639616b3e5
Folds in Shay's rulings on the spike: - ERRATA sections added to ADR-0249, ADR-0250, and both acceptance-evidence docs (statuses stay Accepted): corpus label 'real GSM8K dev holdout' -> 'CORE-authored GSM8K-style corpus (ADR-0119.2)'. Mechanism claims (wrong=0, PARITY, conservation, atomicity, chain-of-custody, 200/200) all stand; reframe not retraction. Covers the #74 review comment that repeated the label. (Also fixed ADR-0250 evidence status -> RATIFIED, which the ratification housekeeping missed.) - Two instruments named + never-blended scoreboard: authored = mechanism-correctness (200/200 stands as that); official holdout = generalization (5/500 symbolic / 0/500 corridor). public/dev-1 floors relabeled MECHANISM floors, composition-disclosed. - Censored-incidence fix: parse-INDEPENDENT tagging of 100 official cases (documented heuristic). 70/100 carry frontier markers (rate 31, fraction 28, compare_mult 20) -> (c) refuted, frontier COMMON not rare; the table is the reader-arc curriculum by real-world frequency. - Reader-practice-loop design with the four non-negotiable constraints (real held-out data only; wrong=0 gates consolidation; sealed test untouched; holdout_dev split disjoint from the measured metric). Sequencing: capability->practice->calibration-> serve; reader is the critical path for all three; serve-landing de-prioritized (0/500). Spike verdict for ruling; nothing built.
150 lines
9.1 KiB
Markdown
150 lines
9.1 KiB
Markdown
# ADR-0249: Reader→Hamiltonian Compiler — the Composition Frontier
|
||
|
||
**Status**: **Accepted**
|
||
**Date**: 2026-07-18
|
||
**Authors**: Joshua Shay + multi-model R&D (implemented Fable 5)
|
||
**Depends on**: ADR-0243 (cognitive lifecycle — `relax_to_ground`, `ProblemHamiltonian`), ADR-0244/0245 (content-addressing, f64 relaxation, `_cached_eigh`), ADR-0217 (R2 finite-integer constraint compiler — front-end precedent), ADR-0175/0191–0193 (calibrated-learning / composition-wall doctrine)
|
||
**Design record**: `docs/research/reader-hamiltonian-compiler-spike-2026-07-18.md`
|
||
**Acceptance evidence**: `docs/handoff/ADR-0249-Acceptance-Evidence.md`
|
||
|
||
---
|
||
|
||
## ERRATA (2026-07-18) — corpus mislabel corrected; status stays Accepted
|
||
|
||
This ADR and its evidence repeatedly call `evals/gsm8k_math/dev` the **"real GSM8K dev holdout."**
|
||
That label is wrong: per ADR-0119.2, `dev` (and `public`) are **CORE-authored, GSM8K-*style*** problems
|
||
in CORE's own grammar — **not** real GSM8K. Real GSM8K lives in `train_sample`, `holdout_dev/v1`, and
|
||
the sealed test. Wherever this ADR says "real GSM8K dev holdout," read **"CORE-authored GSM8K-style
|
||
corpus (ADR-0119.2)."**
|
||
|
||
**Nothing about the mechanism is retracted.** Certified relaxation, conservation, atomicity,
|
||
chain-of-custody, `wrong=0`, and PARITY against a symbolic fold of the *same* compiled program all
|
||
stand — the authored corpus is the **mechanism-correctness instrument**, and the compiler's 200/200 on
|
||
it is exactly that evidence. What falls is only the corpus label. Measured real-GSM8K reach (spike
|
||
`docs/research/dev2-frontier-measurement-spike-2026-07-18.md`): reader parses 5/500, corridor solves
|
||
0/500 — reader-gated. This erratum records that we caught our own mislabel via measurement; that is the
|
||
honest-measurement doctrine working, not a defect hidden.
|
||
|
||
---
|
||
|
||
## 1. What this ADR is — and the thesis
|
||
|
||
The generalized-lift instrument (ADR-0246 seam S4) recorded a single blocking
|
||
scope limitation: *no reader→Hamiltonian compiler existed beyond ≤5-atom
|
||
propositional problems and quadratic wells, so real problems (GSM8K arithmetic,
|
||
NL deduction) could not be ingested by the corridor at all.* This ADR builds
|
||
that compiler and nothing more — it does not touch serving, does not activate
|
||
any gate, and does not claim the corridor now beats symbolic solvers.
|
||
|
||
**Thesis — composition = certified turn sequences.** The corridor's native
|
||
space is 32-dimensional (the Cl(4,1) blade basis; `_MAX_ATOMS = 5` is exactly
|
||
2⁵ = 32, structural not incidental). A problem is therefore *not* compiled into
|
||
one large Hamiltonian. It is compiled into a **turn program** — an ordered
|
||
sequence of small relation-wells, each solved in one certified relaxation turn,
|
||
with the working state flowing turn-to-turn on the field. Composition depth
|
||
lives in the certified chain, not in matrix size. This is the "wall is
|
||
COMPOSITION" lesson (ADR-0193) finally getting its mechanism.
|
||
|
||
## 2. Components (five phases, each merged under its own PR)
|
||
|
||
| Phase | Module | What it does |
|
||
|---|---|---|
|
||
| P1 | `core/physics/quantity_kernel.py` | Quantities as null points on the Cl(4,1) conformal line; add/scale by constants as translator/dilator versor sandwiches; projective scale-invariant decode. |
|
||
| P2 | `core/physics/relation_compiler.py` | `output = scale·input + offset` → quadratic-well `ProblemHamiltonian` via versor transport (reuses `compile_quadratic_well`). |
|
||
| P3 | `evals/logic_cnf_compiler.py` | Structural formula→CNF (reuses `logic_canonical` parser; eliminate iff/implies → NNF → distribute) → `PropositionalProblem` + query `Clause`. |
|
||
| P4 | `evals/turn_program.py` | `MathProblemGraph` → ordered turn program; chained-relaxation executor; content-addressed tamper-evident turn ledger. |
|
||
| P5 | `evals/generalized_lift_instrument.py` | `arithmetic-chain` domain: corridor vs symbolic fold on the real GSM8K dev holdout; honest coverage recorded. |
|
||
|
||
## 3. Anti-hollow discipline (load-bearing)
|
||
|
||
The compiler must never *evaluate* the arithmetic or logic it encodes — else the
|
||
corridor confirms a precomputed answer rather than solving, and the instrument
|
||
is hollow. Enforcement:
|
||
|
||
- The relation/turn compilers apply the problem's **structure** as versor
|
||
operators (dilator for scale, translator for offset); the substrate's
|
||
geometric product performs the arithmetic. No Python `scale·input+offset`.
|
||
- Compiled wells are **bare projectors** — metadata `{curvature, target_digest}`
|
||
only; they carry no answer and no coefficients.
|
||
- In a turn chain the accumulator flows as a **field state and is decoded exactly
|
||
once, at the end**; intermediate values are never decoded to drive the next
|
||
turn.
|
||
- Ablation pins confirm the answer is produced by relaxation (the start state
|
||
decodes to the input, only the relaxed state to the answer, with byte-identical
|
||
Hamiltonians).
|
||
|
||
## 4. The Ring-2 correction (verified in-tree)
|
||
|
||
The design spike proposed "chain turns through the Ring-2 residual protocol."
|
||
Grounding against the code corrected this: `run_residual_protocol`
|
||
(`core/ports/residual_protocol.py:222`) is **zero-bound / non-mutating** — its
|
||
stage-5 recertification re-witnesses the subject and raises
|
||
`recertification_witness_changed` if it moved. Arithmetic turns *mutate*. The
|
||
protocol therefore certifies a *fixed* state's admissibility, not a state
|
||
*transition*, and cannot be the vehicle for the mutating turn.
|
||
|
||
Resolution (no hollow integration): the per-turn certificate is the relaxation's
|
||
own `RelaxationCertificate`; the tamper-evident *sequence* is recorded with the
|
||
Ring-2 chain-integrity **pattern** — a content-addressed `TurnRecord`,
|
||
`GENESIS_DIGEST`-linked, with `verify_turn_chain` mirroring `verify_replay_chain`.
|
||
|
||
## 5. Reproducibility — three tiers, honestly bounded (spike §4.6)
|
||
|
||
- **Tier 1 (content address): solved.** SHA-256 over explicit `<f8` LE bytes.
|
||
- **Tier 2 (matrix construction): enforced.** `geometric_product` is a
|
||
fixed-order scatter-add (no BLAS/reduction) ⇒ IEEE-754 hardware-deterministic;
|
||
it silently truncates to f32 unless handed f64, so every construction is
|
||
explicit f64 and pinned with golden-bytes canaries.
|
||
- **Tier 3 (eigensolve): bounded, not overclaimed.** The affine well
|
||
`H = c(Id − ψψᵀ)` has a 31-fold-degenerate excited eigenspace whose LAPACK
|
||
basis is build-dependent; propagator bytes are not cross-hardware identical.
|
||
Mitigation: the ground state is analytic (the target is constructed), so the
|
||
basis-invariant `phase_correlation/2` agreement is pinned, never the bytes.
|
||
|
||
## 6. Tier-1 scope + recorded frontier
|
||
|
||
Tier-1 is the **affine single-accumulator** envelope: add / subtract / multiply /
|
||
divide with constant `Quantity` operands and positive scale; propositional
|
||
deduction over ≤5 atoms. Everything else — multi-entity, transfer, rate,
|
||
comparison, fraction, partition, non-positive scale, >5-atom deduction, unknown×
|
||
unknown — is **refused with a typed error and recorded**, never silently dropped
|
||
(no-silent-caps). >5-atom deduction via ROBDD-partitioned turn programs is the
|
||
next arc, not this one (a sequencing choice, not an impossibility).
|
||
|
||
## 7. Honest measurement (the done-when)
|
||
|
||
On the sealed real GSM8K dev holdout (50 cases): **26 are Tier-1 ingestible and
|
||
solved wrong=0**; the corridor matches the symbolic fold on every one (PARITY,
|
||
delta = 0 — the field matches arithmetic, it does not beat it); the 24 refused
|
||
are all multi-entity, the recorded Tier-2 frontier. This is the honest-NULL
|
||
protocol working as designed: the deliverable is a *decidable measurement* and
|
||
*recorded real-data coverage*, not an inflated lift claim. The recognition
|
||
domain's genuine +9 lift stands separately; arithmetic is honestly PARITY.
|
||
|
||
## 8. Governance
|
||
|
||
Off-serving (A-04): every module lives in `core/physics/` or `evals/` and is
|
||
never imported by `chat/runtime.py`. No gate is activated; no manifold is
|
||
mutated (I-03). Reuses ratified contracts (`compile_quadratic_well`,
|
||
`HamiltonianCompileError`, `logic_canonical`, the Ring-2 chain pattern) rather
|
||
than forking them. Local-first CI: smoke green in-worktree at every phase.
|
||
|
||
## 9. What this ADR does NOT claim (honest register)
|
||
|
||
- Not a serving change; the symbolic `generate/` path remains the serve path.
|
||
- Not a lift over symbolic solvers on arithmetic (PARITY is the honest result).
|
||
- Not full GSM8K coverage — 52% of the dev holdout is Tier-1; the rest is the
|
||
recorded frontier.
|
||
- Not an activation of any admission gate (ADR-0246 §3.7 stays uncalibrated).
|
||
|
||
## 10. Ruling record (§8 — Shay)
|
||
|
||
**RATIFIED 2026-07-18 by Joshua Shay — ADR-0249 ACCEPTED for commit `220039f7`.**
|
||
The four design rulings (affine-only Tier-1; >5-atom deduction deferred; single
|
||
ADR; field wedge independent) are affirmed. The honest-by-construction measurement
|
||
is affirmed as the definitive outcome: 0 wrong, 26 correct, 24 strictly refused on
|
||
the real GSM8K dev holdout — the engine flawlessly executes what it mathematically
|
||
understands and halts exactly at the frontier of what it does not; PARITY with the
|
||
symbolic fold confirms the Cl(4,1) mechanics map to mathematical reality without
|
||
probabilistic drift. Next arc selected: Tier-2 arithmetic (multi-entity / transfer
|
||
/ rate / comparison / fraction), same honest-by-construction rigor.
|