core/docs/handoff/curriculum-license-loop-2026-07/DIVISION-OF-WORK.md
Shay 923e60c136 docs(arc): reconcile the arc's records with what the units measured
Every correction here is to something I wrote earlier in this arc. Original text
is preserved in place; nothing was rewritten to look right in hindsight.

ADR-0264 §4.2 — CORRECTED block. I sized the eleven band ceilings from per-TERM
exclusivity; `resolve_domain`'s predicate is per-PAIR, which is looser, so every
ceiling was understated and one verdict was wrong: systems_software_causal is
720, not 630, i.e. reachable rather than impossible. Seven of eleven bands
cannot reach 657, not eight. Downstream conclusions survive — physics·modal (480)
is still impossible and philosophy_theology·modal is still the target.

ADR-0264 R5 + DIVISION-OF-WORK ×2 + compile_premises docstring — the
"verdict-identical" claim is exact only where the full family was READABLE. Over
MAX_PREMISE_SENTENCES, full-family compilation refuses outright, so narrowing is
verdict-IMPROVING there. Read literally the unqualified phrasing implies a
17-chain family keeps declining, which is backwards — the opposite of the
unblock.

E2 rename staleness, 4 sites: ADR-0256 (Accepted) named
`gold.py::assert_corpus_sound`, which #119 deleted; the plan's E2 item and
DIVISION-OF-WORK's §2 row, §4 heading and "renaming after that is strictly more
work" reasoning were all still written as pending.

DIVISION-OF-WORK §0b — new status table: which PR each unit is, what stacks on
what, and the four things the units found that the plan did not predict.

Adds the arc-close brief from ARC-CLOSE-TEMPLATE.md. Its §3 (falsified
assumptions) is the section that cannot be recovered from the diff: nine
entries, five of them corrections to my own claims. Records that the
articulation trigger is not merely unmet but STRUCTURALLY unmeetable — the
previous arc left it unrecorded, which is what kept it re-triggerable — and
names four pieces of stale reasoning still in the tree, with files, since that
is the class of staleness nobody greps for.

[Verification]: in-worktree on canonical CPython 3.12.13 with `uv sync
--locked`: smoke 569, deductive 291 — both baseline, since the only non-doc
change is one docstring.
2026-07-26 12:33:30 -07:00

22 KiB
Raw Blame History

Division of Work — Curriculum License Loop

Arc: curriculum-license-loop-2026-07 · Written: 2026-07-25 · By: Opus 5 Plan of record: docs/plans/curriculum-license-loop-2026-07-25.md (binding) Read before any unit: that plan §1§3, then AGENTS.md.

This file is the contract for who does which unit and why. It is written to be read cold, by a tier that was not present for the analysis. If it conflicts with a memory or a chat instruction, the plan of record and AGENTS.md win.


0b. Status 2026-07-26 — every code unit in this arc is open as a PR

Superseding §0 below (kept for the reasoning, which still holds). All units are open, unmerged, awaiting Shay's authorization.

unit PR branch state
E2 rename #119 MERGED
R5R7 query-scoped premises #120 MERGED
C practice producer #121 feat/curriculum-practice-producer open
D proposal-queue reseal #122 feat/proposal-queue-reseal (on #121) open
E1 code half (audit-ledger R7) #123 feat/audit-ledger-r7 open
parity measurement + core rust build fix #124 fix/rust-build-env-and-parity-report open
polarity R1R4/R8 #125 feat/curriculum-polarity (on #121) open
doc reconciliation + arc close this branch docs/arc-reconciliation open

Merge order: #121 first (#122 and #125 stack on it; retarget both to main after it lands). #123 and #124 are independent of everything. Merge-commit, not squash. #122 and #125 both touch core/cli_proposal_queue.py in different hunks.

Still Shay's, unchanged: Phase F content (philosophy_theology · modal), Phase G the flag, E3 the ADR vocab sweep (Haiku), and the four open rulings — of which outcome mix is now blocking rather than academic (#121).

What the units found that the plan did not predict — all four are in the PRs and the research notes, listed here because a status table is where someone looks:

  1. A committed curriculum ledger is necessarily an EARNING ledger. The plan's Phase-1 exit criterion ("a real, still-unearned ledger") is unreachable: a band clears θ_SERVE on correct NON-COMMITMENTS alone, so sealing licenses four bands on ~99% non-entailed evidence. Hence #121 ships the writer and commits no artifact, and #122's verb refuses to grant a license without an explicit flag.
  2. §4.2's ceiling table was understated (per-term vs per-pair routability); systems_software · causal is reachable, not impossible. Four bands can reach 657, not three. ADR-0264 §4.2 now carries the correction.
  3. then and therefore are taught lemmas that collide with the argument reader's control grammar — 164 atoms refuse, in the Phase F target band. The vocabulary boundary is never screened against the reader's grammar.
  4. The Rust lane-parity method cannot fail. Sabotaging every Rust kernel still yields 11/11 lane pins matching, so §6B's premise needs re-earning.

0. Status at handoff — 2026-07-25, Opus units complete

The three Opus units are done and pushed. Read the plan of record's AMENDED block first: Phases A and B falsified four premises, one of which reorders the remaining work.

Unit Branch State
A — negative-curriculum epistemology docs/adr-negative-curriculum ADR-0264, pushed. Proposed, awaiting ratification
B — volume-honesty invariant test/volume-honesty-invariant (stacked on A) pushed. 13 tests, mutation-checked, in smoke
E1 Opus half — the R7 assertion docs/e1-r7-assertion-spec this branch. E1-R7-ASSERTION-SPEC.md

Merge order: A, then B (B is stacked on A), then E1's spec — merge-commit, not squash. E1's spec branch is independent of A/B and can merge in any order.

What changed for the Sonnet units. Three things, all of which make the remaining work different from §4 as originally written:

  1. A new implementation unit precedes everything: ADR-0264 R5R7, query-scoped premise compilation. MAX_PREMISE_SENTENCES = 16 caps a family at 16 chains; at 17 the band declines every question, including ones that work today. Until this lands, no curriculum band can earn a license at any corpus size, so building Phase C first would produce a correct instrument reporting that every band is at its ceiling. R5 is fully specified and verified verdict-identical over 8,520 questions — it is a loud unit (the physics lane pins every verdict). (Precisely: identical wherever the full family was READABLE, and verdict-improving over the cap, where full-family compilation refuses outright. See ADR-0264 R5's precision note.)
  2. Phase F retargets to philosophy_theology · modal. physics · modal has a hard ceiling of 480 against a 657 threshold. Do not author physics modal volume.
  3. Phase C's producer must call the Phase B audit. AUDIT_SOURCES in tests/test_volume_honesty.py carries a declared None for curriculum_serve that fails the moment its ledger exists. That is deliberate: it is the forcing function against repeating the deduction pattern.

Standing "do not fix" — the deduction ledger exposure. 21 of 25 ratified deduction_serve bands do not clear θ_SERVE on distinct evidence (worst: 28 distinct cases inflated to 720 committed). It is measured, pinned in both directions, and must not be repaired by any unit below. The ledger is SHA-sealed, ratified, and gating a live flag; re-sealing it is Shay's ratification. If a unit's work makes tests/test_volume_honesty.py fail, that is a real signal — report it, do not update the inventory to match.

Revised order: E2 → R5R7 → C → D → E1(code) → parity.


1. The tiering criterion — read this first, it is the whole rationale

Units are not assigned by difficulty, and not by model prestige. The evidence does not support that split. In the 2026-07-24/25 window Sonnet 5 (Tier S) produced the six-worktree public_demo archaeology, found the derived_close_proposals.py::DEFAULT_SINK bug that wrote outside the repository, and found an engine-refused misclassification while building the instrument meant to measure something else. In the same window Opus 5 mis-diagnosed the CGA hot path by multiplying a microbenchmark by a call count, and made three wrong readings of this branch's own commits that survived only until they were checked.

Model tier was not the variable. Verification discipline was.

So the split is by one question:

If this unit is done wrong, does a gate catch it, or is the error silent?

  • Gate-caught — a test fails, an import breaks, a lane goes red, a hash drifts. The work is safe to execute at the cheapest competent tier, because being wrong is loud. Iterating against a red gate needs no design authority.
  • Silent — the system stays green while asserting something false. No downstream check can recover it, because the wrong thing has become the reference. These units need the tier that can reason about what the gate cannot see, and they must be finished before the units that depend on them.

Two units in this arc are silent-failure units because they are themselves the gate. Those are the ones worth spending expensive tokens on. Everything else is loud, and spending expensive tokens on it is waste.

Corollary for whoever picks this up: if you find yourself about to make a judgment call that no test would catch, you have crossed a tier boundary. Stop and escalate (§6). That is not a failure — it is the mechanism working.


2. Ownership at a glance

Unit Phase Owner Failure mode
Negative-curriculum epistemology ADR A Opus silent
Volume-honesty invariant (tests first) B Opus silent
R7 assertion — what to assert E1 Opus silent
Articulation feasibility study §6 of plan Opus silent
assert_corpus_sound rename E2 Sonnet #119 loud
Query-scoped premise compilation (R5R7) new, gates C/D/F Sonnet loud
Practice producer implementation C Sonnet loud
core proposal-queue reseal verb D Sonnet loud
R7 code move + suite registration E1 Sonnet loud
Rust parity measurement plan §6 Sonnet loud
ADR status vocabulary sweep E3 Haiku loud
Content authorship (≈219 relations) F Shay
Capability-index scoring / stratification / hash_surface rulings plan §5 Shay
curriculum_serving_enabled flip G Shay

3. Opus units — expectations

A. Negative-curriculum epistemology (ADR) — do this first

Why Opus. This is the highest-risk unit in the arc. If the epistemology is wrong, the corpus encodes falsehoods and wrong=0 still passes — because the lane's gold derives from the same corpus, and evals/curriculum_serve/oracle.py is independent in code but reads the same rows. Corpus-level epistemology errors are structurally unfalsifiable by every gate this repo has. Nothing downstream notices; the whole arena inherits the error and certifies it at volume.

Binding constraints (from the plan, not re-litigable):

  • Absent edge ⇒ UNKNOWN, never "no". This is the open-world reading and it is the reason curriculum serving is honest at all.
  • Explicit negative edge ⇒ REFUTED. These are different states and the corpus must distinguish them.
  • Polarity is a row-level field or intent value, never a new operator_family. Bands key on (domain, operator_family) (chat/curriculum_surface.py:184), so a modal_negative family creates a new band at n=0 rather than adding refuted volume to curriculum_physics_modal.
  • The oracle must learn polarity independently of the serving path. Its code-level independence is the entire evidentiary value; do not let the two share a polarity helper.

Done when: ADR is Accepted, the row schema shape is fixed, and the UNKNOWN-vs-REFUTED boundary is stated precisely enough that Phase B can enumerate against it. No code.

B. Volume-honesty invariant (tests, written before any producer)

Why Opus. This unit is the anti-gaming gate, so there is nothing above it to catch a mistake. S6 already proved the exposure: conservative_floor cannot distinguish "657 independent facts" from "16 facts asked 41 times each" (docs/research/curriculum-volume-quantification-2026-07-24.md §1). An unbounded generator therefore clears θ=0.99 at n≥657 without earning anything, and the resulting license reads as identical to a real one. Getting the bounding rule wrong silently invalidates all of Phase F's authoring effort.

Expectations:

  • Tests come first and must fail for the right reason before any producer code exists. The test file is the specification of the invariant; that is why it is the deliverable and not a by-product.
  • At most one committed case per distinct (relation, direction, polarity, question-shape) tuple. No paraphrase padding.
  • Assert the bound structurally: N ratified chains cannot yield more than a bounded multiple of N committed cases.
  • Mutation-check it. Pad the generator ⇒ red. A pin that cannot fail is worthless; this is the same discipline tests/test_adr_status_governance.py was held to.

Done when: tests exist, fail correctly, fail under mutation, and the module interface they imply is unambiguous enough that Phase C needs no design decisions.

E1 (Opus half). The R7 assertion

Why Opus. The code move is trivial and specified — docs/specs/runtime_contracts.md:448-456 already says "defer the single sink emission to the pipeline serve boundary." What is not specified is what to assert, and the default sink is None, so nothing exercises this path. A wrong move is therefore silent. This is also the exact seam that regressed silently for two days in the #96 register-axis incident, so it gets the careful tier by history as well as by reasoning.

Done when: a written assertion exists — the sink receives the post-override event, and its trace_hash matches a pipeline replay — precise enough for Sonnet to implement against.


4. Sonnet units — expectations

Every unit here has a loud failure mode. You are expected to iterate against the gate until it is green, not to re-derive design. If a unit seems to require a design decision, that is the signal in §6.

Order: E2 → R5R7 → C → D → E1(code) → parity. E2 first because Phase C adds a third assert_corpus_sound caller and renaming after that is strictly more work. R5R7 second because nothing downstream can earn anything until it lands (§0).

Both landed (#119, #120), so this ordering constraint is spent. The names are now assert_practice_gold_sound (no args, practice gold) and assert_lane_cases_sound(domain, cases) (lane contract); Phase C added its caller after the rename, as intended. Remaining units are order-independent except for the stacking recorded in §0b.

R5R7. Query-scoped premise compilation — do this before C and D

teaching/curriculum_premises.py::compile_premises emits every chain in the family, and read_verb_argument refuses past MAX_PREMISE_SENTENCES = 16. So a 17-chain family answers nothing. Three rules, all specified in ADR-0264 §2:

  • R5 — compile a scope that is a superset of the query-atom rows and a subset of the family. Default: term incidence (rows whose subject or object is one of the query's two terms). If that would still exceed the cap, narrow to the query-atom rows. Never truncate arbitrarily, and never refuse for size.
  • R6 — an empty scope is UNKNOWN; only an empty family is the empty_curriculum refusal. Get this wrong and 8,463 of 8,520 questions flip from unknown to declined.
  • R7 (curriculum)premise_count keeps reporting family size, because it is user-visible. Carry the compiled-scope size as a separate field. (Unrelated to audit-ledger R7 in E1 — same numeral, different register.)

Why this is loud, and why it is not the ADR-0261 §5.1 premise-dropping sin. Narrowing is verdict-identical, not merely sound: each compiled premise mints one independent propositional atom, so a scope containing the query atom decides the query exactly as the full family does — wherever the full family was readable. Over the cap it is verdict-improving: full-family compilation refuses compiled_premises_unreadable there, so the narrowed scope answers questions the unscoped compiler could not. That is the unblock, not a deviation from it. Verified over 8,520 routable questions with zero mismatches (docs/research/curriculum-premise-scope-2026-07-25.md probe 3). Gold is preserved: physics stays entailed 14 / unknown 12 / declined 6.

Gate: the curriculum_serve lane at wrong=0, and the 20 tests in tests/test_curriculum_serve.py. If any verdict moves, stop — the equivalence argument has a hole and that is escalation §6.3.

E2. assert_corpus_sound rename — DONE, #119

Renamed to assert_practice_gold_sound / assert_lane_cases_sound. Two functions, one name: evals/deduction_serve/practice/gold.py:1163 (no args, practice corpus) and evals/curriculum_serve/runner.py:71 (domain, cases, lane contract). Not a missing contract — a name collision, flagged in BRIEF-CLOSE-assessment-verification-2026-07-25.md §9.5. Gate: imports break.

C. Practice producer

evals/curriculum_serve/practice/{__init__,generator,runner}.py. Mirror evals/deduction_serve/practice/ topology exactly: producer in evals/, sealed artifact next to its reader in chat/data/.

Mirror the topology, NOT the volume strategy. The deduction producer sets a flat CASES_PER_BAND = 720 and fills the quota by cycling a template × vocabulary space that is as small as 28 distinct instances — which is the exposure Phase B measured. Follow the estimation producer instead (evals/determination_estimation): 660 distinct cases per class, zero repeats, chosen just above the 657 a perfect record needs. Emit at most one case per distinct query atom (ADR-0264 R9) and let the count fall where the corpus puts it. Then register the producer in AUDIT_SOURCES (tests/test_volume_honesty.py) — the declared None there fails as soon as chat/data/curriculum_serve_ledger.json exists, which is how this unit is stopped from repeating the pattern.

Reuse, do not rewrite — this is why the unit is small:

  • evals/curriculum_serve/oracle.py — already an independent decision procedure; wrap as the gold tether.
  • core/learning_arena/engine.py::run_practice — the ADR-0199 pure fold.
  • core/ratified_ledger.py::{seal_artifact, tally_dict, write_sealed_ledger} — extracted by ADR-0263 for exactly this third instance.
  • core/reliability_gate::{Action, Ceilings, ClassTally, license_for}.
  • chat/curriculum_surface.py — the solver under test.

Determinism is required: no clock, no RNG, byte-identical sealed output across runs, so the artifact is safe to commit and SHA-verify on load. Read evals/deduction_serve/practice/runner.py docstring for the pattern.

Gate: Phase B tests; wrong=0 on the physics lane.

D. core proposal-queue reseal

teaching/ratification.py, core/cli_proposal_queue.py. Explicit verb — never auto-reseal on ratify (doctrine forbids ratification automation, and per-row reseal would churn the ledger's content_sha256). ratify_chain records the enqueue; a human runs the seal. Follow the existing status|list|show|review|ratify verb patterns.

Gate: CLI tests; the verb produces a SHA-sealed chat/data/curriculum_serve_ledger.json, and curriculum_serve_license() reads a real but still-unearned ledger instead of an absent one.

E1 (code half). R7

Move the emission per Opus's written assertion. Do not improvise the assertion.

Rust parity measurement

PYO3_PYTHON=/opt/homebrew/bin/python3.12, cargo test in core-rs/, then rerun the hash-pinned lanes under CORE_BACKEND=rust. cargo 1.97.1 and the registry cache are present locally; the cloud session was blocked only by sandbox network policy. Report the measurement; do not act on it. A parity failure is a silent-class finding about determinism — escalate rather than "fix."

Suite registration (all units)

Anything you add goes into a curated suite in core/cli_test.py. Two files from PR #113 landed in no suite and ran only under full, which gates nothing — that was the fifth instance of this pattern in three arcs. Do not make it six.


5. Haiku unit

E3. ADR status vocabulary normalization

27 unparseable status lines plus draft/ratified/active/accepted. variants across 312 ADR files (inventory: docs/research/adr-status-governance-2026-07-25.md §6). Normalize to the intended lifecycle, then widen tests/test_adr_status_governance.py from its deliberately narrow scope to a closed vocabulary.

Bulk mechanical with an objective acceptance criterion: the widened pin goes green. Do not change any ADR's decision — only its status token. If a file's intended status is genuinely ambiguous, list it and stop; do not guess.


6. Escalation — when a cheaper tier must stop

Stop and hand back up when any of these is true. Escalating is the mechanism working, not a failure; guessing past one of these is the only real error available here.

  1. You are about to make a call no test would catch. By definition you have crossed into a silent-failure unit.
  2. A gate is green but you do not believe it. That is the signature of a reference that has become wrong — exactly the ADR-0256 shape, where the record asserted "not yet decided" about a live decision and nothing failed.
  3. The measurement contradicts the plan. Say so with the numbers. Both prior arcs improved by having their premises falsified in writing; a falsified plan premise is a deliverable, not an interruption.
  4. A parity or determinism check fails. Never "fix" a hash drift. Report it.
  5. Scope exceeds the unit as written. Flag it in the PR/handoff rather than improvising, per the standing Tier-S/O constraint.

7. Standing constraints (non-negotiable, all tiers)

  • wrong=0. No flag flips — those are Shay's ratification only.
  • Forgejo-primary; no gh, no tea. SSH:29222 often times out — push over the authenticated HTTPS remote instead.
  • Branch and worktree per unit; both deleted immediately after merge.
  • Pre-push gate in-worktree on canonical Python 3.12.13 with uv sync --locked: core test --suite smoke -q then --suite deductive -q. Baselines: smoke 555 / deductive 285 at arc start; smoke 569 after Phases A+B (+1 from ADR-0264 joining the derived ADR pin, +13 from tests/test_volume_honesty.py).
  • [Verification]: line in every PR description, naming the interpreter.
  • One PR per coherent unit. Compare-URL PRs; Shay merges — no merge automation, and green PRs sit until explicit authorization.
  • Never scripts/verify_lane_shas.py --update — surgical single-line pin edits only. Three arcs depend on that rule.
  • New tests go into a curated suite (§4).

8. At arc close

Fill in docs/handoff/ARC-CLOSE-TEMPLATE.md → this directory. The Falsified assumptions section is the one that cannot be recovered from the diff — state which parts of the plan of record turned out wrong, including your own claims. Also record whether the articulation trigger fired, which the previous arc left unrecorded.