Bug in custom implementation of FSRS-5

[RichuSuresh]

Hi! I was wondering if any of you guys could help me fix an issue(?) in my implementation of FSRS on a web-app I'm working on. It seems to be producing intervals that don't line up with the intervals on the simulator https://huggingface.co/spaces/open-spaced-repetition/fsrs4anki_previewer

Specifically, I'm trying to test with 7 "easy" repeat ratings when reviewing a card on the same day. The intervals I get from my implementation are as follows:
15.691049999999997
37.027032023230845
87.37472001232312
206.18292312605706
486.54116182357825
1148.1178875512842
2709.276803579464
(all in number of days)

However, the simulator gives the following:
0.0d,16.0d,5.0m,3.4y,25.5y,100.0y,100.0y,100.0y

I've checked my implementation against the formulas given in The Algorithm https://github.com/open-spaced-repetition/fsrs4anki/wiki/The-Algorithm and they all seem to line up correctly. I'm not sure where I'm going wrong. I saw that the actual implementation in anki has some sort of random fuzz to it? I haven't implemented any such fuzz so maybe that's why my values are different?

I've pasted my implementation below, it includes a for loop at the bottom which outputs the results for 7 easy ratings so you guys can reference. I'd appreciate any help on this, thankyou!

from datetime import datetime, timedelta, timezone

w = [0.40255, 1.18385, 3.173, 15.69105, 7.1949, 0.5345, 1.4604, 0.0046, 1.54575, 0.1192, 1.01925, 1.9395, 0.11, 0.29605, 2.2698, 0.2315, 2.9898, 0.51655, 0.6621]

def FSRS(card, grade):
    gradeDict = {
        "forgot": 1,
        "hard": 2,
        "good": 3,
        "easy": 4
    }
    DECAY = -0.5
    FACTOR = 19/81
    REQUESTED_RETENTION = 0.9
    grade = gradeDict[grade]
    today = datetime.now(timezone.utc)
    if "lastReview" not in card:
        card["lastReview"] = today

    lastReview = card["lastReview"]
    

    daysSinceLastReview = (today - lastReview).days

    if "difficulty" not in card:
        card["difficulty"] = w[4] - math.exp(w[5] * (grade - 1)) + 1
    else:
        changeInDifficulty = -w[6] * (grade - 3)
        originalDifficulty = card["difficulty"]
        linearDamping = originalDifficulty + changeInDifficulty * ((10 - originalDifficulty)/9)
        card["difficulty"] = w[7] * (w[4] - math.exp(w[5] * (4 - 1)) + 1) + (1 - w[7]) * linearDamping
    
    if "retrievability" not in card:
        card["retrievability"] = 1
    else:
        card["retrievability"] = (1 + FACTOR * (daysSinceLastReview/card['stability']))**DECAY
    
    if "stability" not in card:
        card["stability"] = w[grade - 1]
    elif "stability" in card and daysSinceLastReview <= 0:
        originalStability = card["stability"]
        card['stability'] = originalStability * math.exp(w[17] * (grade - 3 + w[18]))
    elif grade == 2 or grade == 3 or grade == 4:
        originalStability = card["stability"]
        card['stability'] = originalStability * (
            math.exp(w[8]) *
            (11 - card['difficulty']) *
            originalStability**-w[9] *
            (math.exp(w[10] * (1 - card['retrievability'])) - 1) *
            (w[15] if (grade == 2) else 1) *
            (w[16] if (grade == 4) else 1) +
            1
        )
    else:
        originalStability = card["stability"]
        card['stability'] = w[11] * (card['difficulty']**-w[12]) * (((originalStability + 1)**w[13]) - 1) * math.exp(w[14] * (1 - card['retrievability']))
    

    nextInterval = (card['stability'] / FACTOR) * ((REQUESTED_RETENTION**(1/DECAY))-1)
    print(nextInterval)
    nextInterval = today + timedelta(days=nextInterval)
    card['nextInterval'] = nextInterval
    card['lastReview'] = today

    return card

card = {}

for i in range(7):
    result = FSRS(card, 'easy')
    card = result

[L-M-Sherlock]

Your result is correct. But you shouldn't allow the user to press easy in the same day twice.

[L-M-Sherlock]

Yes, they should be able to do that. It's useful if they learn some very different subjects in the same time.

[RichuSuresh]

interesting! And should they be able to do this on the fly? So let's say they complete a review of a deck in day 1, then can they change the parameters on day 5?

If so, should the FSRS values be recalculated based on these parameters? Or can they remain the same and the algorithm would automatically adjusts to the changes?

If the algorithm allows the FSRS values to adjust themselves automatically, how robust is this? What if the user decides to change the parameters (like retention rate) every day?

[L-M-Sherlock]

In Anki, we recommend users optimizing their parameters once a month. FSRS hasn't provided any auto-optimization feature.

[RichuSuresh]

ahh I see. And when you say "optimise", are you referring to the method used to adjust all of the fsrs values (difficulty, stability, nextInterval e.t.c.) based on the new parameters the user input (the new list of default parameters and the new desired retention rate)?

If so, how does this work?

[L-M-Sherlock]

I mean the method used to optimize the parameters of FSRS. Here are the optimizers:

https://github.com/open-spaced-repetition/fsrs-optimizer

https://github.com/open-spaced-repetition/fsrs-rs

You need prepare review logs.