docs: backfill PROGRESS.md + capability_roadmap.md with Phase 6 + ADR-0027..0110 (audit #2) (#121)

PROGRESS.md was last updated mid-Phase-5 (~ADR-0021). The pack-layer
chain (ADR-0027..0045), forward-graph+surface composer chain
(ADR-0046..0089), and evidence-governed domain layer (ADR-0091..0110)
were all absent.

capability_roadmap.md was last updated 2026-05-17 and predated the
expert-demo promotion contract entirely.

Added to both:

- Phase 6 — Evidence-Governed Domain Layer section covering ADR-0091
  contract chain through ADR-0110 expert-demo arc, including the
  refused-amended-succeeded narrative and the current ledger state
  (math at expert-demo, four others at reasoning-capable).
- Backfill subsections for ADR-0027..0045 (pack-layer chain) and
  ADR-0046..0089 (forward-graph + surface-composer + register +
  anchor-lens + contemplation-loop work) marked retrospective.

capability_roadmap.md also gains a '2026-05-22' status update at the
top alongside the existing 2026-05-17 entry. The full Phase 0-5
exit-criteria rewrite remains queued separately.

No code, no tests, no claims changes.
This commit is contained in:
Shay 2026-05-22 13:21:22 -07:00 committed by GitHub
parent 95e5e93d02
commit 95d3b4a45a
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 156 additions and 2 deletions

View file

@ -702,6 +702,99 @@ Per-surface bit-identity gates landed (2026-05-16):
--- ---
## Phase 6 — Evidence-Governed Domain Layer
**Status:** IN PROGRESS (opened 2026-05-21, first promotion landed 2026-05-22)
**Depends on:** Phase 5 corpus flywheel + pack-layer chain (ADR-0027..0045)
**Roadmap:** see `docs/capability_roadmap.md` Phase 6 entry.
This phase ratifies the distinction between *contract-passing* (`reasoning-capable`) and *demonstrated* (`expert_demo=true`) at the capability ledger surface.
### Contract layer (all accepted)
- [x] **ADR-0091** Domain Pack Contract v1 — nine predicate checks on every ratified pack
- [x] **ADR-0092** Reviewer Registry v1 — schema-validated YAML reviewer roster
- [x] **ADR-0093** Domain Contract v1 implementation — runtime validator + ledger enforcement
- [x] **ADR-0094** Proposal source provenance — discriminated `ProposalSource(kind=...)`
- [x] **ADR-0095** Miner-sourced teaching proposals — `teaching/from_miner.py` + SHA-pinned `miner_loop_closure` lane
- [x] **ADR-0096** Fabrication-control eval lane — phantom / cross-pack / sibling-collapse refusals (SHA-pinned)
- [x] **ADR-0098** Demo composition contract
- [x] **ADR-0099** Public showcase demo (deterministic, byte-equal, <30s)
- [x] **ADR-0104** Curriculum-sourced teaching proposals — `teaching/from_curriculum.py` + SHA-pinned `curriculum_loop_closure` lane
- [x] **ADR-0105** Sealed-holdout encryption via age — dev-mode plaintext fallback preserved
### Reasoning-capable ratifications
- [x] **ADR-0097** `mathematics_logic` reasoning-capable (now superseded by ADR-0110 expert-demo, see below)
- [x] **ADR-0100** `physics` reasoning-capable
- [x] **ADR-0101** `systems_software` reasoning-capable
- [x] **ADR-0102** `hebrew_greek_textual_reasoning` reasoning-capable (first multi-pack ratification: 4 packs)
- [x] **ADR-0103** Hebrew + Koine Greek fluency lane attachment to ADR-0102 packs
### Expert-demo arc (the contract demonstrated end-to-end)
- [x] **ADR-0106** Expert-Demo Promotion Contract — domain-aware, reviewer-signed, replay-deterministic
- [x] **ADR-0107** `mathematics_logic` expert-demo deferred — first promotion attempt **honestly refused** on two named blockers (metric-shape uniformity; `inference_closure` 40% pass)
- [x] **ADR-0108** Proposed-ADR sequencing — meta-decision pinning the post-ADR-0105 frontier
- [x] **ADR-0109** Lane-shape-aware threshold amendment — 8 lane ids → 5 shapes (`cognition_shape`, `accuracy_shape`, `inference_shape`, `refusal_shape`, `symbolic_logic_shape`); unknown lanes fail-closed; cognition-shape thresholds preserved bit-identical
- [x] PR #117 — fix intent-classifier regression that had broken `inference_closure` (`_CORRECTION_CUE_PREFIX_RE` guard)
- [x] **ADR-0110** `mathematics_logic` expert-demo promoted — **first domain at `expert_demo=true`** in project history; signed claim digest reproduces byte-for-byte from on-disk lane results
### Contract demonstration narrative
The ADR-0106 contract refused once (0107), amended once cleanly (0109), and succeeded once (0110). External readers can now distinguish the two ceilings (`reasoning-capable` vs `expert-demo`) by inspecting the ledger.
### Current ledger state (per `core capability ledger`)
| Domain | Status |
|---|---|
| `mathematics_logic` | **`expert-demo`** ✓ |
| `physics` | `reasoning-capable` |
| `systems_software` | `reasoning-capable` |
| `hebrew_greek_textual_reasoning` | `reasoning-capable` |
| `philosophy_theology` | `reasoning-capable` |
### Open within Phase 6
- [ ] Second expert-demo promotion (any of `physics` / `systems_software` / `hebrew_greek_textual_reasoning` — all now eligible under ADR-0109 shape rules)
- [ ] Multi-reviewer threshold signing (open candidate frontier item from ADR-0105)
### Pack-layer chain — ADR-0027 through ADR-0045 (backfill)
The pack-layer architecture that Phase 6 builds on was established as a sequence between Phases 4 and 5 work but never registered here. Captured retroactively:
- [x] **ADR-0027** Identity packs — `IdentityManifold` loaded from swappable, content-addressed pack
- [x] **ADR-0028** Identity surface wiring — packs carry `surface_preferences` consumed by assembler
- [x] **ADR-0029** Safety packs — sibling to identity, never-swappable, fail-closed at startup
- [x] **ADR-0033** Ethics packs — third pack layer; swappable like identity, propositional like safety
- [x] **ADR-0035** Turn-loop verdicts — `SafetyCheck` + `EthicsCheck` auto-invoked
- [x] **ADR-0036** Safety refusal policy — typed runtime refusal
- [x] **ADR-0037** Ethics refusal opt-in
- [x] **ADR-0038** Hedge injection — soft remediation
- [x] **ADR-0039** Audit completeness — `TurnVerdicts` bundle + stub-path emission
- [x] **ADR-0040** Telemetry sink — structured JSONL turn-event sink
- [x] **ADR-0041** CLI verdicts + fan-out
- [x] **ADR-0042** Audit-tour demo — `core demo audit-tour`
- [x] **ADR-0043** Pack measurements — identity-divergence + refusal-calibration runners
- [x] **ADR-0044** Medical-ethics pack — worked-example domain ethics pack
- [x] **ADR-0045** Long-context comparison evidence
### Forward-graph + surface-composer chain — ADR-0046 through ADR-0089 (backfill)
Also retroactive (the bulk happened between 2026-05-18 and 2026-05-20):
- [x] **ADR-0046/0047** Forward graph constraint — PropositionGraph → AdmissibilityRegion *before* generate runs
- [x] **ADR-0048..0066** Pack-grounded surface composers for every intent shape (DEFINITION, RECALL, COMPARISON, PROCEDURE, CORRECTION, NARRATIVE, EXAMPLE)
- [x] **ADR-0063/0064** Cross-pack resolver + cross-pack teaching corpora
- [x] **ADR-0068..0072** Register substrate (terse / convivial / formal); orthogonal axis — register holds `trace_hash` CONSTANT
- [x] **ADR-0073** Anchor lens substrate — substantive variation axis; **opposite** invariant from register (lens moves `trace_hash` DISTINCT)
- [x] **ADR-0078** Composer/graph atom equivalence telemetry
- [x] **ADR-0080** Contemplation Loop Phase 1 — read-only frontier-compare miner, `SPECULATIVE`-only findings (landed 2026-05-22)
- [x] **ADR-0083** Transitive chain surface — bounded multi-hop teaching-grounded surface
- [x] **ADR-0089** Discourse planner + compound-intent dispatch
---
## Open Scope Decisions ## Open Scope Decisions
| Decision | Status | Deadline | | Decision | Status | Deadline |

View file

@ -2,7 +2,7 @@
**Status:** Draft, derived from `docs/sessions/SESSION-2026-05-15-capability-gates.md` **Status:** Draft, derived from `docs/sessions/SESSION-2026-05-15-capability-gates.md`
**Owner:** Joshua Shay **Owner:** Joshua Shay
**Last updated:** 2026-05-17 **Last updated:** 2026-05-22 (Phase 6 added; status update appended below)
## Status update — 2026-05-17 ## Status update — 2026-05-17
@ -35,7 +35,22 @@ Closed gates (vs the Phase 1 "Foundational Triple"):
A full rewrite of the phase exit-criteria against current state is queued for A full rewrite of the phase exit-criteria against current state is queued for
the next planning pass. the next planning pass.
This document walks CORE from its present state through the gating framework defined in the 2026-05-15 session. It is organized into six phases. Each phase has entry criteria, work items, exit criteria, and a benchmark discipline contract. ## Status update — 2026-05-22
Phase 5 (Curriculum Era) work landed in volume during the 2026-05-17 → 2026-05-22 window. The roadmap gains a new Part-II phase (**Phase 6 — Evidence-Governed Domain Layer**, see below) that did not exist in the 2026-05-15 draft because the substrate it ratifies (multi-domain packs, signed reviewer registry, expert-demo gate) did not yet exist either. The phase is documented here as accepted and partially landed, not proposed.
Major chains accepted since 2026-05-17:
* **Pack-layer chain — ADR-0027 through ADR-0045.** Identity packs (0027/0028), safety packs (0029), ethics packs (0033/0036/0037), reviewed verdicts (0035), audit completeness (0039), telemetry sink (0040), CLI verdicts + fan-out (0041), audit-tour demo (0042), pack measurements (0043), medical-ethics worked example (0044), long-context comparison evidence (0045).
* **Forward graph constraint → pack-grounded surface — ADR-0046 through ADR-0070.** PropositionGraph becomes AdmissibilityRegion before generate (0046/0047); pack-grounded surface composers for DEFINITION/RECALL/COMPARISON/PROCEDURE/CORRECTION/NARRATIVE/EXAMPLE intents (00480066); cross-pack resolver and teaching-corpus chain (0063/0064); register terse v1 (0070); seeded variation (0071); register telemetry + tour (0072).
* **Anchor lens substrate — ADR-0073 + sub-ADRs.** Substantive variation axis sibling to register, with the opposite invariant (lens moves trace_hash DISTINCT; register holds it CONSTANT). Both orthogonal-axis demos CI-pinned simultaneously.
* **Transitive chain surface and definitional groundwork — ADR-00780089.** Composer/graph atom equivalence telemetry (0078); transitive chain surface (0083); compound-intent dispatch + discourse planner (0089).
* **Contemplation Loop Phase 1 — ADR-0080 (accepted 2026-05-22).** Read-only frontier-compare miner emits `SPECULATIVE` findings only; routes through `teaching/review.py`; no pack mutation. The first system-emitted gap-finding surface.
* **Evidence-governed domain layer — ADR-0091 through ADR-0110.** See Phase 6 below.
A full Phase 05 exit-criteria rewrite against this expanded state remains queued; that pass is independent of the Phase 6 documentation.
This document walks CORE from its present state through the gating framework defined in the 2026-05-15 session. It is organized into six phases (now seven with Phase 6 below). Each phase has entry criteria, work items, exit criteria, and a benchmark discipline contract.
The benchmark discipline is the spine of the plan. Without it, the phases become aspirational. With it, "are we there yet" becomes a CLI question. The benchmark discipline is the spine of the plan. Without it, the phases become aspirational. With it, "are we there yet" becomes a CLI question.
@ -323,6 +338,52 @@ The phase has no single exit criterion. Instead, each domain becomes its own sub
**This phase has no estimated duration.** It is the phase the project lives in after the engineering era ends. Frontier-LLM parity on breadth happens *inside* this phase if it happens at all — likely measured in years, not weeks, and at whatever sample efficiency Phase 4 demonstrated. **This phase has no estimated duration.** It is the phase the project lives in after the engineering era ends. Frontier-LLM parity on breadth happens *inside* this phase if it happens at all — likely measured in years, not weeks, and at whatever sample efficiency Phase 4 demonstrated.
### Phase 6 — Evidence-Governed Domain Layer
**Entry criteria:** Phase 5 corpus flywheel operational (curriculum + miner sourcing actively producing reviewed proposals); pack-layer chain (ADR-0027..0045) closed; forward-graph-constraint and surface-composer chains (ADR-0046..0089) shipping.
**Goal.** Distinguish *contract-passing* from *demonstrated* at the ledger surface. A pack that satisfies the nine ADR-0091 predicates earns a `reasoning-capable` ledger row; promotion to `expert_demo=true` requires a reviewer-signed evidence-bundle digest that reproduces byte-for-byte from on-disk lane results.
**Substrate (all accepted as of 2026-05-22):**
| Layer | What it ratifies | ADR |
|---|---|---|
| Domain Pack Contract v1 | Nine predicate checks on every ratified pack. | 0091 |
| Reviewer Registry v1 | Schema-validated reviewer roster; primary vs. domain-scoped. | 0092 |
| Domain Contract v1 implementation | Validator + ledger enforcement at runtime. | 0093 |
| Proposal source provenance | Discriminated `ProposalSource(kind=...)` for every teaching proposal. | 0094 |
| Miner-sourced proposals | `teaching/from_miner.py`; SHA-pinned `miner_loop_closure` lane. | 0095 |
| Fabrication-control eval lane | Phantom / cross-pack / sibling-collapse refusals must clean. | 0096 |
| Demo composition contract | Demos compose from shipped modules; no parallel mechanism. | 0098 |
| Public showcase demo | Deterministic, byte-equal, <30s. | 0099 |
| Reasoning-capable ratifications | `mathematics_logic` (0097), `physics` (0100), `systems_software` (0101), `hebrew_greek_textual_reasoning` (0102). |
| Fluency lane attachment | Hebrew + Koine Greek lanes attached to ADR-0102 packs with dev/public/holdout. | 0103 |
| Curriculum-sourced proposals | `teaching/from_curriculum.py`; SHA-pinned `curriculum_loop_closure` lane. | 0104 |
| Sealed-holdout encryption | age-based; dev-mode plaintext fallback preserved. | 0105 |
| **Expert-demo promotion contract** | **Domain-aware, reviewer-signed, replay-deterministic gate.** | **0106** |
| Lane-shape registry | 8 lane ids dispatch to 5 shapes; unknown lanes fail-closed. | 0109 (amends 0106) |
**Worked promotion narrative.** The contract has been demonstrated end-to-end:
1. **ADR-0107** — first promotion attempt (`mathematics_logic`) honestly refused by the contract on two named blockers (metric-shape uniformity assumption; `inference_closure` substantively failing at 40% pass).
2. **ADR-0109** — threshold rules amended with explicit lane-shape registry; cognition-shape thresholds preserved bit-identical; four new shapes added (`accuracy_shape`, `inference_shape`, `refusal_shape`, `symbolic_logic_shape`); unknown lanes fail-closed.
3. PR #117 fixed the intent-classifier regression that had broken `inference_closure`.
4. **ADR-0110**`mathematics_logic` promoted to `expert_demo=true` under the amended contract. Signed claim digest reproduces from on-disk lane results; first domain at expert-demo.
**Exit criteria (cumulative; each can land independently of the others):**
- ☑ Contract definition (ADR-0091..0093) and reviewer trust root (ADR-0092) accepted.
- ☑ Negative-control fabrication lane (ADR-0096) ratified and SHA-pinned.
- ☑ Four reasoning-capable domain ratifications (ADR-0097/0100/0101/0102).
- ☑ Expert-demo promotion contract accepted (ADR-0106 + ADR-0109 amendment).
- ☑ First worked promotion lands (ADR-0110 — `mathematics_logic`).
- ☐ Second worked promotion lands (next domain: TBD; `physics` / `systems_software` / `hebrew_greek_textual_reasoning` all eligible under the now-amended contract).
- ☐ Multi-reviewer signing (currently single-recipient; the open candidate frontier item from ADR-0105).
**Why this is its own phase.** Phases 15 measure capability *internally*. Phase 6 is the first layer that measures *what the system has actually demonstrated to an external reader* and forces the ledger to distinguish the two. Every prior phase makes claims about the substrate; Phase 6 makes the substrate's claims auditable.
**This phase has no estimated duration.** Each new domain promotion is a discrete unit of work; the contract is the durable artifact.
--- ---
## Part III — Cross-Cutting Considerations ## Part III — Cross-Cutting Considerations