Codex — Agentic Execution Plan

Drafted 2026-05-17. Operational sibling to FASTTRACK_COVERAGE_ROADMAP.md (which says what to cover next) and FASTTRACK_EQUIVALENCE_PLAN.md (which says what equivalence means). This document says how to agentically execute the corpus build at maximum throughput without compromising any quality gate that the existing orchestration protocol defines.

This is not a new model. It composes the existing roles (audit / gap- analysis / production-plan / production / integration / verification / weaving / continuity-verification / quality-sampling / flow / watchdog — see docs/specs/ORCHESTRATION_PROTOCOL.md) into a sustainable pipeline, fills the seven policy gaps the existing docs left open, and gives the operator a concrete dispatch protocol.

Read alongside:

docs/specs/ORCHESTRATION_PROTOCOL.md — agent role definitions, file-write discipline, phase model P1–P5.
docs/specs/QUALITY_RUBRIC.md — the 27/27 validator and tier-content gates.
docs/specs/UNIT_SPEC.md — three-tier unit format.
docs/specs/CONTINUITY_SCAFFOLD.md — Pass W weaving and Pass V continuity verification.
docs/specs/FASTTRACK_FLOW_SCAFFOLD.md — flow phases PA–PE.
docs/plans/FASTTRACK_EQUIVALENCE_PLAN.md — seven-layer equivalence model.
docs/plans/FASTTRACK_COVERAGE_ROADMAP.md — wave-by-wave book sequencing.

§1 Mission

Push the Codex corpus toward Fast Track equivalence-completeness on all 93 books at the highest sustainable throughput consistent with zero relaxation of the existing quality gates. Specifically:

27/27 validator must pass on every shipped unit (no exceptions).
All 7 equivalence layers must verify per book (no shortcuts to "topic overlap only").
Pass W and Pass V must run on every batch (no skipping the seam-stitching).
Quality-sampling 5% spot-check runs every batch (no silent drift).
Every substitute text and every BUY-only fallback must be logged and justified against the new rules in §6 below.

If throughput and quality conflict, throughput yields. The throughput target is high, not maximum: maximum throughput would require relaxing gates, which is forbidden.

§2 Concurrency budget (the throughput envelope)

The existing protocol (ORCHESTRATION_PROTOCOL.md §5.1) allows unlimited parallel production agents because they write to disjoint unit paths. The actual bottleneck is the integration agent (single-threaded) and the operator's review bandwidth. The concurrency budget below is calibrated so neither becomes a queue:

Role / pass	Max parallel	Sustained pace	Bottleneck check
Audit (P1)	12	6/day	Operator review of audit §1 output
Gap-analysis (P2)	6	4/day	Reads shipped corpus — context-heavy
Production-plan (P3)	4	4/day	Operator sign-off on punch-list before P4
Production (P4)	8 per batch, 2 batches/day	16 units/day	Integration drain rate
Integration	1 (mandatory single)	16 units/day	Single-threaded; budget 10 min/unit
Verification (P5)	2	1 book/day	Per-book equivalence audit
Pass W (weaving)	1 per batch	1 batch/day	Runs after each P4 batch
Pass V (continuity verify)	1 weekly	weekly	Curriculum-wide; serialized
Quality-sampling	1 per batch	5% of P4 output	Concurrent, read-only
Flow (PA–PE)	4	2 flows/day	Independent of unit pipeline
Watchdog	1 continuous	hourly tick	Stuck-book alerts

Sustained daily throughput at these limits: ~6 audits (P1), ~16 shipped units (P4), ~1 verified book (P5), ~2 flow updates. Weekly: ~30 audits, ~80 units, ~5 books verified.

Why not more. Two reasons. (1) Integration is single-threaded by design — the only writer to shared files (CONCEPT_CATALOG.md, deps.json, connections.json). Pushing P4 above 16 units/day fills the integration queue. (2) Pass W and Pass V are serial quality gates. Bypassing them to "catch up" is the exact compromise this plan forbids.

§3 Phase hand-off protocol (auto vs. manual)

ORCHESTRATION_PROTOCOL.md §3.3 names the phases but does not specify hand-off triggers. Fill that gap explicitly:

Hand-off	Trigger	Auto / manual	Operator action
P1 → P2	Audit §1 written + ≥3 sources cited	Auto dispatch P2 agent	Spot-check audit §1 within 1 day
P2 → P3	§2 coverage scorecard populated for all 7 layers	Manual approval	Operator reviews gap report; can reject and re-run
P3 → P4	§3 punch-list complete + dependencies declared	Manual approval	Operator confirms unit decomposition; rejects if punch-list compresses tiers
P4 → Integration	Producer commits unit at expected path	Auto queue	Integration agent picks up FIFO
Integration → 27/27	Integration script writes frontmatter + edges	Auto validator runs	Failure routes back to producer with defect list
27/27 pass → ship	Validator returns all-pass	Auto flip status `draft` → `approved` → `shipped`	None unless quality-sampling flags
Batch complete → Pass W	All P4 units in batch shipped	Auto dispatch Pass W	Operator reads `weaving_report.md` post-hoc
Pass W → Pass V	Weekly cron OR ≥3 batches since last Pass V	Auto dispatch Pass V	Operator reads continuity report; addresses deficits
Per-book Pass W+V both clear	Triggers P5	Auto dispatch P5	Operator signs off `verified` status
P5 verify → `equivalence-covered`	All 7 layers ≥ threshold	Manual approval	Operator signs off in `manifests/campaign.json`

Operator-only gates (3): P2→P3, P3→P4, and P5→equivalence-covered. Everything else autopilot. The three manual gates are where quality is most defensible: gap fidelity, punch-list scope, and book-level closure.

§4 Daily / weekly operator loop

The operator (Tyler or a designated driver agent) runs this loop:

Morning (~30 min):

Read overnight watchdog_report.md and quality_sampling_report.md.
Resolve any stuck-book alerts (>4h owner lock; see ORCHESTRATION_PROTOCOL.md §5.3).
Review yesterday's weaving_report.md. Flag any forced-anchor-phrase insertions that look mechanical (drift signal).

Mid-day (~60 min): 4. Read any P2 gap reports pending approval. Approve or send back. 5. Read any P3 punch-lists pending approval. Approve or send back. 6. Dispatch next day's audit batch (queue up to 12 P1 agents for tomorrow).

End of day (~30 min): 7. Read campaign.json rollup: how many audits closed, units shipped, books verified. 8. Sign off any books reaching verified state. 9. Update MEMORY.md if a feedback pattern emerged.

Weekly (~2 hours): 10. Read full continuity_report.md from Pass V. Address any metric below threshold by queueing a targeted Pass-W run. 11. Re-read FASTTRACK_COVERAGE_ROADMAP.md and adjust Wave priorities based on actual throughput vs. plan. 12. Sample 10 random shipped units for prose-density read (master tier only); confirm Lawson-Michelsohn / Hörmander calibre maintained. 13. Commit the _archive of the week's reports.

Total operator time: ~2 hours/day + 2 hours/week = ~12 hours/week. This is the human-in-the-loop budget. Anything beyond this means the pipeline is over-dispatching and must be throttled.

§5 Wave 5 execution schedule (next 14 days)

Concretely sequenced against the throughput envelope in §2. Targets the Wave 5 audit list from FASTTRACK_COVERAGE_ROADMAP.md §3.

Day 1 (today, kickoff):

Dispatch P1-audit batch: 2.03 Chatterjee, 2.02 Woit, 3.38 May, 3.40 May Simplicial, 3.42 Ravenel. All FREE; PDFs already in reference/fasttrack-texts/. Five parallel agents.
Operator: read the existing Brown 1.05 punch-list and dispatch a P4-production batch of the 7 priority-1 units (one batch of 4 agents on units 02.01.06 is shipped — skip; 02.01.07 compact-open, 02.01.08 cofibration, 02.01.09 fibration, 03.12.00 fundamental group).

Day 2:

Receive Day 1 audits. Operator reviews each audit §1 briefly (15 min × 5 = 75 min).
Auto-dispatch P2 gap-analysis for all 5. (6 max — comfortable.)
Receive P4 batch from Day 1 + Pass W on the batch. Quality-sampling pulls 1 unit.
Dispatch next P4 batch: 03.12.0a fundamental groupoid, 03.12.0b Seifert-van Kampen, plus 2 production agents on Apostol Vol I priority-1 (any of its 18 punch-list items not yet started).

Day 3:

Receive Day 2 P2 reports. Operator manual-approves each (5 min × 5 = 25 min).
Auto-dispatch P3 production-plan agents. (Note: these update the per-book plan stub's §3 punch-list — they are not the punch-list itself; the punch-list was already in the Wave 5 audit's §1.)
Dispatch P1-audit batch: 2.01 Landau QM, 2.10 Landau Stat Phys 1, 0.4 Arnold ODE, 1.02 Landau Mechanics, 1.03 Landau Classical Fields. Five more parallel agents. (All BUY — see §6 fallback rule.)

Days 4–7:

P3 punch-lists from Day 1's audits land; operator approves; P4 production begins on top-priority items from each new book.
Continued P4 batches at 2/day × 8 units = 16/day target.
Pass W after each batch; quality-sampling on each batch.
Day 7 weekly checkpoint: Pass V continuity run.

Days 8–14 (Wave 5 closeout):

Wave 5 P1 audits complete (10 books). All 10 should be at "plan-stub with priority-1 punch-list" state by Day 10.
P4 production active on at least 3 of those 10 books in parallel (different sections → no integration collision).
Wave 5 first books reach P5 verification by Day 12 (Chatterjee is short — realistic).
Day 14: Wave 5 complete. audits_done rises from 13 → 23. books_unaudited 48 → 38.

Wave 5 success criterion: 10 audits closed, ≥80 units shipped, ≥1 book at verified state, zero quality-gate failures, zero stuck-book alerts unresolved.

If by Day 7 the pace is materially below this, throttle audit dispatch and route capacity to integration / weaving until the queue drains. Speed is never an excuse to compress quality.

§6 The seven policy gaps — explicit fills

The Explore audit (2026-05-17) found seven operational gaps in the existing specs. This plan fills each:

6.1 Substitute-text policy

Rule. A plan stub may anchor against a substitute text only if:

The FT-listed canonical text is genuinely inaccessible (out of print, no library availability, no acceptable scan) and
The substitute covers ≥80% of the canonical text's theorem inventory (verified by a side-by-side TOC diff in the plan stub's §1) and
The substitute's notation crosswalk is recorded so a reader can move between texts.

The four existing substitute plans (Hatcher for May 3.38; Cannas da Silva for parts of Sternberg 1.10 and Woodhouse 3.09; Forster for Ahlfors-Sario 1.06; Arnold Mathematical Methods for Arnold ODE 0.4 and Sternberg 1.10) do not satisfy this rule and must be supplemented with canonical-text plan stubs as part of Waves 5–10.

Naming. Substitute plans are renamed <author>-<topic>--substitute-for-<ft-entry>.md so the file name itself flags the substitution.

6.2 Continuity metric thresholds (Pass V)

CONTINUITY_SCAFFOLD.md defines six metrics but leaves several thresholds informal. Lock them at:

backward-reference density ≥ 8 / 1000 words (Master+Intermediate avg)
forward-promise density ≥ 3 / 1000 words (Master only)
lateral-connections ≥ 3 / unit (Master only)
synthesis-claims: ≥ 2 / Intermediate unit, ≥ 4 / Master unit
registered anchor-phrase reuse ≥ 95% of cross-references
throughline-coverage = 100% (no chapter starts without an inbound connection registered)

A unit failing any metric in Pass V does not lose shipped status, but it enters the Pass-W rework queue.

6.3 Per-unit token / cost budget

Defer to a thinking-budget posture, not a hard ceiling. Hard ceilings encourage cutting corners. Instead:

Beginner sections target ~600–900 words. Hard cap: 1500.
Intermediate sections target ~1200–2000 words.
Master sections target ~1500–3000 words.
Each unit's production agent is given a thinking budget of "as much as needed; the rubric is the gate." Cost ledger (per ORCHESTRATION_PROTOCOL.md §9) records actuals.

If a unit exceeds 4× the upper-target word count, it is split into multiple units rather than published as a monolith.

6.4 Review bottleneck resolution

REVIEWER_PLAN.md flags expert review as the scaling constraint. This plan resolves it as follows:

27/27 validator is the primary gate for production units. Human review beyond that is optional, sampled.
Quality-sampling agent (5%/batch) is the routine human-substitute.
Operator manual review is reserved for: every P2 gap report, every P3 punch-list, every P5 equivalence verification, and any unit the quality-sampler flagged in the prior 7 days.
Expert (human, external) review is NOT in the critical path. It runs as a separate audit lane, in arrears, on a sample of verified books — not per unit. The expert's job is to catch systemic drift the validator misses, not to gate ship.

6.5 Flow-update triggers

When a unit changes, its parent flows must be re-validated. Trigger:

Any edit to a shipped unit's body or frontmatter sets flow_dirty: true in manifests/connections.json for every flow containing that unit.
Flow-PD agent (verification, see FASTTRACK_FLOW_SCAFFOLD.md §5) wakes on the next flow batch and re-runs verification on flow_dirty: true flows.
Pass W on the unit's batch must clear before flow_dirty is cleared.

6.6 BUY-only book fallback

For books unobtainable via the project's existing channels (purchase, library, ILL, institutional access), an audit agent may produce a reduced audit with the following constraints:

§1.1 (theorem inventory) is built from publicly available TOC, preface, sample chapter, and citation network (Google Scholar references to the text).
§1.2 (exercise inventory) is omitted; replaced with a note "exercises inventoried after sourcing — see NEED_TO_SOURCE.md".
§1.3 (worked examples) populated from any sample chapter or publisher preview.
§1.7 (distinctive perspective) populated from peer reviews, bibliography essays, and citation context.
The audit is marked audit_completeness: reduced in manifests/production/plan.json.
A reduced audit can still trigger P2 / P3 / P4 — the punch-list proceeds against the partial inventory. Re-audit upgrade is queued in NEED_TO_SOURCE.md.

Hard rule: P5 verification on a reduced-audit book cannot mark the book equivalence-covered. It can only mark it equivalence-partial. The book must be re-audited from a full source before final closure.

6.7 Master-tier prose calibre enforcement

The existing rubric prohibits sloppy phrasings but does not name a positive standard. Codify: master-tier prose must be at the calibre of Lawson-Michelsohn Spin Geometry and Hörmander The Analysis of Linear Partial Differential Operators. Specifically:

Declarative sentences. No hedged futures ("we will see", "the right thing to do is").
No metacommentary ("Note that…", "It is interesting that…").
No real-time error correction in prose ("Actually, what I meant was…").
No valedictory closing ("We have now completed our discussion of…").
No rhetorical italics for emphasis.
Honest structural signposting: a paragraph that says "We prove X" must prove X immediately after.

Quality-sampling agent scores master-tier prose against this standard; units scoring < 4/5 enter the Pass-W rework queue.

§7 Campaign ledger format (`manifests/campaign.json`)

The existing campaign.json is sparse. Extend it (without breaking existing readers) to carry the per-book state machine this plan needs:

{
  "books": {
    "1.05a": {
      "title": "Nonabelian Algebraic Topology",
      "author": "Brown, Higgins, Sivera",
      "audit_completeness": "full",
      "phase": "P1-done",
      "owner": null,
      "owner_locked_at": null,
      "audit_path": "plans/fasttrack/brown-higgins-sivera-nonabelian-algebraic-topology.md",
      "punch_list_count": 10,
      "p1_units_shipped": 0,
      "p1_units_total": 5,
      "p2_units_shipped": 0,
      "p2_units_total": 2,
      "p5_verified": false,
      "blocked_by": ["1.05"]
    }
  },
  "wave": 5,
  "wave_started": "2026-05-17",
  "next_review_at": "2026-05-24"
}

The integration agent updates this on every state transition. The watchdog agent reads it to detect stuck books (owner_locked_at older than 4h with no progress).

§8 Quality enforcement: the no-compromise contract

Restated as a contract the operator can point to when an agent (or themselves under pressure) is tempted to relax:

Never disable a 27/27 check. If a unit can't pass, fix the unit; do not patch the validator.
Never ship a unit without registered prereqs. Use pending_prereqs: true and queue the prereq unit instead of letting an edge dangle.
Never close a P5 verification without all 7 layers above threshold. Use equivalence-partial for partial closure and queue the gap.
Never skip Pass W on a batch. Skipping it creates a weaving debt that compounds across batches.
Never skip the weekly Pass V. This is the only metric that catches global drift.
Never let a substitute-text plan masquerade as a canonical one. §6.1 naming convention is enforced at audit-agent dispatch.
Never mark a reduced-audit book as equivalence-covered. §6.6.
Never let watchdog alerts sit overnight. A stuck book is a pipeline-blocker.
Never let quality-sampling reports go unread. Drift caught at 5% is recoverable; drift caught at 50% is a rewrite.
Never trade master-tier prose calibre for throughput. §6.7 is enforced at quality-sampling.

§9 Failure modes and recovery

Failure	Symptom	Recovery
Integration queue overflow	P4 units sitting in `draft` >4h	Pause P4 dispatch; let integration drain; investigate per-unit time
27/27 false-positive (validator passes a bad unit)	Quality-sampling catches it	Patch validator; re-run on shipped corpus; identify all affected units
Pass-W introduces forced anchor phrases (mechanical drift)	`weaving_report.md` shows >20% identical sentence templates	Pause Pass W; investigate weaving prompt; rerun with diversified phrasing
Substitute-text drift (no canonical plan written)	Audit ledger shows substitutes outnumbering canonical	Block new substitute audits; route capacity to the canonical text being substituted for
Stuck book (owner_locked_at >4h)	Watchdog alert	Operator inspects last agent transcript; resets owner; re-dispatches with bug-fix prompt
Author voice drift on flows	Flow-PD verification scores <4/5 on author voice	Run Pass W with explicit author-voice corrective; re-verify
Cost ledger spike	Daily token spend >2× sustained average	Throttle dispatch; investigate which agent / book is the cost driver

§10 Success metrics (review weekly)

Throughput: units shipped this week ≥ 80; books closed this week ≥ 5.
Quality: 27/27 first-pass rate ≥ 90%; quality-sampling avg score ≥ 4.2/5; Pass V all six metrics above threshold.
Queue health: integration queue ≤ 16 units; no stuck books; no unread reports.
Coverage: weekly delta in audits_done ≥ 6; weekly delta in shipped units ≥ 80.
Cost: sustained token spend within 1.5× the prior-month average.

If three consecutive weeks meet all five, increase concurrency budget (§2) by 25%. If any week misses two metrics, decrease by 25%.

§11 First concrete dispatch (executable today)

To start: queue exactly this batch.

P1 audit batch (5 agents, parallel, dispatched together):

Agent	Book	Source	Output path
audit-a	2.03 Chatterjee QFT	FREE PDF in `reference/fasttrack-texts/02-quantum-stat/Chatterjee-QFTLectureNotes.pdf`	`plans/fasttrack/chatterjee-qft-lecture-notes.md`
audit-b	2.02 Woit Quantum Theory	FREE PDF in `reference/fasttrack-texts/02-quantum-stat/Woit-QuantumTheoryGroupsRepresentations.pdf`	`plans/fasttrack/woit-quantum-theory-groups-representations.md`
audit-c	3.38 May Concise Algebraic Topology	FREE PDF in `reference/fasttrack-texts/03-modern-geometry/May-ConciseAlgebraicTopology.pdf`	`plans/fasttrack/may-concise-algebraic-topology.md`
audit-d	3.40 May Simplicial Objects	FREE PDF in `reference/fasttrack-texts/03-modern-geometry/May-SimplicialObjects.pdf`	`plans/fasttrack/may-simplicial-objects.md`
audit-e	3.42 Ravenel Complex Cobordism	FREE Green Book (author PDF)	`plans/fasttrack/ravenel-complex-cobordism.md`

Each audit agent's prompt template:

Subject: P1 audit of <book> (FT entry <#>)

Read <PDF path>. Produce a plan stub at <output path> following the
template in plans/fasttrack/brown-higgins-sivera-nonabelian-algebraic-topology.md.

Required sections:
  §1 What the book is for — distinctive perspective, 5-8 numbered points
  §2 Coverage table — every named theorem/concept mapped to a Codex unit
     or "GAP" with priority
  §3 Gap punch-list — priority-ordered new units to write, with three-
     tier estimates and notation crosswalk requirements
  §4 Implementation sketch — realistic hour estimate, originator-prose
     citations, notation crosswalk
  §5 What this plan does NOT cover — explicit non-goals
  §6 Acceptance criteria — per-book threshold for FT equivalence
  §7 Sourcing — local PDF path and license note

Do NOT write any content/ units. Audit only. Cite ≥3 sources beyond
the primary text. Hard limit: 3 hours of agent time.

When done, return:
  - path to written plan stub
  - punch-list count by priority (P1/P2/P3/P4)
  - any books that should block this one (e.g., prereq audits not done)
  - flag if substitute-text per §6.1 of AGENTIC_EXECUTION_PLAN
  - flag if reduced-audit per §6.6 of AGENTIC_EXECUTION_PLAN

P4 production batch (4 agents, parallel, dispatched concurrent with P1 audit batch):

Agent	Unit	Source plan
prod-a	`02.01.07` Compact-open topology and function spaces	brown-topology-and-groupoids.md §3 P1 #2
prod-b	`02.01.08` Cofibration and homotopy extension property	brown-topology-and-groupoids.md §3 P1 #3
prod-c	`02.01.09` Fibration (Hurewicz / Serre)	brown-topology-and-groupoids.md §3 P1 #4
prod-d	`03.12.00` Fundamental group	brown-topology-and-groupoids.md §3 P1 #5

Each production agent's prompt template:

Subject: P4 production of unit <id> — <title>

Source plan: <plan path> §3 entry <#>.
Write the unit at content/<section>/<chapter>/<id>-<slug>.md per
UNIT_SPEC.md §4. Three tiers required (Beginner / Intermediate /
Master) marked [Beginner] / [Intermediate+] / [Master].

Quality contract:
  - master-tier prose at Lawson-Michelsohn / Hörmander calibre
    (AGENTIC_EXECUTION_PLAN.md §6.7)
  - declarative sentences; no hedged futures, metacommentary,
    real-time error correction, valedictory closings, rhetorical italics
  - all named theorems proven (no [...] placeholders)
  - ≥3 cross-references in Master Connections, each classified by
    connection type per CONTINUITY_SCAFFOLD.md §3
  - ≥1 primary-literature reference beyond reference archive
  - Lean module if applicable; lean_mathlib_gap ≥30 words if not

After writing, run scripts/validate_unit.py on the unit. If 27/27 does
not pass, fix and re-run. Return the unit path and 27/27 result.

Hard limit: 4 hours of agent time.

Both batches run in parallel today. If both clear cleanly, the pipeline is calibrated and Wave 5 proceeds per §5 schedule.

§12 What this plan is NOT

Not a content production guide. Production-agent prompts deepen the punch-list units; this plan just orchestrates them.
Not a replacement for FASTTRACK_EQUIVALENCE_PLAN.md — it implements it.
Not a substitute for operator judgement. The three manual hand-off gates (P2→P3, P3→P4, P5→equivalence-covered) require a human (or driver agent acting in operator role) to read and approve.
Not a one-shot artifact. As Waves close and metrics arrive, this plan is amended (especially §2 concurrency budget and §10 success metrics).

§13 Open questions for the operator

Before kickoff, Tyler should decide:

Driver agent or self-drive? Does Tyler want to be the operator each day, or should this plan also specify a "driver" agent that performs the operator loop (§4) under sampled human review?
Concurrency budget calibration. §2 numbers are conservative first-pass estimates. After Wave 5 closes they should be re-calibrated against actual integration drain rate.
External expert review cadence. §6.4 puts expert review in arrears. Is there an actual expert (mathematician, physicist) to engage, and on what cadence?
BUY-only sourcing escalation. §6.6 specifies the reduced-audit fallback. Is there a budget line for purchasing the BUY-only physics texts (Landau-Lifshitz set, Weinberg set, etc.) to avoid running on reduced audits indefinitely?
Substitute-text remediation backlog. §6.1 says the 4 existing substitutes need canonical-text companions. Should those be inserted into Wave 5 (delaying Section 2 unblock by ~5 days) or deferred to Wave 7?

These questions do not block dispatch of the §11 first batch; they shape Wave 6 onward.