Why Not Skip Straight to Higher Elo?

The opponent that finds your bracket fastest is not always the experiment that estimates your rating best

The first games looked absurdly easy.

A strong assisted chess agent was playing a deliberately weakened Stockfish. It won, the ladder moved up, and another block began. Why spend paid model calls confirming what we already seemed to know? Why not jump straight to 2000, 2400, or wherever the model probably belonged?

That instinct is partly right. It also hides two different statistical jobs.

1. Bracketing: find an opponent near the player's strength quickly.
2. Estimation: collect enough information to estimate that strength and its uncertainty.

A high starting anchor can make the first job much cheaper. It can make the second job much worse.

We did not answer the question by changing a live experiment. We built a provider-free simulation, froze its seed and design, and asked what would have happened under several prospectively declared starting rules. The result is a useful no-free-lunch theorem in miniature: skipping higher is efficient when your prior is right, expensive when it is wrong, and not automatically precise even when an oracle tells you approximately where to start.

The ladder we actually use

The exploratory ladder operates in blocks of 10 clean games. The language model plays Black. After each block, its score is

score = (wins + 0.5 × draws) / 10

The transition rule is fixed:

• score at least .90: raise the Stockfish anchor by 400 Elo;
• score from .80 through .8999: raise it by 300;
• score from .65 through .7999: raise it by 200;
• score from .35 through .6499: repeat that anchor once; and
• score below .35: stop.

After a repeat, the repeat block alone decides whether the ladder advances or stops. Provider-error records do not count as clean games.

This rule was designed to move rapidly through obvious mismatches and spend more games near an uncertain boundary. But its starting anchor was 1320. For a player around 3000 in this local protocol, several early blocks will be close to foregone conclusions.

The question is not whether those blocks cost games. They do. The question is what we give up when we avoid them.

The audit we ran instead of improvising

We crossed eight known true ratings—1100, 1320, 1500, 1700, 2000, 2400, 2800, and 3200—with two draw propensities, 0.20 and 0.70. For each of those 16 scenarios, we simulated four prospective designs:

• start at 1320;
• start at 1720;
• start at 2120; and
• start at an impossible oracle reference near the known Black-adjusted fair opponent.

Every design used the same downstream transition policy. Every visited block entered the same all-anchor Davidson win-draw-loss likelihood used by the current rating analysis. Each cell received 150 replicates: 9,600 simulated ladders in total. No provider and no chess engine were called.

The oracle matters because it separates two ideas that are easy to conflate. It knows approximately where the fair opponent is, so it has almost no bracketing problem. If “start near the truth” were also sufficient for precise rating, the oracle should dominate everything.

It did not.

Higher starts bought speed

Across the deliberately uniform scenario grid, the fixed starts produced this summary:

Starting anchor	Mean games	Mean rating RMSE	Fail-closed profile coverage	Overshoot / undershoot scenario cells
1320	39.88	113.19	.9521	2 / 12
1720	29.47	277.00	.9483	6 / 8
2120	22.37	477.62	.9600	10 / 6

Moving from 1320 to 1720 saved 10.41 games on average. Moving to 2120 saved 17.51. If the only objective were “reach a stopping point with the fewest model calls,” higher would look excellent.

But mean RMSE moved in the opposite direction. Starting at 1720 more than doubled it. Starting at 2120 more than quadrupled it.

The coverage column does not rescue the high starts. Coverage can remain high because an interval is extremely wide. An experiment can honestly admit “we have almost no idea” and cover the truth most of the time. Precision and coverage are different properties.

These cross-scenario averages are not a leaderboard and not a deployable policy score. The grid gives equal weight to every chosen truth; it is not an empirical prior over the next model. Its purpose is to expose the tradeoff.

What overshooting looks like

Consider a true local rating of 1100. With the fixed 35-Elo Black correction, the fair opponent is about 1065. A 2120 start overshoots that opponent by 1055 points.

In all 150 replicates at each draw propensity, the ladder stopped after the first 10-game block. That sounds efficient. Statistically it was brutal:

Draw propensity	Mean games	Saturated blocks	Rating bias	Rating RMSE	Mean profile width
.20	10.00	87.33%	-540.61	578.28	1468.97
.70	10.00	66.67%	-567.10	606.95	1475.02

The estimator often saw one nearly all-loss block, hit or approached its lower rating boundary, and had to infer both rating and draw propensity from that tiny record. It stopped quickly because the mismatch was obvious. It did not learn the player's strength precisely.

This is the asymmetric danger of a high start. Undershooting a strong player spends extra games climbing. Overshooting a weak player can terminate almost immediately with little more than “substantially weaker than this anchor.” If the goal is a finite point estimate rather than a coarse upper bound, those are not equivalent errors.

When skipping really does help

Now take a true rating of 3200. The Black-adjusted fair opponent is about 3165, so both 1320 and 2120 undershoot. Here the higher start removes games that are very unlikely to discriminate among plausible high ratings.

Draw propensity	Start	Mean games	Saturated blocks	RMSE	Mean profile width
.20	1320	69.47	51.15%	76.09	289.86
.20	2120	49.80	32.80%	69.54	288.12
.70	1320	73.13	37.92%	66.56	293.38
.70	2120	52.40	19.97%	82.96	295.37

Starting at 2120 saved 19.67 or 20.73 games, depending on draw propensity. It also removed many saturated blocks. Rating error and interval width stayed in the same broad range, although the RMSE direction was not uniform across the two draw settings.

This is the real case for skipping higher: independent evidence has already made weak-player scenarios implausible, so low anchors mostly generate near-certain outcomes. The claim is conditional. It is not “higher is always better.” It is “higher is cheaper when the prior is trustworthy and the cost of overshoot is acceptable.”

Why easy games can still help a likelihood

An Elo shortcut often uses only the last block:

rating = opponent + color correction
       + 400 × log10(score / (1 - score))

Under that shortcut, an early 10-0 block looks disposable after a later 5-5 block appears. The later block is closer to the decision boundary; surely it is the only one that matters.

Our publication estimator does something different. It fits the complete W-D-L record over every visited anchor and estimates draw propensity alongside rating. Earlier blocks constrain which strength and draw combinations could have produced the entire path. A saturated block carries less local information than a balanced one, but “less” is not “zero.” Several anchors can also stabilize a fit that would otherwise rest on one or two noisy blocks selected by the stopping rule.

This explains the apparent oracle paradox.

For a true rating of 2800 with draw propensity .20, the oracle-near reference started at 2800 and used 18.53 games on average. Only 1.08% of its blocks were saturated. Yet its rating RMSE was 209.09 and its mean profile width was 401.36. At true 3200 with the same draw propensity, it used 19.47 games, with RMSE 101.31 and mean width 370.00.

The oracle solved bracketing. The unchanged policy then saw a near-even block, usually authorized one repeat, and stopped with roughly two blocks. That was enough to locate the neighborhood and not always enough to estimate rating plus draw propensity tightly.

An oracle starting point is therefore a lower reference for bracketing work—not a lower bound on statistical error.

What the simulation does not prove

The audit is deliberately favorable to our estimator. It generates outcomes from the same Davidson family it later fits. It treats the 35-Elo Black penalty and configured Stockfish anchors as known. Every game is Black. It excludes provider failures, opponent drift, opening diversity, within-game dependence, and model misspecification. Ordinary profile intervals are not corrected for adaptive stopping or starting-anchor selection.

The 16 scenarios are a stress grid, not a probability distribution. Reporting their unweighted average does not say that 1320 is optimal for the next model. If we were almost certain the next system exceeded 2500, the low end of the grid would deserve little prior weight. If that confidence were marketing optimism rather than independent evidence, deleting it would hide exactly the failure mode the audit exposes.

The maximum Monte Carlo standard error for a binary rate in this audit is 0.0408. Small coverage differences should not be narrated as discoveries.

The rule for future ladders

We will not change the active ladder or discard its lower anchors. Those games were generated by its declared measurement procedure and remain part of its rating likelihood.

A future ladder may start higher when all four conditions are met:

1. Independent calibration: the starting anchor comes from evidence not generated by the ladder outcomes it will evaluate.
2. Declared uncertainty: the prior range is recorded, not only its most flattering point estimate.
3. Asymmetric risk: the design explicitly prices the cost of overshooting a weak player against the cost of making a strong player climb.
4. Protocol versioning: the new start rule is frozen before outcomes, and every block it subsequently observes remains in the likelihood.

If the plausible range spans 1100 through 3200, 1320 is slow but robust. If independent calibration nearly rules out the weak end, 1720 or 2120 can save substantial paid games. If precision is the objective, starting higher may also require a fixed minimum number of informative blocks rather than reusing a stopping rule built mainly for bracketing.

That final distinction is the answer.

Skip easy anchors prospectively when independent evidence justifies it. Do not confuse a fast bracket with a precise rating, and do not rewrite a live protocol after its results begin to arrive.

---

Reproducibility note

The frozen report is prospective_start_anchor_audit.json, SHA-256 03c2970552d674337a7a56bb480d5a8d149ea9cf065717ed153bbb2242491b6d. Its provider-free generator is prospective_start_anchor_simulation.py, and the separately versioned protocol and quantitative memo are in PROSPECTIVE_START_ANCHOR_AUDIT.md. The report binds the generator, exact downstream simulator, rating model, and live transition-rule source by SHA-256. Rebuilding with seed 20260805 and 150 replicates per cell reproduces the report byte for byte.