higher-bounty quests; anomalies quarantine payouts pending review, with an appeal path (contributor trust is a product surface, not just a defense).
• Economics guardrail: bounties pay on verified coverage units, never hours (Reward Economy alignment) — the incentive design is itself the first fraud control.

Done when gold-quest coverage spans all active quest families and a red-team exercise (someone on the team actively cheating) is caught by the pipeline.

C5 — Influence scorer v1 + annealing validation

The two-tier valuation stack, minimum version (Engine Architecture §1):

• Fast tier: a small model scoring every incoming clip for expected training value — v1 is trained on cheap proxies (atlas rarity, quest-verification outcome, curation tier) and upgraded to MATES-style locally-probed influence once Track A produces a model to probe against.
• Slow tier (ground truth): one annealing ablation per cycle — mix a candidate slice at high weight into a partially-trained checkpoint, measure eval-slice deltas (the Llama 3 pattern). Track A is the first annealing cycle; C5 turns its harness into a repeatable job.
• Calibration loop: scorer ranking vs annealing outcome, reported each cycle; the scorer earns authority only as its calibration holds. This loop is also what makes the atlas learnability-weighted (Klindt's broken-TV caveat): cells whose data demonstrably moved the model get priority; cells that stay flat despite data get retired as noise, not farmed as novelty.
• Deferred: gradient attribution (EK-FAC/TRAK) — forensic use only, later; per-clip Shapley — research luxury, not v1.

Done when every clip carries a value score at ingest, and one full calibration cycle (scorer prediction → annealing verdict → scorer update) has run.

What v1 deliberately defers

Model-disagreement triggers (needs deployed models); automated quest review; live bounty markets; gradient attribution; naturalness-aware automatic mix ratios (measure first — Founding Experiment step 7). Deferral is a scope decision, not a disagreement with the architecture.