Add BUDGET.md

calibration/BUDGET.md

@@ -1,118 +1,37 @@
-# Prime Intellect budget ledger — 100-turn reward calibration
-
-- Starting balance (2026-08-10): **$1,176.87**
-- Calibration spending cap: **$250.00**
-- Balance reserved for the later full-scale evaluation: **$926.87**
+# Prime Intellect budget ledger — v2.1 final calibration
+
+- Starting settled balance (2026-08-11): **$1,049.90**
+- Final hard calibration cap: **$650.00**
+- Final settled balance (2026-08-12): **$439.26**
+- Total wallet spend: **$610.64**
+- Amount under cap: **$39.36**
+- Remaining full-evaluation reserve: **$439.26**
 - Authoritative accounting source: `prime --plain wallet`
-- Balance after native smoke: **$1,176.15**
-- Smoke wallet spend: **$0.72** (trace-recorded: $0.7153)
-- Balance after stopping the stable Luna screen tail: **$1,167.48**
-- Total wallet spend through the broad screen: **$9.39**
-- Broad screen retained 91/96 completed rules (at least 10 per family); its five
-  remaining in-flight tail cases were cancelled once neither possible outcome
-  could change the adaptive-stage decision. Trace-recorded completed cost:
-  **$8.2253**.
-- Balance during the five-rule v5 global-history probe (2026-08-10 15:25
-  America/Mexico_City): **$1,160.14**.
-- Total wallet spend through the completed 42-rule adaptive supplement and the
-  in-flight global probe: **$16.73**. This leaves **$233.27** inside the
-  calibration cap and **$1,160.14** for calibration plus the later full-scale
-  evaluation.
-
-## Stage caps
-
-| Stage | Maximum wallet spend | Gate before continuing |
-|---|---:|---|
-| Native 100-turn model/context smoke | $10 | No protocol, context, or systematic tool-use failures |
-| One-deal Luna screen (96 rules maximum) | $50 | Every family has candidates near the reward target |
-| Terra + DeepSeek Pro confirmation (97-rule expanded maximum after the targeted gate) | $120 | Per-family reward bands are estimable and candidates remain |
-| Sol sentinels, repeat deals, provider overhead | $70 | Only boundary/selected rules; stop at total cap |
-
-The workflow checks the wallet after every paid stage. Trace-reported cost is
-diagnostic; the wallet delta is binding because provider errors and aborted
-requests may not appear in accepted traces.
-
-## Cost-control rules
-
-- Candidate generation, validation, equivalence checks, and observability run locally.
-- No broad Grok or Sol screen during calibration; those models belong in the
-  later confirmatory evaluation.
-- No paid stage expands automatically. It must satisfy its quality gate and
-  remain inside both its stage cap and the $250 total cap.
-- Context-limit/provider-failed traces are not scored as zero-reward episodes.
-- Repeated deals are allocated to selected or high-variance rules, not to the
-  full candidate pool.
-
-## Completed stages
-
-### Native Luna smoke
-
-- 8 valid episodes, one per family; the preceding turn-0 tool-server failures
-  made no model calls and are excluded.
-- Maximum valid plays: 97; maximum model calls: 103.
-- Maximum final request context: 68,391 tokens (6.5% of Luna's 1.05M window).
-- Valid-episode solve rate: 0.125; mean reward: 0.1138. This was an intentionally
-  hardest-sentinel smoke, not a dataset estimate.
-- No context, provider, card-supply, scoring, or stopping failures.
-
-### Broad Luna reward screen
-
-- 91 completed rules; 10–12 observations per family.
-- Family mean reward ranged from 0.000 (chunk/global) to 0.726 (static), proving
-  that aggregate calibration would reproduce the original family-confounding
-  problem.
-- Periodic-cycle reward was already inside the 0.20–0.42 family band. Six
-  targeted replacements were allocated to each of the other seven families,
-  for a 42-rule adaptive supplement rather than another full-bank pass.
-
-### Adaptive Luna supplement
-
-- 42/42 targeted rules completed, six each for the seven families outside the
-  broad-screen band.
-- The supplement produced useful intermediate/late candidates for first-order
-  transition and static-predicate rules. Conditional, compositional, and the
-  original global-history templates remained mostly at the floor, so these
-  traces are used only for candidate triage and are not reported as final model
-  estimates.
-- Some old-protocol supplement episodes ended after 94–99 valid plays because
-  invalid actions consumed the 115-call framework cap. The final v5 protocol
-  exposed that even a 150-call cap could stop one adversarial rollout at 98
-  valid plays. The final v6 protocol uses a 250-call cap; every reported
-  confirmation episode is rerun under v6.
-
-### Locked-v6 Luna/Terra confirmation
-
-- The common 80-rule stage retained 79 valid Luna and 77 valid Terra traces.
-  Trace-recorded cost was **$14.52** and **$30.21**, respectively.
-- Missing tail cells were not treated as zero-reward model failures. A
-  three-episode static-predicate repair stage was allocated so final selection
-  can require three valid model observations per rule.
-- Initial confirmation showed useful intermediate and late behavior in most
-  families, but compositional, higher-order, and global-history means remained
-  near the floor. This triggered a final **17-rule** supplement rather than a
-  broad repeat: four compositional, nine dynamic global-history, and four
-  higher-order candidates.
-
-### Final common-panel confirmation
 
-- Wallet balance immediately before the DeepSeek confirmation decision
-  (2026-08-10 17:45 America/Mexico_City): **$1,106.23**. Cumulative settled
-  wallet spend was **$70.64**, leaving **$179.36** inside the calibration cap.
-- The targeted Terra screen produced a turn-71 solution on a dynamic
-  global-history rule (reward 0.30), satisfying the evidence gate to run the
-  97-rule DeepSeek V4 Pro confirmation.
-- No Sol screen and no broad repeat-deal stage will be run during calibration;
-  that budget remains reserved for the fresh-deal full evaluation.
-- The DeepSeek V4 Pro stage produced 96/97 traces. It was stopped after every
-  family had at least eight valid three-model rules; the one remaining static
-  tail could no longer affect eligibility or selection.
-- A one-rule Luna redundancy repair was attempted while the last DeepSeek
-  traces were in flight. It failed before turn 1 with `ToolsetError`, made no
-  valid game progress, and is excluded from calibration aggregates.
-- Final settled wallet balance (2026-08-10 19:30 America/Mexico_City):
-  **$1,049.90**. Total calibration wallet spend was **$126.97**, or **50.8%**
-  of the $250 cap. This leaves **$123.03** of unused calibration headroom and
-  **$1,049.90** for the locked-dataset evaluation and later work.
-- Trace-recorded cost for the selected-rule confirmation records was
-  **$108.15**. The wallet delta remains authoritative and includes screens,
-  provider overhead, and stopped/invalid work.
+## Cost controls used
+
+- Paid calibration used Gemini 3.1 Pro and GPT-5.6 Sol only; the expensive
+  original panel remains reserved for the final evaluation.
+- One deal per candidate, with targeted repairs only after family/tier review.
+- Local compilation, behavioral uniqueness, acceptance balance, and
+  observability filters preceded all paid screening.
+- Wallet delta is the binding total. Per-trace usage is diagnostic because
+  provider failures and interrupted work are not always represented completely
+  in trace cost totals.
+- Inference stopped after the final Gemini confirmation wave produced repeated
+  upstream 400s and request timeouts. Those partial traces are preserved as
+  right-censored evidence; no budget was spent retrying them again.
+
+## Ledger
+
+| Checkpoint | Balance | Cumulative spend | Decision |
+|---|---:|---:|---|
+| Start | $1,049.90 | $0.00 | Begin local validation and minimal provider diagnostics |
+| Heartbeat diagnostics | $1,033.87 | $16.03 | Isolate Gemini signed-state failure |
+| Protocol v10 diagnostic | $943.33 | $106.57 | Reject 16k per-play completions as too expensive |
+| Protocol v11 lock | $789.87 | $260.03 | Required tool use removes thought-only early stops |
+| Screens and targeted repairs complete | approximately $446 | approximately $604 | Freeze paid broad screening; only final Gemini check remains |
+| Final settled checkpoint | **$439.26** | **$610.64** | Stop inference; $39.36 below cap |
+
+The last line is authoritative. Intermediate approximate lines are retained to
+explain stage decisions, not to supersede the settled wallet delta.