Add FINAL_REPORT.md

calibration/FINAL_REPORT.md

@@ -1,235 +1,224 @@
-# Eleusis 100-turn reward-calibration report
+# Eleusis v2.1 final calibration report
 
 ## Executive result
 
-We built and locked a **64-rule test dataset with exactly eight rules in each of
-eight semantic families**, then calibrated it under a 100-turn protocol using
-model-balanced normalized reward rather than pass rate. The selected suite has
-mean reward **0.240**, solve@100 **37.5%**,
-solve@30 **18.8%**, and a late-solve rate of
-**18.8%**. Among solved episodes,
-**50.0%** were solved only after the old 30-turn
-cutoff. Family mean-reward spread is **0.253**.
-
-The dataset is published as provisional because these preregistered gates were missed: `every_family_reward_0.20_to_0.42`, `family_reward_spread_at_most_0.15`, `overall_reward_0.25_to_0.40`. The misses are exposed rather than averaged away.
+We froze a **32-rule test split**: eight semantic families, four explicit
+complexity tiers per family, and one rule at every tier. Under the 100-turn v11
+protocol, GPT-5.6 Sol achieved **50.0%
+solve@100** and **0.253 normalized reward**. Only
+**12.5%** of rules were solved by the former
+30-turn horizon; **75.0% of successful
+solves occurred on turns 31–100**. This is direct evidence that the extra
+horizon carries outcome-relevant search rather than merely padding the task.
+
+Every family retained at least one unsolved rule, and GPT Sol family reward
+ranges from **0.150** to
+**0.443**. Thus, low aggregate reward does not
+come from one impossible family cancelling one easy family.
+
+Gemini confirmation is **incomplete** because Prime's Gemini gateway failed
+during the last economical confirmation wave. Of the 32 locked rules, Gemini
+has 11 exact endpoints, 11
+right-censored prefixes, and 10 rules without paid
+evidence. Its exact-only reward (0.496) is
+selection-biased toward episodes that completed and is **not** a dataset-wide
+Gemini score. Provider and harness timeouts are never relabelled as benchmark
+failures. This leaves one intentionally visible missed gate:
+`gemini_full_exact_32_rule_confirmation`.
 
 Published dataset: [https://huggingface.co/datasets/nph4rd/eleusis-calibrated-rules](https://huggingface.co/datasets/nph4rd/eleusis-calibrated-rules).
 
-These values are selection-conditioned calibration estimates. They describe the
-locked dataset construction; the planned full-scale evaluation should use new
-deals and the broader model panel for confirmatory model claims.
-
-![Reward by family](reporting/reward_by_family.png)
-
-*Figure 1. Family-balanced reward with rule-bootstrap uncertainty. The shaded
-band is the preregistered 0.20–0.42 family target.*
-
-## What we changed
-
-This is a dataset calibration exercise, not a new environment. The environment
-changes are limited to what the 100-turn dataset requires:
-
-- The turn limit is 100 and reward is `(101 - first_correct_turn) / 100`, or zero
-  when unsolved.
-- The framework allows up to 250 model calls so invalid actions, which consume
-  no game turn, cannot shorten the intended 100-valid-play horizon.
-- The legacy starter, hand, and primary two-deck order are preserved exactly.
-  Deterministically seeded reserve shoes are appended only when a long game needs
-  more cards.
-- The prompt now states the 100-turn horizon, explains the reserve shoe, asks the
-  model to maintain a compact evidence table, and requires one `play` call on
-  every response until the engine says game over.
-- Metrics now separate solve@30, solves on turns 31–100, early abandonment,
-  exhausted horizon, invalid actions, context usage, and provider/context errors.
-- Context or provider failures are excluded from reasoning aggregates. They are
-  protocol failures, not zero-reward induction failures.
-
-The dataset generator now exposes controlled ladders over static clause count,
-modular cardinality, transition sets, conditional branching, period/chunk
-length, history window, global summaries, and composition. Structured
-suit–rank modular laws were added after the first screen showed that ordinary
-static thresholds were still solved immediately.
-
-## How calibration was run
-
-1. We generated **720 fully held-out candidates** and validated compilation,
-   starter existence, acceptance balance, behavioral uniqueness, and split
-   leakage locally.
-2. Greedy and random version-space probes screened observability at 100 turns.
-3. A native family-stratified smoke verified the extended deck, scoring,
-   trace persistence, and context growth.
-4. A one-deal Luna screen completed 91 candidates (at least 10 per family). It
-   revealed family reward from 0.000 to 0.726, so those results were not averaged
-   into a misleading aggregate.
-5. We first spent on a **42-rule adaptive supplement**: six simpler rules for each
-   floor-zero/too-hard family and six harder rules for each too-easy family.
-   Periodic rules received no supplement because their screen mean was already
-   in band. The exact adaptive decisions are in
-   `stages/luna_supplement_v2_summary.json`.
-6. Two-model confirmation then exposed a residual floor in compositional,
-   higher-order, and especially global-history rules. We generated a final
-   **17-rule targeted supplement**: four simpler compositional,
-   four higher-order, and nine dynamic first-anchor/last-card global rules.
-   Anchor-only predicates were deliberately excluded because their behavior is
-   static within one episode and does not identify global reasoning.
-7. From the combined screen we chose candidates per family, prioritizing
-   intermediate and post-30 solutions, low early-solve rate, target reward, low
-   dispersion, and template diversity. Up to **97**
-   candidates were run under the locked v6 prompt on Luna, Terra, and DeepSeek
-   V4 Pro.
-8. We averaged deals within model, models within rule, and eight rules within
-   every family. We selected eight confirmed rules per family and locked the
-   dataset before the later full-panel evaluation.
-
-Calibration spent **$126.97** from Prime Intellect, leaving
-**$1049.90**. This is below the $250 calibration cap and leaves
-the rest for the confirmatory evaluation.
-
-![Calibration funnel](reporting/calibration_funnel_100turn.png)
-
-*Figure 2. Local screening and an adaptive one-model stage constrain paid model
-work. Stage counts are not all nested: the supplement adds targeted candidates.*
+The dataset is structurally finished and fully confirmed on GPT Sol. It is
+appropriate to freeze for the planned fresh-deal, full-panel evaluation, but
+that evaluation—not this censored construction sample—must provide the final
+two-model and cross-model estimate.
+
+![Family reward and solve rate](reporting/family_reward_and_solve.png)
+
+*Figure 1. Complete 32-rule GPT Sol results. Reward stays below 0.5 and at
+least one rule remains unsolved in every family.*
+
+## What we did
+
+This iteration primarily changed the **dataset**, not the game or score.
+Reward remains `(101 - first_correct_turn) / 100` for a correct rule and zero
+at the 100-turn horizon. We:
+
+1. Built and locally validated 104 candidates spanning static, first-order,
+   conditional, periodic, chunk/run, higher-order, global-history, and
+   compositional families.
+2. Preserved four family-specific structural tiers instead of pretending there
+   is one universal rule-complexity scalar.
+3. Narrowed to a 72-rule economical screen, then added targeted static and
+   first-order repairs. Global history and compositional hybrid retained three
+   alternatives per tier because the preceding iteration under-resolved them.
+4. Evaluated GPT Sol across the screen and repairs at one fixed 100-turn deal
+   per rule. We used completed Gemini episodes when exact and retained failed
+   prefixes only as censoring evidence.
+5. Locked one item at every family × tier cell, optimizing for both incomplete
+   solve coverage and late solutions rather than pass rate alone.
+6. Saved the dataset, raw traces, exact/censored tables, Catppuccin plots,
+   report, presentation chapter, and browser trace viewer.
+
+![Calibration funnel](reporting/calibration_funnel.png)
+
+*Figure 2. Cost-aware narrowing from the broad local bank to the balanced
+32-rule lock. The final suite avoids pseudo-replication by retaining one rule
+per family tier.*
+
+The Prime wallet moved from **$1049.90** to
+**$439.26**: **$610.64 spent**,
+**$39.36 under the $650 cap**. We stopped paid
+inference after the Gemini gateway failure wave rather than burn the remaining
+cap retrying transport errors before the much larger final evaluation.
+
+## Protocol accommodations for 100 turns
+
+- The initial deal is unchanged; deterministic reserve shoes extend card
+  supply only when the 100-turn game needs them.
+- Up to 250 model calls allow invalid-action recovery while preserving 100
+  *valid* game turns.
+- A neutral user heartbeat follows each tool result, preserving Gemini's
+  provider state; required tool choice enforces the already-stated one-play
+  contract while the game is live.
+- One continuation is allowed only after a provider reports a completion-length
+  cutoff. The full transcript is retained, with no hidden compaction.
+- A provider, context, or harness failure is a reliability failure—not an
+  unsolved rule. Exact metrics require a clean terminal `game_over`.
+
+Reward and verifier semantics were not changed.
 
 ## What we found
 
-### Reward balance by family
+### Family difficulty is broadly low, while pass rate and reward remain distinct
+
+| Family | Sol reward | solved ≤30 | solved 31–100 | solved ≤100 | Gemini exact / censored / missing | Gemini reward, exact only |
+|---|---:|---:|---:|---:|---:|---:|
+| Chunk / run | 0.193 | 0.0% | 50.0% | 50.0% | 3 / 0 / 1 | 0.697 |
+| Compositional hybrid | 0.378 | 25.0% | 50.0% | 75.0% | 3 / 1 / 0 | 0.430 |
+| Conditional transition | 0.150 | 0.0% | 25.0% | 25.0% | 0 / 0 / 4 | — |
+| First-order transition | 0.237 | 0.0% | 50.0% | 50.0% | 0 / 4 / 0 | — |
+| Global history | 0.163 | 0.0% | 25.0% | 25.0% | 2 / 2 / 0 | 0.000 |
+| Higher-order history | 0.182 | 25.0% | 0.0% | 25.0% | 0 / 0 / 4 | — |
+| Periodic cycle | 0.443 | 50.0% | 25.0% | 75.0% | 3 / 0 / 1 | 0.693 |
+| Static predicate | 0.278 | 0.0% | 75.0% | 75.0% | 0 / 4 / 0 | — |
+
+Periodic and compositional rules have the highest GPT Sol solve rates
+(75.0% and
+75.0%), but
+their rewards remain only 0.443
+and 0.378
+because solutions are slower and the top tiers still defeat the model. This is
+exactly why both metrics are needed: pass rate asks whether a rule is solved;
+reward asks how much experimental search it took.
+
+![Exact model-family reward](reporting/model_family_reward_heatmap.png)
 
-| Family | Rules | Reward | Solve ≤30 | Solve 31–100 | Solve ≤100 | Late share of solves | Intermediate rules |
-|---|---:|---:|---:|---:|---:|---:|---:|
-| Chunk / run | 8 | 0.306 | 25.0% | 16.7% | 41.7% | 40.0% | 8 |
-| Compositional hybrid | 8 | 0.123 | 12.5% | 4.2% | 16.7% | 25.0% | 0 |
-| Conditional transition | 8 | 0.247 | 16.7% | 29.2% | 45.8% | 63.6% | 7 |
-| First-order transition | 8 | 0.332 | 29.2% | 20.8% | 50.0% | 41.7% | 3 |
-| Global history | 8 | 0.081 | 4.2% | 12.5% | 16.7% | 75.0% | 3 |
-| Higher-order history | 8 | 0.215 | 20.8% | 8.3% | 29.2% | 28.6% | 3 |
-| Periodic cycle | 8 | 0.334 | 29.2% | 16.7% | 45.8% | 36.4% | 2 |
-| Static predicate | 8 | 0.281 | 12.5% | 41.7% | 54.2% | 76.9% | 4 |
+*Figure 3. Exact completed endpoints only. Every cell states n/4; dashes are
+missing evidence, not zero reward.*
 
-The family view is the primary calibration result. Figure 1 shows whether each
-family lands in the intended reward band; Figure 3 exposes the individual rule
-points so an apparently good family mean cannot hide an immediate-solve plus
-impossible-rule mixture.
+![Gemini evidence coverage](reporting/gemini_evidence_coverage.png)
 
-![Rule reward dispersion](reporting/rule_reward_dispersion.png)
+*Figure 4. Exact, right-censored, and not-run Gemini items by family. Global
+history and compositional hybrid—the two requested refinements—have evidence
+for every locked rule.*
 
-*Figure 3. Every point is a selected rule; horizontal ticks are family means.*
+Global history is now distinctly difficult: GPT Sol solves one of four rules
+(reward 0.163); Gemini is exactly unsolved through 100 on both upper tiers and
+is right-censored only at turns 96 and 99 on the lower tiers. Compositional
+hybrid preserves a useful ladder: lower-tier compositions are solvable, the
+short-window tier defeats both models through turn 100, and the long-periodic
+tier takes GPT Sol 77 turns while Gemini remains right-censored through 94.
 
-### The extra horizon carries measurable signal
+### The old 30-turn horizon would erase most positive results
 
 ![Horizon outcomes](reporting/horizon_outcomes_by_family.png)
 
-*Figure 4. Solves are separated into the legacy horizon and genuinely late
-turns. Gray mass is censored unsolved at turn 100.*
-
-![Solve-time CDF](reporting/solve_time_cdf.png)
-
-*Figure 5. Cumulative solution timing by confirmation model; the dashed line is
-the old turn-30 boundary.*
-
-The benchmark is long-horizon only to the extent that Figures 4 and 5 show a
-substantial turn-31–100 contribution. A low reward produced only by early solves
-and zeroes would fail the family late-solve gates and remain visible in Figure 3.
-Early abandonment is reported separately because a model that ignores the
-required action protocol has failed persistence, but did not consume all 100
-experimental opportunities.
-
-### Model profiles are not interchangeable
-
-| Model | Episodes | Reward | Solve ≤30 | Solve 31–100 | Solve ≤100 | Early abandon | Recorded cost |
-|---|---:|---:|---:|---:|---:|---:|---:|
-| `deepseek/deepseek-v4-pro` | 64 | 0.216 | 17.2% | 14.1% | 31.2% | 39.1% | $31.19 |
-| `openai/gpt-5.6-luna-pro` | 64 | 0.185 | 14.1% | 15.6% | 29.7% | 39.1% | $11.55 |
-| `openai/gpt-5.6-terra` | 64 | 0.320 | 25.0% | 26.6% | 51.6% | 12.5% | $22.17 |
-
-![Model/family reward](reporting/model_family_reward_heatmap.png)
-
-*Figure 6. The selection objective is family balance after equal model weighting;
-the heatmap retains model-specific strengths and weaknesses.*
-
-### Complexity scales difficulty, but not with one universal ruler
-
-![Complexity versus reward](reporting/complexity_vs_reward.png)
-
-*Figure 7. Designed complexity generally lowers reward, but the vertical spread
-within a level shows why calibration must be conditional on rule family.*
-
-Static clause/cardinality, transition alternatives, branch count, period,
-window, and composition are useful scaling axes. They are not exchangeable:
-the same designed level can demand very different search strategies. The
-dataset therefore stores both semantic family and structural features instead
-of collapsing them into one synthetic complexity score.
-
-## Context windows and compaction
-
-The Eleusis/Null harness sends the full transcript on every request. It does not
-automatically summarize, truncate, or compact. In the measured current panel,
-100 turns do not approach the documented roughly one-million-token windows of
-[GPT-5.6](https://developers.openai.com/api/docs/models) and
-[DeepSeek V4](https://api-docs.deepseek.com/quick_start/pricing):
-
-| Model | Long traces | Observed final p90 | Turn-100 projected p90 | p90/window | Context errors |
-|---|---:|---:|---:|---:|---:|
-| `deepseek/deepseek-v4-pro` | 86 | 136,623 | 168,341 | 16.8% | 0 |
-| `openai/gpt-5.6-luna-pro` | 87 | 104,516 | 107,496 | 10.2% | 0 |
-| `openai/gpt-5.6-terra` | 85 | 31,989 | 42,237 | 4.0% | 0 |
-
-![Context usage](reporting/context_window_usage.png)
-
-*Figure 8. Median, p90, and maximum request context as a fraction of each
-model's documented window; the warning threshold is 50%.*
-
-No compaction is used because it would introduce a new, model-dependent memory
-intervention into an inductive reasoning benchmark. If a smaller-context model
-hits its limit, that rollout is marked as a protocol failure and excluded from
-reward, not silently scored zero. If future horizons require compaction, the
-correct design is a separately versioned deterministic evidence-checkpoint
-protocol rerun on every compared model.
-
-## Is this now a good benchmark?
-
-**Inductive/scientific reasoning:** the environment is a strong instrument, but
-this dataset release is not yet fully calibrated. The interaction requires
-models to propose executable hypotheses, choose discriminating interventions,
-update from accepted and rejected evidence, and converge under behavioral
-equivalence. However, Compositional hybrid and Global history remain below the family reward
-floor. The equal-family analysis exposes that failure instead of letting easier
-families hide it.
-
-**Long-horizon reasoning:** yes at the protocol level, with encouraging empirical
-evidence. Half of solved confirmation episodes occurred after turn 30, and every
-family had at least one late solve. The decisive evidence is that post-30 mass in
-Figures 4 and 5, not the nominal value “100.” A fresh-deal evaluation is still
-needed before treating the exact rate as a benchmark estimate.
-
-## Limitations and next evaluation
-
-- Rule selection and the calibration estimates use the same three-model panel;
-  do not present these numbers as an unbiased leaderboard.
-- One deal per model/rule is a cost-efficient calibration design, not a precise
-  estimate of sampling variance. The full evaluation should use fresh deals.
-- The rule language is reviewed and finite. Performance can improve through
-  benchmark-specific priors as the dataset becomes familiar.
-- Context safety is established for the current million-token panel, not for
-  every deployable model.
-- Any missed gate listed above remains a known limitation of this version.
-
-Do **not** spend the full-scale evaluation budget on this provisional split yet.
-First run a narrow v2.1 interpolation stage only for the two under-band families:
-simple anchor-gated transitions for global history, and compositions that sit
-between the immediately solved static-plus-transition rule and the floor-level
-periodic/window hybrids. Screen those candidates on Luna, confirm only survivors
-on Terra and DeepSeek, replace rules without touching the six in-band families,
-and rerun the registered family gates. The full-panel evaluation should begin
-only after that revised split is locked; its fresh-deal results must not feed
-back into item selection.
-
-## Artifacts
-
-- `PROTOCOL.md` — preregistered reward, horizon, family, and context gates.
-- `BUDGET.md` — wallet ledger and stage decisions.
-- `reporting/calibration_summary.json` — machine-readable final estimates.
-- `reporting/confirmation_records.csv` and `reporting/rule_calibration.csv`.
-- `reporting/selected_rule_ids.json` and `final/` — locked dataset.
-- `CONTEXT_PREFLIGHT.md` — context model and failure policy.
+*Figure 5. GPT Sol outcomes split into solved by 30, solved on turns 31–100,
+and unsolved at 100.*
+
+![Solve-time curve](reporting/solve_time_cdf.png)
+
+*Figure 6. The curve grows from 12.5% at turn 30 to 50% at turn 100. Twelve of
+sixteen successful episodes occur only after the dashed line.*
+
+This makes the suite empirically long-horizon for GPT Sol. The scientific
+reasoning loop is also substantive: models must propose executable hypotheses,
+choose interventions from a constrained hand, use positive and negative
+evidence, revise, and finally commit to an extensionally exact rule.
+
+### Complexity scaling is explicit and mostly directional, not universal
+
+| Family | Structural scaling axis | GPT Sol reward, tier 1 → 4 |
+|---|---|---|
+| Chunk / run | chunk length and state: 2 → 5 → 6 → 11 | 0.18 → 0.00 → 0.59 → 0.00 |
+| Compositional hybrid | component count, gated state, and temporal span | 0.73 → 0.54 → 0.00 → 0.24 |
+| Conditional transition | previous-card gate specificity and branch interaction | 0.60 → 0.00 → 0.00 → 0.00 |
+| First-order transition | modular delta → conjunction → suit/rank-indexed transition table | 0.61 → 0.34 → 0.00 → 0.00 |
+| Global history | anchor state → running balance → cumulative-rank state | 0.00 → 0.65 → 0.00 → 0.00 |
+| Higher-order history | lookback/window length and aggregate statistic | 0.73 → 0.00 → 0.00 → 0.00 |
+| Periodic cycle | period length and phase alphabet: 2 → 4 → 7 → 10 | 0.74 → 0.80 → 0.23 → 0.00 |
+| Static predicate | modular width → gated dual modulus → suit-specific table | 0.52 → 0.49 → 0.10 → 0.00 |
+
+![Complexity tiers](reporting/complexity_tier_reward_by_family.png)
+
+*Figure 7. Every point is a retained rule. Six families show a strong downward
+direction across tiers; chunk/run and global history retain non-monotonic
+single-deal noise rather than hiding it with a fitted curve.*
+
+One observation per rule cannot prove a smooth item-response function. The
+tiers are defensible structural interventions—period length, chunk length,
+history depth, gating, or table cardinality—while empirical reward is noisy and
+model-specific. The full evaluation should preserve those tier labels and add
+fresh-deal variance estimates.
+
+### Context did not approach the configured ceiling
+
+![Context use](reporting/context_tokens.png)
+
+*Figure 8. Reported input sizes from exact endpoints. GPT Sol peaked at
+70,053 tokens; Gemini exact episodes peaked at
+11,535. The harness ceiling is 700,000.*
+
+The harness performs no automatic compaction. At this horizon, observed traces
+are far below the configured ceiling. If a future model has a smaller context
+window, the preflight must reject the run or use a separately versioned,
+deterministic compaction policy applied to every compared model. Silent
+model-specific summarization would change the task.
+
+## What to expect in the final full evaluation
+
+- GPT Sol should land near 50% solve@100 and roughly 0.25 reward on this fixed
+  deal, with substantial family and seed variance.
+- Models stronger at literal periodicity may score better on the two lower
+  periodic tiers while still losing reward through the long-period items.
+- Global-history upper tiers and higher-order windows are likely to be floor
+  items for several models; monitoring family solve and reward separately will
+  expose that rather than letting the aggregate hide it.
+- Gemini's dataset-wide result remains an open confirmatory question because
+  construction evidence is incomplete. The full evaluation should be the next
+  paid measurement, on fresh deals, without changing this lock afterward.
+
+## Limitations
+
+- One deal per rule/model was the deliberate low-cost calibration design; it
+  does not estimate seed or sampling variance.
+- Selection and GPT Sol point estimates use overlapping evidence and are
+  construction estimates, not a leaderboard result.
+- Gemini transport failures make its final-suite confirmation incomplete. The
+  exact/censored/missing split is preserved in every artifact.
+- Publishing executable rule code risks future contamination.
+- Structural tiers are designed scaling axes, not a fitted universal notion of
+  complexity.
+
+## Artifact index
+
+- `reporting/calibration_summary.json` — machine-readable metrics and gates.
+- `reporting/selected_rules.csv` — locked rules and evidence columns.
+- `reporting/calibration_records.csv` — exact episodes only.
+- `reporting/right_censored_records.csv` — failure prefixes, kept separate.
+- `reporting/selected_rule_ids.json` and `final/` — immutable 32-rule lock.
+- `PROTOCOL.md`, `BUDGET.md`, and `RUN_MANIFEST.md` — protocol and audit trail.
 - `outputs/trace_viewer/index.html` — manual trace inspection.
-- `outputs/reporting/presentation.html` — extended Catppuccin presentation.
+- `outputs/reporting/presentation.html` — full presentation with this chapter
+  appended after the HF-style evaluation and prior calibration.