# Blind Machine curated protocol bundle — cohort_histogram.
#
# This manifest is the ONLY file the Rails platform parses. The author's pure
# functions in server.py / local_project_owner.py / local_data_owner.py own the
# computation; the numbered stage files are kit-owned shims that call them (see
# docs/rfcs/0002). This file owns the governance + review surface (display
# metadata + the gate values that let compute be authorized without unpacking the
# tarball). See docs/protocol_structure.md for the schema and
# docs/protocol_catalog.md §3 for this protocol's definition.
#
# cohort_histogram is deliberately NON-GENOMIC: it aggregates a bucketed count
# over an ordered, published bucket definition (demographic age bands in the v1
# demo). It reuses the flagship's additive minimal-params BFV loop verbatim to
# prove the additive primitive is generic beyond genotypes — genomics is the
# demo domain, not the only possible one.
name: cohort_histogram
version: 1.0.0

builder:
  type: uv
  project: env

crypto: tenseal-bfv             # display hint; real deps live in env/uv.lock
computation: additive_bfv       # pure coordinate-wise homomorphic vector add
min_contributors: 25            # catalog §3 (higher than the flagship's 20)

# The published bucket definition. Every contributor MUST encode their bucket
# membership against this identical ordered definition or the blind histogram is
# meaningless — that invariant is what folding this manifest into the bundle
# SHA-256 enforces. `length` is the number of buckets B; each contribution is a
# one-hot vector h in {0,1}^B with sum(h)=1. The v1 demo uses ordered categories
# (demographic age bands); a numeric protocol would instead publish `edges`.
input:
  type: onehot_vector
  length: 10                    # B = number of buckets
  value_domain: [0, 1]          # one-hot membership; exactly one coordinate is 1
  constraint: one_hot           # sum(h) = 1 per contributor
  buckets:
    kind: ordered_categories    # or `ordered_edges` with `edges: [...]` for numeric bins
    seed: blind-v1-histogram-buckets
    labels:                     # the published, ordered bucket definition
      - "0-9"
      - "10-19"
      - "20-29"
      - "30-39"
      - "40-49"
      - "50-59"
      - "60-69"
      - "70-79"
      - "80-89"
      - "90+"

output:
  shape: aggregate_count_vector # counts in Z^B (+ append-1 sentinel -> N)
  exactness: exact              # BFV is exact in Z_t; one-hot sums never overflow
  tolerance: 0

release_policy:
  aggregate_only: true          # only the summed histogram is ever decrypted
  allowed_runs_per_project: 1   # run cap (differencing mitigation, not a defense)

resources:
  max_memory_mb: 8192
  max_wall_seconds: 3600
  max_ciphertext_mb: 512

# Collaboration scenario (docs/rfcs/0002 §4.1): who runs which role's functions.
# The author writes pure functions grouped BY ROLE into the files below; the
# numbered stage shims that call them are kit-owned framework code, materialized
# at run time (never signed into the bundle). `aggregate` = 1 project owner + N
# data owners.
scenario:
  name: aggregate
  roles:
    project_owner: { file: local_project_owner.py, functions: [ keygen, decrypt, decode ] }
    data_owner:    { file: local_data_owner.py,    functions: [ encode, encrypt ] }
    server:        { file: server.py,              functions: [ compute ] }

# Content-addressing digests. Computed at registry ingest — `bundle` by
# protocols/hash_bundle.py, `env_lock` over env/uv.lock + env/.python-version +
# pinned runner metadata. Left null in the committed bundle ON PURPOSE: the
# canonical bundle digest is computed over the raw bytes of every signed file
# INCLUDING this manifest, so a computed `bundle` value here would be
# self-referential (its own input). The authoritative digest instead NAMES the
# bundle (protocol-<name>@<digest>) and is recorded out-of-band
# (.blind-signature / DB), never inside the file it would have to hash. See
# protocols/README.md § "Why the manifest ships null hash placeholders".
hashes:
  bundle: null
  env_lock: null