docs/design-journal.md

# Design Journal

Internal engineering history for FUTURE-TS. The papers, CHANGELOG, and README
describe the benchmark that shipped. This file captures the internal decision
path that got it there, so future contributors can understand why specific
pieces exist without having to reconstruct the history from git.

## Naming

The repository shipped publicly as FUTURE-TS `v0.1.0`. Internally the work
happened in three overlapping waves, labeled `v1`, `v2`, and `v3` in commits
and scratch notes.

- **internal v1** — the original scaffold. Schemas, tier-aware validation,
  capability vector, scoring pipeline.
- **internal v2** — integrity and scoring repair. Closed the scoring
  loophole, added paired point/probabilistic metrics, added the pretraining
  data manifest (honest-but-optional), added the rank-based aggregate with
  bootstrap CI, separated capR from capA, deprecated capD because no task
  actually carried a utility-kind metric, renamed the headline scalar to
  `tier_weighted_score`.
- **internal v3** — public-preview readiness. Promoted the pretraining manifest
  to a validator-enforced requirement on a new strict benchmark bundle,
  restored capD with an actual newsvendor task, added the sealed runner MVP,
  added the multi-budget runner, added the TSFM.ai catalog manifests, and
  added the external submission onramp.

The public `0.1.0` preview consolidates those internal waves. The benchmark
bundle that had been called `benchmarks/v1-strict/` was renamed to
`benchmarks/v1/`. The permissive original 24-task surface stayed at
`benchmarks/v0/` for regression fixtures.

## Why capD was deprecated and then restored

The scaffold defined capD as the mean skill on tasks whose primary metric
had kind `utility`, but no task in the v0 or v2 empirical suites actually
carried such a metric — so capD was null in every real report. Internal v2
removed capD on the principle that a dimension that is always null is
weaker evidence than one that is absent.

Internal v3 restored capD once we had a concrete utility-kind task:
`public_dev_inventory_holding_cost`, a newsvendor framing with an explicit
asymmetric cost function (`h=0.5` per over-forecast unit, `p=2.0` per
under-forecast unit). A symmetric MAE minimiser loses to a slight
over-forecaster by design, so capD has genuine discriminatory power rather
than being a label wrapping the same signal as capF.

## Why `overall_score` survives as an alias

We renamed the headline scalar to `tier_weighted_score` in internal v2. We
kept `overall_score` in the report format as an identical-value alias
because external consumers of the report format already referenced the old
key. The names are intentionally equal; do not rename either one alone.

## Why the sealed runner is an MVP rather than a service

The local runner enforces the integrity contract that matters for
cross-model capE comparability and for the `execution_mode="sealed"`
property: CPU / wallclock / memory caps, Linux network-namespace isolation,
platform-signed timestamps, and integrity-hash rebuild against the
predictions the subprocess produced.

Kubernetes job isolation with GPU quotas, cryptographic attestation of
container digests, per-prediction platform-measured latency, and
artifact-bucket immutability are follow-ups for a hosted multi-tenant
service. The MVP intentionally does not implement them because the
benchmark package should run reproducibly on any workstation; the hosted
service is a separate program that consumes this package.

## Why the multi-budget runner is a wrapper rather than a rewrite

The single-budget TSFM.ai runner is ~1700 lines. Refactoring it to support
multi-budget execution natively would have been risky and invasive. The
wrapper in `future_ts.multi_budget` monkey-patches `_forecast_parameters`
per budget, calls the existing runner once per budget with a budget-specific
`context_length`, and merges the per-budget submissions with recomputed
integrity hashes. Budgets that require actual weight updates (`peft`, `ft`)
are skipped with a warning because the hosted inference API cannot provide
them; they become operable once the sealed runner accepts model-author-side
training at submission time.

## Why the strict benchmark is separate from the v0 benchmark

The v0 benchmark surface predates the pretraining-manifest system and is
used by regression tests, scaffolding fixtures, and older tutorial content.
Forcing the manifest requirement onto v0 would break historical submissions
and documentation examples that are useful exactly because they are
permissive.

`benchmarks/v1/` is the strict public-preview surface that submissions should
target. It sets `require_pretraining_manifest=true` so missing manifests are a
structural rejection rather than a silent `pretraining_overlap=None`.

## Sealed runner relative-path bugfix

An early iteration of the sealed runner passed the submission script into
the subprocess as a path relative to `working_dir`. CI caught that when a
caller passed a repo-relative script path and a working directory that was
not a descendent of the repo, the subprocess `cwd=workspace` made the
relative path resolve against the workspace instead of against the repo —
so the script could not be found. The fix: resolve both
`submission_script` and `working_dir` to absolute paths before
`subprocess.Popen`. There is a regression test
(`test_relative_script_path_is_resolved_before_exec`).

## Task paths in the v1 bundle

`benchmarks/v1/benchmark.json` references task cards at
`../v0/tasks/*.json`. The tasks themselves are shared across `benchmarks/v0`
and `benchmarks/v1`; only the bundle configuration (strictness, task list,
tier weights, track list) changes. This keeps the single source of truth
for task cards in `benchmarks/v0/tasks/` regardless of how many benchmark
bundles reference them.