Referential Sieve and H-group: an Ideological Review and Convention Design Proposal

This document compares the most common 2p systems, Referential Sieve and H-group. I analyze their relative strengths with the explicit purpose of proposing design goals for a hypothetical system to incorparate power from both.

In my ideal world, I'd like these analyses to be community efforts. But due to Hanabi's complexity, it is impossible to prove objectively what is a strong convention: in fact, as far as I'm aware, "strong" doesn't even have an objective meaning when applied to an individual convention. In lieu of proof, we can only theorize, with our evidence being hueristics backed up by games played. In addition to limiting knowledge, lack of proof means individuals often disagree, limiting even shared community beliefs. That said, it's still worthwhile to have our understanding documented. Although this work is entirely my point of view, I would enjoy converting it to (or starting a new) a community document given sufficient interest. As part of that, I would expect some (many) of my thoughts to be overidden due to others' disagreement.

With that said, let's get into the comparison.

Advantages: Referential Sieve

These are traits with clear benefits included in Referential Sieve but not in H-group, ordered roughly by how clear the benefit is.

Tempo Clue Focus

Referential Sieve has fill-in clues. Clues which reveal a card as playable do not have other conventional meaning (beyond what is meant from good touch). This means that when the only safe action in hand is to play a touched card, Referential Sieve can always get it (in no variant). On the other hand, H-group sometimes cannot, since H-group clues prioritize focusing new cards over filled-in cards.

Starting hand chop moves

After the good 1s have played, new cards drawn are much more likely to be trash than old cards. Referential Sieve takes advantage of this fact by having the default discards start out coming from the deck, whereas in H-Group the default discards are from the starting hand.

Unlock promise

Unlock promise is amazing. It is efficient in the short and long term and theoretically well justified. Every serious two player system should use some version of it. It's hard to be confident in exactly how far to take it, but free choice between exactly known possibly-connecting playables and known trash at 2+ pace is almost surely a sufficient condition.

Maximal play clue predictibility

Despite (except in pathological cases) Referential Sieve always being able to tell new cards to play, the new card is deprioritized as a play clue focus. This means that referential sieve players can be confident that newly-drawn cards won't interfere with play signals, while still always being able to tell new cards to play.

Referential information

The empathy of the possible clue space is determined by the hand, not conventions. If we want to maximize the information conventions provide, we desire to minimize the overlap between information provided by empathy and information provided by convention. Referential clueing means play signals are often given to untouched cards, significantly limiting redundant information.

Universal rank saves

I…think I prefer universal rank saves in 2p? I'm actually pretty unsure on this. Being able to save any card is great. Not having control over what other cards a save clue happens to touch is less great. Universal rank saves are definitely better later on. It's unclear to me which is better early in the game.

Advantages: H-Group

These are traits with clear benefits included in H-group but not in Referential sieve.

Discard-oldest

In comparison to discard-newest, discard-oldest gives the team more time to prepare for chops, allowing flexibility in what turn save clues can be given, leading to safer low clue states. It also scales better to difficult variants, where the extra time allows for better planning to save awkward cards.

Tempo

H-group asks for play signals to be given quickly. This is to avoid discarding good cards in their own hand. Referential sieve does not get tempo as quickly, leading to situations where you either risk losing playables or risk wasting a clue on a poor lock hoping you hold the unlocker.

Scream discards

I like scream discards. I wish free choice between playing and discarding had an officially defined meaning in Referential Sieve beyond discard modulation opportunities.

Direct information

When cards are play signaled, they are often exactly known. This has several benefits. First, the team will less often lose delayed playable cards on chop. Second, the team will not bomb dupes as often. Third, with exactly-promised cards, players can more often find reasons to disbelieve the signal due to context, but context deserves its own point.

Also, in variants such as Duck, Cow & Pig, and Throw it in a Hole, direct information is important because the team gains significantly from knowing exact identities of the cards their play clues focused.

Context

I find basic H-group context much easier than ref sieve, and H-Group teams tend to be able to consistently read deeper into context than Referential Sieve teams. Why is that?

1. Tempo expectations make late play clues often unbelievable.
2. More generally, stronger behavioural expectations means players can make more confident reads about how their teammates would handle situations.
3. Direct info means that players only have to rule out one possibility rather than many to determine if a clue is unbelievable.

On the other hand, tempo is not expected in referential sieve. Blocked play signals tend to remain blocked. The main consistent way to achieve focus inversion is to start with a (costly) lock.

Stronger context and particularly the ability to have the clue's empathy make the meaning unbelievable scale well for handling the challenges posed by Hanabi variants.

Play clue flexibility

Because rank and color are both play clues, H-group conventions provide more choice with regards to what cards get picked up when a play signal is given.

Shared positives between both systems

Good Touch Principle

Good touch principle is good. It is important to bias conventions towards providing information to good cards. The exact identities of trash cards are much less useful to distinguish.

Single card focus

When a clue is given, it carries conventional meaning about a single card in addition to its Good Touch meaning. This is useful to limit mental drain.

Tempo clue chop moves

1s order chop moves

Elimination notes

Every convention system attempting to be strong will have some discards which look too bad to be believable: these discards promising another copy somewhere is a good convention.

Shared negatives between both systems

Good Touch Principle

Good touch principle traditionally does not allow for cross-hand dupes to be good. But there are many gamestates, especially towards the end of the game, where the team wants to hold on to both copies. In these gamestates it is bad for clues to interpret cross-hand dupes as conventional trash.

Priority

I have problems with priority implementations more than I do with priority inherently. But it's part of my identity to oppose priority at this point, so I can't stop now.

Questionable characteristics of both systems.

Discard-default

There are some strong score-hunting mixed play/-default systems out there, waiting to be found. I would recommend Permission to play for anyone looking for a launching point to build off of.

Clued-unclued categorization

Why exactly do we have such a distinction, particularly in pink and rainbow heavy variants?

Bluffs

Are fill-in reclue bluffs really necessary? It may be viable to change a lot of them to become fill-in sacrifice signals.

Summary

Concept	Referential Sieve	H-Group	Comparison
Information	Referential	Direct	RS wins on empathy efficiency. H wins on dupe/bomb avoidance, saving delayed playables, and focus inversion.
Chop	Newest	Oldest	RS wins on chop moving the starting hand. H wins on discard-oldest.
Tempo Clue Focus Priority	On	Off	RS wins. Being able to get the only safe action in hand is good.
Consequence of Delaying Play Clues	Change the expected discard	Change the expected play	RS more often can save cards without using a save clue. H-Group can give tempo without chop moving trash and more consistently focus inverts.

The Future of 2p

Referential sieve continues to be optimized, but with its tempo issues and discard-newest being impossible to address without fundamentally modifying the system, I personally would appreciate alternative options. Turbo is a proposed option for the future of 2p. While I like Turbo, I find it chop moves too much, and feels quite similar to Referential Sieve. I would like to play a 2p score-hunting system that does not yet exist incorporating each system's strengths.

I propose the system be designed to satisfy the following traits:

• The starting hand is chop moved.
• Chop is discard-oldest card not chop moved.
• Immediately signaling unclued (or otherwise awkward) playable cards while good touching is expected at relatively high clue counts (probably at least 2-3 available tokens) and does not chop move.
• Fill-in play clues chop move.
• Prompts do not chop move.
• Older (especially starting hand) chop-moved cards are easier to discard signal compared to newer chop-moved cards.
• Play clues are referential (probably/often).
• Future clue availability is predictable, at least for the clue giver, but ideally for the receiver as well.
• Frequent focus inversion effectively deals with blocked signals and assists mitigating downsides of hanabi variants.

Things that feel worth exploring but have uncertain value

• Less strict dichotomy than referential sieve's color=play, rank=save
• Bluff relaxation in favor of known sacrifices

Things that could be worth exploring but have highly uncertain value:

• Play-default states
• Clued-unclued categorization in certain variants
• Asymmetric Alice/Bob conventions

And, that's it! If you have thoughts, ideas to add, disagreements, or would like to experiment, reach out! I'm sodiumdebt in the Hanabi Central Discord server and have posted this in the Convention-Development forum channel there.