Teacher Operating System (TOS)

A set of AI assistants that help K–12 teachers create and double-check the everyday materials of teaching — lesson and unit plans, assessments and rubrics, slide decks, curriculum maps, IEP/504 supports, intervention (MTSS) plans, family messages, and coaching/PD resources.

Everything it makes is aligned to your state's standards, built with differentiation in mind, and flagged for your review — it's a strong starting point, not a final decision. It never uses real student information; examples use placeholders only.

What you can do with it

• Draft a standards-aligned lesson or unit plan
• Build assessments, rubrics, and answer keys
• Turn a lesson into a slide deck
• Map curriculum across a unit or a year
• Get differentiation and accommodation ideas for diverse learners (special education, English learners, gifted)
• Write family communication — and translate it
• Plan interventions / MTSS supports
• Prepare coaching and professional-learning materials

Who it's for

Classroom teachers first — plus special educators, instructional coaches, and school administrators. You don't need to be technical to use it.

How you use it

Ask in plain language — e.g., "Make me a 5th-grade fractions lesson aligned to my standards" — and the right assistant takes it from there. Two doors in, both written for non-technical readers:

• On Claude (Claude Code or the Cowork desktop app — same plugin; or claude.ai in the browser): implementation/claude/README.md — the full experience, including the setup wizard and your personal requirements map.
• On ChatGPT: implementation/gpt/web/README.md — one file to drag in, plus an optional Reference Pack with the verified Florida data.
• Verified tools inside either chat (optional, ~2 minutes): implementation/mcp/README.md — give your assistant eight read-only tools so it looks standards up instead of recalling them: search the verified corpus, check that a cited code is real, catch a misquoted standard, validate an artifact. Four ways in, depending on which app you use.

For the curious: docs/END_TO_END_WALKTHROUGH.md is a technical validation log of a real start-to-finish run (shell commands and all — not a getting-started guide).

Under the hood (for the curious or technical)

You don't need any of this to use TOS — but here's how it's built:

• A coordinator (skills/core/teacher-core) reads your request and sends it to the right specialized skill.
• Specialized skills do the work (lesson planning, assessment design, special-education support, and so on), grouped under skills/ as core/, educator/, operations/, and small single-purpose atoms/.
• Shared engines under shared/ give every skill one source of truth for standards, differentiation, and quality.
• A quality check runs before anything is called "done," and an automated consistency checker (tools/sync_check.py) keeps the shared rules each skill carries identical to the originals.

Verified standards (Florida)

A wrong standard code — or a real code quoted wrongly — is the most damaging mistake this system could make, so it is checked by code, not by eye.

# Does every standard this artifact cites actually exist, and does it fit the grade band?
python3 tools/verify_standards.py --input my-artifact.json

# Is this restatement of a standard faithful to the official text?
python3 tools/verify_standards.py --compare MA.3.NSO.1.1 --text "Read and write numbers to 10,000."

• Fully offline. It reads a committed corpus of 6,574 Florida standards (B.E.S.T. / NGSSS, including access points) — no network, no account, no API key.
• Honest about what it knows. A code absent from the corpus is reported as not_found and blocks — every Florida subject is corroborated against CPALMS above the trust threshold, so absence is evidence. (The old best-effort escape hatch survives in code as a safety net: if an overlay is ever reverted below threshold, that subject's absences degrade to advisory rather than accusing real standards of fabrication.) A standard Florida has withdrawn resolves as retired — a warning, never a "fabricated" verdict. CCSS and NGSS are scheme-only: structure is checked, existence is not.
• Checked against CPALMS, code by code — the whole corpus. Every subject and grade has been checked against CPALMS, Florida's official standards site, in both directions, including a reverse census that catches standards CPALMS has and the corpus lacks. All 6,574 codes match CPALMS's official text exactly (math 1,127 · ELA 719 · science 1,450 · computer science 560 · social studies 2,713 · ELD 5). Two real CPALMS standards absent from the source documents are recorded as provenance-stamped additions, and nine withdrawn computer-science standards as retired — flagged for review, never counted as verified. Results live in shared/standards/resources/florida/data/overlays/ with the CPALMS URL and check date for every entry; the parsed corpus is never overwritten. An adversarial audit re-proved every confirmed label offline from the stored texts before this paragraph was written (docs/audits/2026-08-13-sweep-completion-audit.md).
• What this proves, and what it does not. Our standards documents were themselves downloaded from CPALMS (sources.json), so a match proves parse fidelity — that our extraction of a CPALMS document says what CPALMS's own database says. It is not independent corroboration from a second authority, and a CPALMS-side error would be invisible to it. Florida's administrative rule documents remain the legal source of record.
• "Verified" means the text actually matches. The comparison used to accept a 97 % similarity score — about three characters of slack on a typical standard, enough to let a changed number or a deleted "not" pass as verified. It now requires the two texts to be the same text, ignoring only whitespace. There is no similarity band, no prefix allowance, and no minimum-length rule; anything short of equality is recorded as needing review, never as verified.
• Limits, stated plainly. Verification is against CPALMS, which is also where the source documents came from — so a match proves parse fidelity, not independent corroboration, and a CPALMS-side error would be invisible. Currency decays from the check date (checked_at per entry): CPALMS revised two benchmarks and retired nine standards in the window this work covers. The live, generated breakdown is ledger/cpalms-run-manifest.json (python3 tools/cpalms_verify.py --manifest); trust it over any number written in prose, including this paragraph. Full residual-risk statements: docs/audits/.

Maintainers refreshing the verification (needs network): python3 tools/cpalms_verify.py --subject math --grades K,1,2,3,4,5 --out report.json then --apply report.json to review, --write to record. Nothing is ever applied without that review.

Learn more

• docs/END_TO_END_WALKTHROUGH.md — a real teacher, start to finish (best place to begin).
• docs/ARCHITECTURE.md — how the pieces fit together.
• docs/QUALITY_MODEL.md — how outputs are checked for quality.
• protocol-layer/quality-gates.md — the rules every output must pass before it's considered ready.
• CLAUDE.md — conventions for developers working in this repo.

What these words mean (quick glossary)

• Skill — one focused AI assistant (e.g., the lesson planner).
• Coordinator (teacher-core) — the skill that decides which other skill handles your request.
• Quality gates — a checklist every output must pass before it's considered ready.
• Quality Ledger — a running log of those quality decisions, so there's a record.
• Drift guard (tools/sync_check.py) — an automated check that keeps each skill's copy of the shared rules identical to the originals.
• Decision support — you get a strong draft to review and adjust; you, the educator, always make the final call.