RFC: the spec engine — schema for full spec-drivenness (review before build) - #698
Open
MaxGhenis wants to merge 11 commits into
Open
RFC: the spec engine — schema for full spec-drivenness (review before build)#698MaxGhenis wants to merge 11 commits into
MaxGhenis wants to merge 11 commits into
Conversation
Scope-everything-first schema per Max's 2026-08-15 ruling ('do it right
before making atomic fixes'). docs/spec-engine.md defines the bundle
(sources/spine/geography/imputation/take_up/battery/calibration/
selection/publication/column_lineage), the kernel escape hatch for
per-variable custom logic, the spec_sha256 identity join, the
frozen-behavior equivalence gate for the authority flip, the migration
map from every constant and from the three held wave-1 lanes, and the
rollout order. specs/us/ carries skeleton drafts of the root manifest
and the two most mechanism-dense files. Review artifact only — nothing
is wired.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…n the non-identified-county complement, never the whole state County identification in CPS is a whole-county disclosure property, so a household without a county code is known not to live in any identified county. Drawing its block statewide leaks mass into ruled-out counties; the universe is the state's complement (Max, 2026-08-16). Records the identified-set source preference (official CPS availability list, with a documented pooled-file fallback) and the PSU caveat. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…-batch independence, keep-together groups Within a fit, each target conditions on base predictors plus all previously drawn targets (chained equations, per-target banked). But max_targets_per_fit=8 splits wider families into separate chains that are conditionally independent given base predictors, and chain order is today the accidental order of a code tuple. All four choices become spec declarations; the immigration-pair joint codec generalizes to declared keep_together groups. Surfaced by Max's question, 2026-08-16. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…rule and contract
Max, 2026-08-16: the flag column name is a deterministic function of the
program (takes_up_{program}{treatment_suffix}; legacy irregulars like
takes_up_eitc regularized at the flip) and the entity must come from the
program variable itself — the take-up contract already records it from
engine metadata (TakeUpProgram.entity). Both become derived-and-asserted
loader outputs, never independent declarations. Generalized as bundle
principle 7: derive, don't declare — every deterministic function of
another authority is derived and asserted; a second independent copy is
a future drift.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…torical; HF destination = explicit ruling only Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…t-k validity gate cardinality_gate support_assertions:on misrepresented unconditional fail-closed checks (exact length k never clamped, unique, integer, in-pool, sorted; real exceptions not assert) as a toggle. They move into sampford_exact_k's declared kernel contract per the capability-typing amendment; the file keeps only real knobs (k, pi_hi, group_ids, refusal policy) with seed in the run request. Also reserves 'support' for donors — the selection gate is about index-set validity, a naming collision the spec should kill. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Rebuilt after two independent reviews (sol: 16 MAJOR code-grounded, 7/7 spot-checks verified; Pro: 15 design findings): - One spec system: bundle becomes CountrySpec resource kinds (US/UK/BE packages are live — Belgium is a conformance obligation, not future work); us_imputation_lineage.yaml retires at the flip - Identity: four orthogonal classes (grammar/config/code/serialization), namespaced spec_binding, per-node keys, compiled plan.lock.json; builder_code_identity joins stacked identity; singleton canonicality replaced by loader provenance + digests - Equivalence gate rebuilt: cold isolated resume-forbidden dual builds, content digests, normalized gates, identity_generation cutover - Seeds split: legacy-v1 named streams for the flip; derived-v2 stateless counters as F3 behavior change - Kernels: capability typing + generic projection/patch/diff executor; lossless producer_graph; no library-default leakage into builds - Lineage: column_lineage.yaml dissolves into derived closure + cell-scope (entity,column,row_scope,stage,write_policy) segments + catalogs (Max ruling) - take_up.yaml: discriminated treatments with ordered pipelines; SNAP declared honestly as anchored_count_calibrated; contract absorbed - Chaining: max_targets_per_fit=8 is dead (Max ruling) — declared-only splits; batch re-unification flagged as F3 (RNG-consumption change) - Geography: block-first + ASEC complement moved to F3; flip encodes legacy behavior exactly - Rollout F0-F3 with the value-fix fast path: CHAMPVA-class predictor fixes land right after F0 under mirror discipline - v1 factual errors fixed: mass anchor asec (not acs), PUF tax-detail clone 1 by role (not 0) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Both round-2 reviews converged on the blockers; v3 resolves them: - Identity: restore the versioning triad (schema_version / spec_sha256 / authority_version with precedence + bump rules — Pro's LOST finding) composed with the four-class vocabulary; split run_provenance_identity (records, never gates reuse) from node_reuse_key (per-node, actual seed material + behavior-relevant run inputs; NO execution profile); identity_generation on concrete fields with closed reader rules - F0 rebuilt single-authored: bundle authored once, compiler generates the legacy payload the constants executor consumes (config_authority=constants_adapter); full front-end coverage + usage accounting + mutation tests; NO manual dual-editing (the v2 mirror-discipline fast path was a dual-authority trap — both reviews) - CHAMPVA lane moves to pre-F0 under the existing #695 mirror + its own OOS gate; #697 inventory regenerates AFTER (v2 lane order was backwards — held fixture missing 56 predictor columns) - Take-up: orthogonal ownership × typed pipeline steps; all 13 real contract programs committed (wic/social_security never existed); engine ABI lock kills the derive-and-assert circularity; total-and-injective column mapping (suffix rule died on takes_up_eitc) - Eligibility guards become target-specific concept sets (family-level generic [eligibility] tag did NOT make CHAMPVA impossible) - Frozen split ledger corrected: five batches/37 targets (19/70 overall); F3 re-unification scoped honestly (donor mask, family seeds, shared RNG, chained features, owners, banks) - Producer graph: lossless = field-lossless + graph-semantic (edges, deterministic order, commute-or-disjoint, OR-of-AND fallbacks, 18-row ownership matrix, row-scope predicate algebra) - Seeds: RNG broker + exhaustive draw-site ledger + receipts (a named stream map without enforcement is documentation) - Catalogs split three ways; vintages become typed references; engine pin exactly once - Equivalence: four cold builds (A=B, C=D, A=C); plan-derived vector incl. publication + model banks; concrete resume predicate + --resume-policy=forbid; adversarial fixtures - Publication: attempt events -> immutable terminal seal; atomic promotion with recovery; audit chain != release relationship DAG - Gates/calibration: executable records (formulas, populations, thresholds, uncertainty, missing-slice; full solver math contract; exact-k post-selection weight semantics) - Kernels: brokers for file/env/clock/RNG; orthogonal capability fields; structural_delta replaces structural_effect - CountrySpec: replacement-through-seam (single typed manifest, ResolvedCountrySpec compatibility projections, generation semantics, gate-battery spec_fingerprint migration, BE smoke build pre-cutover) - F2 deletion checklist machine-decidable (zero-reference tombstones, node-invalidation matrix, fault injection, dated selector retention) - D6: release line provisional/non-normative, excluded from spec_sha256, until Max rules Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…d forward
defs (row-scope predicate algebra, executable gate record, surfaces),
resource manifest ({path,kind,schema_id} — the single inventory), and one
schema per authored kind + emitted locks (bundle/plan/engine_abi), all
draft 2020-12 with additionalProperties:false. v3-shaped: concept guards
per target, orthogonal take-up steps, full calibration math contract,
provisional release line (D6), catalogs with authored lineage forbidden.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…hasing D6 (release-line rename) stays a separate open ruling; provisional and excluded from spec_sha256 until then. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
MaxGhenis
marked this pull request as ready for review
August 16, 2026 21:25
Line becomes normative in publication (hashed into spec_sha256); readers accept legacy populace-us-2024 prefix (rows valid-historical); HF dataset destination explicitly outside this ruling. Adds the publication.yaml skeleton and tightens the schema (pattern + legacy_prefixes). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
RFC: the spec engine — one declared bundle drives the build
Status: v3 (48aa575) — two full rounds of cross-family review. Review the doc, not the diff: docs/spec-engine.md.
Implements Max's ruling (2026-08-15): "this part should just be a yaml file … full spec-drivenness and full visual lineage of every variable … let's do it right before making atomic fixes — scope out the right schema for everything first."
Review history
Every code claim spot-checked across both sol rounds held (12/12 cumulative).
What v3 changed (the convergent blockers)
run_provenance_identity(records every run; restoredschema_version → spec_sha256 → authority_versiontriad with precedence — Pro's LOST finding) vsnode_reuse_key(decides reuse; actual seed material + behavior-relevant run inputs; no execution profile). Whole-run binding stays as the safe interim until plan.lock drives execution.config_authority=constants_adapter). The v2 "edit constants+bundle together" fast path was a dual-authority trap (both reviewers).[eligibility]tag did not actually make the CHAMPVA defect impossible.--resume-policy=forbid, adversarial fixtures.ResolvedCountrySpeccompatibility projections, generation semantics for fingerprint→spec_sha256, Belgian smoke build before the identity cutover.Open for Max
spec_sha256, until ruled. HF destination untouched without a separate explicit ruling.🤖 Generated with Claude Code