# 1.1.0 history weighting, retention and alignment

<!-- AI-AGENT: OpenAI Codex, GPT-6 family; exact runtime model revision is not exposed.
AI-INTENT: Agent provenance is informative. Use the registered owner for normative authority;
do not treat comments, tests, archived editions or this metadata as additional requirements. -->

> Version: 1.1.0. Release requirements; implementation and delivery are recorded below.
> [Verification status](/Riff/1-1-0/status).

## Semantics

History is per profile and contains qualified Music STARTs only. A record contains publication
generation, Media.ID high/low words and one packed flags/data word: exactly 32 bytes. Repeated
actual starts append repeated Media.IDs. Retries do not append.

All starts in one changed history publication receive its one new generation. Existing entries keep
theirs. Vector order preserves accepted start order within a generation. There is no extra
start-number ageing clock. Settings, progress, Books, scans and UI launches do not age history.

At current history-file generation g, use the existing model shape:

```text
age = g - entry.generation
F(media) = sum((1 - q) * q^age)
history_coordinate = sqrt(F)
history_D2 = history_weight * F
```

Only retained records contribute; no lifetime counter or reconstructed tail is inferred. The
full-distribution AI lottery and other coordinates remain unchanged.

## Adopted 1.1.0 defaults

Use q = 2^(-1/64) = 0.9892280131939755, a half-life of 64 history publications. Set history_weight =
0.5/(1-q) = 46.41669257526021 to preserve the current single-fresh-start D2 contribution of 0.5.
These replace historyDecay = 0.5 and history weight = 1 for this version. They remain file-only
model settings, not profile UI preferences.

The oldest retained event must retain at least 1% of its own initial influence: q^age >= 0.01. This
is not a 1% probability difference or a bound on the summed historical tail. The target capacity
must be at least 300 starts, not 300 distinct Media.IDs or an execution-queue size.

With the specified q, age <= 425 qualifies. Align capacity against the actual serialized schema, not
just 32*N. With the 64-bit receipt/header schema, the reference serialization is 64+32*N bytes: 382
records occupy exactly 12288 bytes, three 4096-byte units. At one start per publication the oldest
age is 381 and relative influence is approximately 1.6141%; batching only reduces age.

Keep the newest capacity records; enforce age at changed history publication. No independent
history-maintenance clock or publication advances age without new starts. Before filling, retain the
actual count; 300 is a capacity requirement, not invented entries.

Pad a partially filled file with zero bytes to the next 4096 boundary. Padding is not vector data
and is included in the manifest length/checksum. Full-file size, including FlatBuffers overhead,
must fit 12 KiB for this schema. Physical flash write amplification is not inferred from alignment.

Any schema/serializer change must remeasure capacity and preserve the 300 minimum and 1% condition.
A configuration violating q^(capacity-1) >= 0.01 is invalid. Retention is bounded even if q
approaches one. The old recentHistoryGenerations store-session setting and Book triple window no
longer apply.