Eight rollouts are not eight votes

Eight trajectories enter. Two succeed. Six fail. The obvious verdict is 25% accuracy. The useful question is harder: did those two successes reveal one strategy twice—or two ways of solving the problem that the model should remember?

Reliquary already operates a live market for the first half of that question. Miners generate groups of eight rollouts; the validator admits a fixed-window population, ranks candidates by difficulty, verifies the highest-ranked groups through GRAIL proofs and archives the winners. The market searches for prompts at the current policy's learning frontier: not mastered, not impossible, but learnable now.

This article is about the second half. Difficulty can locate the frontier. It cannot tell us which equally difficult group will improve the next checkpoint, whether two correct trajectories are behaviorally distinct, or which failed trajectory should become a future evaluation case. Calling that missing quantity “novelty” is tempting. It is also dangerous. Novelty is not utility.

1. Start with a receipt, not a vision

At approximately 01:00 UTC on 23 July 2026, Reliquary's public validator state showed checkpoint 50 and open window 24,804. The previous sealed archive, window 24,803, is a compact account of what the production system actually did. Math received 23 pending candidates; Code received 53. The validator made 11 Math and 16 Code proof attempts in total. Two per environment were no-emission, untrained forensic samples, leaving 9 Math and 14 Code attempts on the ranked winner path; eight proven groups were selected per environment.

Across 128 rollout trajectories, the archive reports complete receipt conservation, validator-authoritative reward alignment, no paid-but-unselected group, no late drop and no training quarantine. But the optimizer did not run: the canary had reached its configured checkpoint-50 ceiling. The exact status was training_attempted=false, trained=false, and blocked_reason=training_checkpoint_ceiling.

That distinction matters. The rollout and proof market remains live; it is selecting, proving and archiving training signal. At this snapshot, checkpoint updates were paused by configuration. A serious account of a learning system must say both.

2. Why two out of eight is the discovery peak

Let k be the number of correct trajectories in a binary-reward group of eight and p = k/8. Its population standard deviation is σ = √p(1−p). Reliquary's current difficulty auction ranks an admitted group by σ(1−p): variance identifies contrast, while the second factor favors the rarer-success side of the frontier.

The complete nine-outcome table is more useful than a slogan. At 0/8 and 8/8, every group-relative advantage collapses: there is no contrast. At 4/8, signal is balanced. At 2/8, the two correct paths receive a large positive advantage while the six failures receive a smaller negative one. The score reaches 0.325, the maximum over the admissible interior. At 6/8 the geometry mirrors: two rare failures receive the strongest negative advantages, useful for refinement and evaluation even though they should not be positively imitated.

A 2 / 8 group is not merely a low score on accuracy. It is a high-leverage contrast: two rare successes against six nearby failures.

3. The selected archive sample peaks at two

We derived a frozen aggregate from ten immutable public window archives, 24,793 through 24,802. It contains 138 selected groups and 1,104 trajectories: 395 rewards above 0.5 and 709 at or below that threshold. The sample spans 76 OpenMath groups and 62 OpenCode groups, submitted by 45 distinct miner hotkeys mapped to 24 operators. No address, prompt, completion or identifier is retained in the reproducible aggregate.

Eighty-four groups—60.9% of the sample—had exactly 2/8 rewards above 0.5. Twenty-one had 3/8; twelve had 4/8; ten had 5/8; eleven had 6/8. Ten Code groups contain fractional rewards, so this thresholded mixed-environment distribution is descriptive; it is not the binary OpenMath table projected onto every group. It is production telemetry, not a randomized experiment and not evidence that the rank caused downstream learning.

That concentration creates the next research problem. In binary OpenMath, every 2/8 group shares the same difficulty score. Fractional Code rewards can create additional values, but exact-score plateaus remain. They are ordered by drand-anchored chronological buckets and deterministic sealed tie-breaks. Neither estimates training value.

Two correct completions may be paraphrases of the same derivation. Or they may use distinct evidence, tools, subgoals or verification paths. Difficulty sees the same p, σ and score either way. A future system needs another object: a versioned account of behavior and a causal test for whether preserving that behavior helps the model.

4. Novelty search changes the object of optimization

Classic novelty search was introduced as an answer to deceptive objectives. Instead of asking only how close a candidate is to a target, it maps the candidate to a behavior descriptor b(x) and measures its average distance to nearby behaviors in a complete comparison set C: the archive plus the current comparison population. The archive makes history part of the objective: a behavior is interesting because the search process has not occupied that region before.

Pure novelty deliberately ignores task quality. That can cross a deceptive valley, but it can also spend the budget on difference for its own sake. Novelty Search with Local Competition added quality relative to behavioral neighbors. MAP-Elites made the output concrete: a map of behavior niches, each retaining its strongest known elite. The resulting field is now described as quality-diversity—a repertoire rather than one champion.

For reasoning trajectories, a behavior descriptor cannot simply be “the hidden chain of thought.” Private reasoning is neither a stable interface nor a safe public artifact. A defensible descriptor would combine observables: tool-call sequence, evidence graph, environment transitions, answer/outcome class, verification operations, and a versioned semantic strategy embedding. Verbosity, latency and formatting should be explicit cost dimensions or excluded, otherwise they become cheap ways to look novel.

Correctness tells us whether a trajectory worked. Novelty tells us whether it added a new way to work. Neither tells us, alone, whether training on it helps.

5. Correct first. Different second. Useful before reward.

The safe ordering is lexicographic. Execution and evidence are verified first. Verification and reward eligibility determine whether a trajectory can become a positive training example. Novelty may then distinguish quality-eligible trajectories, while novel failures enter an adversarial evaluation or curriculum archive—not positive imitation. Local competition prevents a globally common strategy from erasing a rare but high-quality one.

This still leaves the causal question. A trajectory can be valid, correct and different yet produce a harmful gradient. Reliquary's next layer therefore treats entropy, hidden-state movement, termination and content distance as observations. The label is downstream change on a fixed, validator-controlled probe set.

6. Foundation live. Novelty influence: 0%.

Reliquary PR #159, merged as commit 36862e4, is now present in deployed image 07b3fac315a2. A direct health receipt at 12:33 UTC on 23 July reported utility telemetry enabled at schema v1, 185 successful writes, zero failures and window 25,005 as the last completed write. Content cooldown was complete from R2 and local state.

Canonical content identity is therefore live: each validator-rendered prompt receives a full SHA-256 identity used for exact-alias deduplication and cooldown. In archived window 25,002, all 16 selected groups carried a content hash and all 16 hashes were distinct. This is not semantic deduplication and does not prove Sybil resistance against coordinated paraphrases.

The private observation path is also live, but deliberately powerless. It retains selected winners and proven no-emission comparisons with exact rewards, termination, likelihood, entropy and representation-shift summaries. Operator identifiers are HMAC-pseudonymized; files are private, atomic and bounded by retention. Any telemetry failure is fail-open. It cannot reject a miner or change proof order, difficulty rank, selection, rewards, cooldown, miner payloads, emission accounting or training. No behavior descriptor, novelty archive, U_step label, U_align label or novelty-shaped rank exists in production. The full boundary is in the utility-foundation specification.

The proposed strongest offline label, U_step, would perform one virtual microstep on candidate group g and measure the change in probe loss. A proposed first-order proxy,U_align, would compare the candidate gradient with the probe gradient. Neither is implemented by the merged telemetry foundation. Any implementation must reproduce the exact checkpoint, mask, forced-token carve, advantage calculation, KL reference and loss used in production. A proxy could activate only if it predicts those labels out of sample.

This is the core discipline: high entropy is not a reward; hidden-state distance is not a reward; novelty is not a reward. They are candidate features in a testable model of utility. The economic oracle remains disabled until downstream improvement survives held-out checkpoints, alternative baselines and live shadowing.

7. Brick by brick, from observation to a September experiment

The next milestone is not “turn on novelty.” It is a sequence in which each step can fail without contaminating emission accounting or training. The immutable observation image and its health receipt are now live. The next gate is evidence: resume a checkpoint-changing training program, because utility cannot be calibrated across learning states without multiple policy revisions. Collect at least 256 complete windows total and three checkpoint revisions in each environment, with at least 99.9% writer success and less than 1% missing utility fields. The write counter alone is not a progress bar for this gate.

Labels should be generated on a separate research GPU. Splits must be chronological by checkpoint, with every content identity and pseudonymous operator confined to one split. Math and Code calibrate separately. The system must replay complete candidate populations, not only historical winners, and compare against difficulty-only, random-within-plateau and arrival-order baselines. Only then should a proposed ordering run live in shadow mode with zero effect on proof order, emission shares, cooldown or training.

The first possible activation is intentionally narrow: calibrated utility may reorder only candidates with exactly equal difficulty. It cannot promote a lower-difficulty group above a higher-difficulty group. The difficulty value would remain only a rank. Under the deployed auction, only the first ranked canonical-content candidate that passes deferred proof can win; every selected group earns one uniform slot, and per-hotkey slot shares feed the EMA used for weights. Activation requires positive top-eight U_step uplift with a block-bootstrap 95% confidence interval above zero, stable direction across held-out checkpoints, proof-path p95 overhead below 5%, and no regression in duplication, concentration, termination or quarantine.

8. The Montréal draft preregistration

By September, the honest target is not an open-ended learner. It is a live, versioned map of verified strategies; a causal label pipeline; and a fixed-budget comparison showing whether archive-aware selection improves training. The draft's primary trained comparison is A, difficulty-only selection, versus D, utility-within-plateau selection, for 300 optimizer steps with identical rollout-generation and retained training-token budgets. A variance-only pilot will set the minimum detectable effect and final paired-seed count before confirmatory outcomes are inspected; ten matched seed pairs are the planning floor. Five pairs are explicitly insufficient for a conventional two-sided exact sign test, whose smallest possible p-value would be 0.0625.

Random-within-plateau and arrival-order remain offline replay baselines; a quality-diversity batch sampler is reserved for a later study. The primary endpoint is the paired change in macro-averaged held-out Math + Code pass@1 at step 300. Inference is seed-level: a paired randomization test plus a cluster-respecting 95% interval, with per-domain results and all runs reported. There is no optional stopping, and neither domain may regress by more than one percentage point. The sample size, exclusions, compute budget, evaluation set and analysis code must be published immutably before collection begins.

A later quality-diversity arm would sample batches across verified niches rather than raw frequency. Its primary outcome remains held-out performance; coverage and QD-score are supporting outcomes, not substitutes. We should release configuration, content hashes, descriptor version, archive snapshot, checkpoints, aggregate figure data, and negative trajectories. The article should be falsifiable by construction: if archive-aware selection does not beat the fixed control, it stays out of the market.

9. More miners are not, by themselves, more ideas

Distributed rollout systems change the economics of exploration, not its objective. The Pluralis Mac post-training run makes that distinction vivid: fourteen heterogeneous Macs across four countries generated asynchronous int8 rollouts while one B200 trained in bf16; held-out PaperSearchQA pass@1 rose from 29% to 63%. Yet an early run collapsed toward short recall answers despite equal success reward. Changing the task mix and loss geometry recovered the desired behavior.

The lesson is not that decentralized workers automatically diversify learning. Workers produce samples. Selection, loss and memory decide what survives. Reliquary's miners are naturally parallel explorers; a versioned archive can make their useful differences persistent. Without explicit descriptors, admission rules, local competition and replay, a larger swarm is only a larger sampler.

10. What could go wrong

Novelty search is attractive precisely because it expands what the system values. That also expands its attack surface. The failure modes should be treated as first-class protocol requirements, not a footnote after launch.

Mitigations follow from the threat model: content addressing and semantic duplicate taxes; frozen, versioned descriptor snapshots; rank percentiles instead of unbounded distances; minimum correctness and utility gates; a separate failure archive; descriptor agreement audits against human strategy labels; per-operator diversity concentration reports; and deterministic archive rebuilds. Raw traces and vectors stay private unless explicitly redacted for research release.

Even a working archive is not an open-ended system. POET showed how environments and solutions can co-evolve and transfer stepping stones; Enhanced POET made the environment space and progress measure richer; XLand evaluated agents against changing task distributions. A 2026 information-theoretic treatment argues that novelty and learnability still do not suffice: new information must sustain useful capability growth. That is a higher standard than “the number of occupied cells went up.”

11. The archive is the objective

Reliquary's current market asks a precise question: which prompts produce contrast for this checkpoint? The live data shows that the answer is clustering at two successes out of eight. That is a meaningful result. It also reveals the plateau where the next mechanism must earn its place.

The next question is not “which trajectory looks strangest?” It is “which verified, behaviorally distinct trajectory creates a useful update that our current repertoire would otherwise lose?” Answering it requires network memory, causal labels, shadow operation and the willingness to keep the feature off when evidence is weak.

More miners give more samples. An archive gives those samples memory. A causal utility test decides whether that memory deserves to shape the model.

Today, Reliquary runs a live difficulty auction over verified rollout groups. The September target is smaller and more serious than a claim of open-endedness: make every candidate reproducible, make useful difference measurable, and bring a shadow archive whose failures are as visible as its discoveries. Brick by brick is not the cautious version of the work. It is the protocol.