Skip to content

ledger_pricing.md

Latest commit

 

History

History
81 lines (68 loc) · 4.1 KB

ledger_pricing.md

File metadata and controls

81 lines (68 loc) · 4.1 KB
Raw

Ledger pricing: post-phase8 changes

Ratified 2026-07-03 (phase8 gate outcome, user-approved): mint pricing moves from score-movement-under-geometric-equilibration to explicit debate standing. Equilibration becomes an opt-in experimental kernel. Emission rate/schedule untouched (still config-gated, still unset).

v3 note (2026-07-08, docs/tool_substrate.md Decision): on the live tool rail, standing has since been retired from mint entirely (mint = usage_term) and Change 2's violator-pays gate is dormant — work units no longer mint at all, and tool mint is not debate-gated. Change 1 (ledger replay determinism) and Change 3 (arm-B render) stand.

v4.1 note (2026-07-09, gradient trust): mint = usage_term still, and drift weight dropped its ε floor for continuous rep-weighted credibility. None of this touches the pricing MODE machinery here (ledger vs equilibrated) — Changes 1/3 stand unchanged. See docs/tool_substrate.md, Decision 2026-07-09.

Fees-only note (2026-07-10, current): the v4.1 rep/ATN mint split and its supply-pegged β cap were both DELETED. The close now computes MONEY ONLY — the ATN epoch pool is burned service fees only (no base emission), distributed pro-rata over usage shares; REP is claimed DAO-side (RepToken) on ratified ATN earnings, never minted by the close. The authoritative payload is schema 3 with a 2-field (agent, amount) merkle leaf. This still does not touch the pricing MODE machinery here — federated_epoch_close keeps pricing="ledger" | "equilibrated", default "ledger", and Changes 1/3 stand. See docs/tool_substrate.md, Decision 2026-07-10.

Change 1 — pricing mode in the epoch close (owner: reconcile agent)

federated_epoch_close(..., pricing="ledger" | "equilibrated"), default "ledger"; "equilibrated" preserves today's behavior bit-for-bit for the experimental kernel.

Ledger mode:

  • Replay applies causal events (sprouts, posts) WITHOUT equilibrate rounds — no geometric propagation, no derived-sprout capture. Per-node score = the existing net_score tree recursion (posts + signed children). Snapshots record these.
  • Mint formula shape unchanged: rise × survival × I(close>0); the novelty factor reads node.n where present else 1.0.
  • Attribution uses causal events only (author-tagged sprouts).
  • Determinism: net_score memoization over co-parented cycles is eval-order sensitive (see test_act_memoization) — ledger mode must fix an evaluation order (sorted node ids) and prove bit-identical double-close in tests.
  • Plumb the mode through WorldService epoch close the same way epoch_emission_rate is plumbed. O(N²) equilibrate leaves the launch hot path.

Change 2 — violator-pays mint gate (owner: reconcile agent)

Replace apply_mint_gate's uniform global ratio (mint_gate.py:207-218): retain the per-(node, agent) attribution breakdown in reconcile, scale it per-node by (1 − gate_strength × violation), THEN aggregate per agent. With the emission pool, suppressed mint now genuinely redistributes from flagged nodes' authors to everyone else — the docstring's promise ("winning a CON debate increases your share") becomes true. Tests must show: violator down, honest agents up post-pool, and no change when no violations.

Change 3 — inference render per phase8 arm B (owner: infer agent)

Phase8: verdicts in the prompt HURT (C−B = −0.28); standing must rank, not fill context.

  • _infer_artifacts standing must work on non-equilibrated worlds (net_score recursion — it already does; add a test constructed without calling equilibrate()).
  • New render_context(result, char_budget=6000) -> str in infer.py: arm-B format — standing-ranked payloads only (problem+resolution text, proportional truncation), NO claims/verdict text. Claims stay in the structured result for programmatic consumers.
  • WorldService.infer_artifacts gains render=False kwarg returning {"result": ..., "context": render_context(...)} when True.

Out of scope

Defender loop (farmability), decay/novelty redesign, emission params, Phase 9. Docs sweep (CLAUDE.md training-loop section, README stalenesss) is a third, code-free task.