Tailoring Contract

Tailoring generates a resume that keeps as many applicable baseline accomplishments as possible within saved per-role limits from saved profile sources and the current employer analysis/requirement-fit ledger. Semantic Judgments Are LLM Determinations defines model authority. Materials Audit owns shared truthfulness and line provenance.

The Contract At A Glance

Work	Authority
Candidate fact, seniority, relevance, prohibited claim, voice and self-talk meaning	Structured model determination
Posting seniority and requirement scope	Accepted job interpretation
Artifact support	Separate claim-verification determination
Quality acceptance	Independent quality-judge verdict
IDs, verbatim quotes, exact numbers/dates/currency, schema, fences	Code
Fit bands, weights and caps over typed requirement verdicts	Code
Line-to-source and line-to-finding display	Persisted ID joins

Generation, factual verification, quality review and the high-fit persona review receive the same summary-quantity and selected-role rendering constraints before their first decision. A rewritten summary uses a quantity only when its cited achievement supports that exact claim; baseline-only quantities are omitted and tenure is not inferred from dates. Only selected experience updates render, so a summary using evidence from an omitted role must supply its employer attribution within the permitted mutable text. These constraints do not add roles, bullets, source permissions or layout changes. Quality reviews evaluate every supplied criterion and report all actionable defects together, while retaining their own pass/fail authority. The shared rules enter determination fingerprints; prior accepted receipts remain immutable.

Provider availability never selects a lexical path. A blocked refresh retains the last accepted generation and records its safe failure code.

Short Version

Canonical facts enter bounded prompts by ID. The generator returns strict JSON with explicit line anchors. The claim verifier independently reads the final lines and affirms their sources, claims and served requirements. The quality judge returns a separate typed verdict. Repair consumes the cited reasons; acceptance commits bytes, provenance and determination references together.

End-to-End Flow

1. Load the saved profile version, employer-analysis generation, accepted job interpretation, requirement-fit report and confirmed tailoring policy.
2. Build the artifact budget, required selections and allowed evidence inventory. Structural infeasibility is actionable; code cannot silently drop pinned evidence.
3. Generate strict line/claim mappings tied to known source and requirement IDs.
4. Assemble and check format, IDs, immutable fields and exact values.
5. Verify final claims and provenance with the model, then judge artifact quality separately. High-fit resumes (normalized fit at least 0.8) retain a separate, cited six-persona review. Its typed verdict decides acceptance; the persona score is diagnostic.
6. Repair from model findings within the attempt budget. No lexical gate or score cutoff replaces the verdict.
7. Extra post-approval voice polish is opt-in (resume_voice_polish, default false). With it disabled, retain the exact selected candidate and its factual, quality and applicable persona receipts, recording a skipped voice audit linked to those reviews. Explicit injected voice ports retain their requested behavior. If voice is requested, the model chooses wording changes; changed text receives fresh claim verification, quality and applicable persona review. An optional rewrite provider or shape failure records a rejected voice attempt and retains the already-verified candidate. A failure of verification for changed text blocks that refresh.
8. Commit the accepted generation and audit together. A failed attempt cannot replace the accepted resume/PDF or its approval.

Local Application Boundary

scoring/tailor.py resolves configured model choices and explicit request overrides into the frozen TailoringLlmPolicy before constructing TailorResumeUseCase. The use case requires this value; it never reads ambient configuration. Saved profile controls and versioned policy history remain canonical inputs, with the existing profile/policy fence at generation commit.

Policy restore is a separate reviewed command. Its tenant, target, expected current version, UUIDv4 command identity and user-requested reason bind the intent. The SQLite writer checks for a matching receipt before comparing the current head under BEGIN IMMEDIATE. A stale new intent refuses; an identical accepted command returns its original revision and the head observed on retry. The new revision stores its versioned command receipt in its rollback provenance field together with the selected target and reviewed predecessor. This keeps receipt and revision atomic within the existing exact-v15 manifest and leaves historical rows and configuration copies intact. History readers decode only recognized structured receipts and preserve legacy provenance. A different payload for the same tenant and command ID is a conflict.

An explicit manual materials request carries its existing low-fit exception through both Tailor and Cover activity payloads to their canonical per-job admission checks. Default and legacy payloads keep the ordinary score floor. The exception changes only the numeric threshold for that request; hard eligibility blockers, current-score and posting checks, artifact ownership, attempt limits, claim verification and quality review still apply. PDF exports consume the accepted materials without changing the score or global policy.

Preparation targets carry the latest canonical Score grade and version, joined by tenant and JobId. The legacy grade on jobs cannot choose review behavior. The target passes its saved grade to the resume review gate, so a normalized fit of at least 0.8 receives its independent six-persona review. A manual low-fit exception leaves that grade unchanged. Unscored targets remain readable for scoring; material admission still requires a current Score.

The generation graph distinguishes selected achievements from canonical context. Its required_generated_evidence_ids manifest requires selected achievement claims and explicit user pins. Role dates, role metadata, education, skills and baseline context remain available sources without forcing them into new summary prose. This structural ownership flag does not decide relevance or grant shipped coverage: the final model verifier must affirm every declared source contribution and served requirement. Unused context cannot become an extra provenance anchor, and an unrepresented context edge receives no automatic generated-coverage credit.

The plan stores baseline_evidence_ids_by_role and the configured max_experience_bullets, separately from explicit pins and target coverage. Each baseline ID of a selected role must occur once in a bound experience claim or once in experience_omissions. Every omission proposal supplies its exact role/source IDs, a rationale and the typed reason clearly_inapplicable or role_bullet_limit. Code rejects missing, foreign, duplicate and contradictory decisions, dropped explicit pins and cap omissions when the role is not full. The configured maximum remains a hard cap; baseline length never raises it.

Generation keeps as many applicable accomplishments as possible up to that cap. The existing independent quality determination reads the full baseline inventory, final lines and omission proposals. It rejects unjustified omissions, underfilled roles and weaker evidence chosen over more useful available proof. Missing job keywords or an active fit edge do not establish irrelevance. Code does not decide applicability or infer semantic equivalence. Multiple related baseline sources and at most one additional achievement source may share a coherent bullet only when independent factual and quality reviews affirm their contributions and preserve agency, action, scope, outcomes, tools and metrics. Retention never grants target coverage or an explicit user pin. baseline_retained identifies a retained factual claim without a target edge. A fully inapplicable baseline can leave role metadata alone after its omission proposals pass quality review.

These selection proposals are part of the existing generation response and quality request, not another paid model stage. Their IDs and decisions remain in the attempt audit; the quality receipt remains authoritative. Current prompts use tailor.v18.applicable-baseline-retention and tailored-resume.v6. Existing accepted artifacts and historical plan metadata remain unchanged.

An overflowing seeded graph is still an evidence menu. If explicit user pins fit but the seeded selection exceeds the configured per-role bullet ceiling, Materials requests a material_evidence_selection determination from the configured generator route before drafting. It selects only admitted requirement/evidence edges, cites their canonical sources, and records a reason for each requirement left uncovered. Code enforces exact IDs, pins and the configured ceiling; the existing artifact-budget guard runs again on the selected plan. The plan audit stores the determination ID. Feasible seeds make no additional call, impossible pins still block before spend, and failed selection preserves accepted materials. Selection never establishes shipped coverage or replaces independent reviews.

Exhausted factual or quality correction returns the non-retryable material_rejected error. It preserves the rejected audit and last accepted artifacts; Temporal does not restart the whole correction budget as though a provider transport had failed. The durable Tailor/Cover stage flag and failure event retain that same non-retryable classification, so automatic preparation cannot start another complete correction cycle after the workflow has stopped. An explicit new request remains available within the durable attempt limit. Adapter retry authority also survives the safe semantic failure protocol: provider_non_retryable stops automatic material retries, while provider_error retains transient recovery. Structural resume rejections pass their candidate and exact validator errors to the next bounded attempt as untrusted user-message repair data, even without a paid-review receipt. They never become generator system instructions. Provider/network faults retain their separate transient retry behavior when classified retryable. A new explicit request can use the existing recovery controls after the owner addresses the recorded finding.

Selected Tailor and Cover cohorts isolate typed per-job prerequisite failures that occur before a stage claim. Missing inputs skip that member, and temporary source unavailability remains a bounded item error while healthy peers finish. This boundary does not claim or terminalize an unowned stage. Cancellation, spend-budget failures and unclassified exceptions still propagate; accepted artifacts and the existing ownership fences remain authoritative.

The effective judge model binds the claim-verification and quality/persona adapters, independently of the generator model pool. Default review shares the operation's default adapter. An explicit reviewer selection binds that route and records its actual provider/model; unavailability fails visibly without substituting the generator adapter. Explicitly injected review ports retain their own authority and are not replaced by composition.

The same composition root injects ArtifactTextWriter, implemented locally by LocalArtifactTextWriter. This small capability creates directories, writes exact UTF-8 text and reports its byte size. The use case retains candidate selection, rejected-candidate inspection content, generation filenames and side-file formatting. Validation, selection, evidence construction and aggregate transitions do not perform file I/O. The writer does not register or accept artifacts: MaterialsRepository registers the generation and UnitOfWork preserves the atomic acceptance/provenance transition. PdfRendererPort still renders before that transition. Candidate write errors propagate before acceptance; diagnostic _JOB.txt and _REPORT.json writes remain best effort after persistence.

Inputs To Tailoring

Inputs are versioned profile facts and confirmed evidence, canonical posting text, accepted job interpretation, employer requirements, model-backed requirement-fit verdicts and user-authored selections/policy. Unconfirmed evidence cards, interview notes and new recollections do not become facts. The prompt does not dump credentials or unrelated private profile fields.

The Actual Model Prompt

The generator and review rubric serialize the tailoring plan as compact JSON. When the complete ordered achievement catalogues in the target profile and coverage graph are identical JSON values, the prompt retains the catalogue once and replaces the duplicate with an explicit reference local to that plan object. Different or empty catalogues stay inline. This encoding preserves every field, record and source ID, does not choose evidence or remove facts, and leaves the canonical plan and persisted metadata unchanged.

Prompts contain minimized canonical source IDs/text, allowed mappings, artifact constraints and repair findings. All source text is untrusted data. Lane binding and BR-050 preflight precede calls; envelopes record schema/prompt versions, provider/model and input fingerprint.

Output Schema

GeneratedResumeDraft contains explicit summary sentences, experience updates, skill updates and generated_claim_mappings. Each mapping records a line ID, location, evidence/requirement IDs, transform and rationale. Code rejects foreign IDs, duplicate IDs and locations that do not map to the shipped line. The verifier cites exact final text and sources; generator assertions alone never establish support or coverage.

Generation describes the renderer's literal bindings: every summary sentence uses executive_profile#0, bullet N uses experience:<entry-id>#N, and a selected skill group uses skills:<category-id>#0. The prompt supplies the exact evidence IDs belonging to each experience entry. The response schema permits only claim labels allowed by the active claim policy; validation independently enforces the same policy and checks every emitted binding.

What "Profile-Row Based" Means

Authored experience, education and skill rows remain canonical. Exact row IDs control selection and rendering. Reframing prose requires source-bound model verification; immutable employer/title/date fields retain literal binding. Selection does not convert a draft evidence card into confirmed history.

Tailoring Plan

target_seniority comes from the accepted job interpretation. Requirement scope and protected-class flags come from that interpretation, without regex overrides or old-analysis fallbacks. required_evidence_ids, selections and budgets are structural/user authorities. Requirement-fit evidence IDs must belong to the supplied profile inventory. No substring “seniority signal” or token overlap adds evidence or coverage.

Attempt Loop And Candidate Selection

Each candidate attempt receives a fresh opaque UUID before provider I/O. The ID stays with its parse/provider failure, reviews, correction feedback and persisted selection. Repeated or common-prefix prompts and parallel attempts still create distinct candidates. The complete request has a separate prompt fingerprint; candidate identity does not decide semantic equality or invalidate determination reuse. Historical IDs remain as recorded, scoped by their generation and attempt.

A generated candidate must pass structural binding, claim verification and the separate quality judge. Model findings supply repair reasons and exact line IDs. Arithmetic may compare accepted typed fit results; it does not assess prose. Provider, spend, JSON, schema and source-binding failures remain distinct and preserve accepted material.

For a structurally valid résumé draft, production factual verification and quality review share the bounded review pool and run concurrently on independent thread-owned database connections. Facts remain the authority for provenance. A quality verdict is used only after every factual/local gate passes and the complete quality request (artifact kind, entity, ordered lines and allowlists, sources and rubric) still has the same fingerprint. A changed request gets a fresh quality determination; a rejected candidate records the unused receipt. Its completed quality findings are retained separately as parallel_quality_feedback for the next generator repair. They do not fill the candidate's judge record, enter accepted review history or authorize approval. Repair evidence selections still intersect the current canonical retry catalogue; a later final judge takes precedence over speculative feedback. High-fit six-persona review still follows the accepted pair. Custom ports and in-memory databases retain sequential execution. Cancellation and shutdown drain admitted siblings before releasing resources.

Bounded Persona Diagnostics

The final high-fit review still assesses every supplied line and the complete rubric from all six personas. New resume-adversarial-v3 requests use a bounded response model: each persona records one explicit verdict, up to three cited source spans and six actionable findings. Rationale and repair strings are limited to 512 characters, quotes to 240, and each citation's literal-value list to four values of at most 64 characters. These limits control response size; they never compute a semantic verdict or permit a pass when a defect remains. The reviewer groups repeated defects, prioritizes repairs and retains fail for any unresolved issue, including issues summarized in its rationale.

The provider receives these nested bounds in its structured-output schema. Structural, source-ID and verbatim-span validation plus the existing bounded correction budget still apply. Every persona and the overall model verdict must pass before acceptance. Scores remain diagnostic. The canonical stored review reader retains its historical wider shape, so old accepted receipts and their original diagnostics remain readable without truncation or rewriting. The prompt revision and response schema participate in determination identity. The response model explicitly names its wider stored_result_schema; generated TypeScript readers publish that canonical shape and current/compatible prompt versions rather than applying new response-size limits to historical receipts.

For each cover-letter draft, factual verification and quality review read separate copies of the same frozen lines, sources and rubric. The production runner overlaps these independent calls in the bounded letter-review pool. Both durable results join before acceptance or an ordered correction. Review IDs, cache fingerprints, model/effort, citation validation and attempt budgets stay the same. Resume checks retain their dependencies on factual grounding and pruning. Custom directly injected reviewers and in-memory repositories run sequentially.

Each bounded resume or cover correction carries an ordered review history: recorded determination IDs and their complete typed quality/persona results. Later generators and reviewers receive that history, including after a changed voice proposal. The reviewer considers earlier repairs and explains any reversal with a feasible correction addressing the competing concerns. Its independent typed verdict still decides acceptance; history cannot force a pass or expand generation permissions.

History is context, never an additional factual or citation source. Current canonical sources and rendered lines retain their exact ID/quote fences. The history participates in determination fingerprints, so changed review context requires a new decision while unchanged factual inputs can reuse verification. Resume attempt audits retain per-candidate and final review context; cover artifact metadata retains the history supplied to its final review. Rejected voice review context remains in the voice audit. Existing attempt/spend limits and accepted-artifact preservation apply unchanged.

Quality and persona reviewers receive the canonical posting plus the same user generation permissions, required roles, bullet limits and current coverage plan as the generator. Their shared rubric also carries the exact canonical evidence IDs eligible for a bounded revision. Review requests cannot authorize a new coverage edge or expand the artifact budget. A required role without active coverage or a bullet pin retains its single positioning-bullet constraint; additional saved evidence outside the permitted revision set cannot become an impossible drafting requirement.

Empty-bullet permissions use the plan's actual achievement inventory, including authored baseline bullets as well as confirmed achievement-ledger records. Standalone prompt construction and field validation share the canonical profile source inventory. Fixed role metadata and unconfirmed ledger records do not become achievement evidence. A required role with baseline evidence cannot be listed as allowing empty bullets solely because its separate ledger is empty; a role with neither evidence nor a bullet pin preserves its fixed metadata.

Required skill-category IDs are a minimum, independent of the full allowed source catalogue. The first prompt exposes all authored categories. Field validation checks required inclusion against that minimum and foreign IDs against the entire catalogue. Text and ResumeDocument share the selection of required plus model-selected known categories in source order. Individual skill pins require their owning category; existing default-all behavior and item-rewriting permissions remain. Source and quality determinations still decide factual support and relevance. Quality correction IDs are checked against admitted revision targets plus canonical sources already bound to the verified rendered candidate. The full profile remains available for factual review; its mere presence cannot grant a new coverage edge. An out-of-scope correction gets exact private ID feedback in the existing bounded repair, without changing the model's quality verdict. Review permissions also record renderer-owned layout and the exact fixed skill category order. Skill-item reordering never grants category or section-layout edits; those remain under the saved profile and template. Seniority does not imply direct reports, and omitted optional roles alone do not require a complete career inventory. Reviewers assess the truthful artifact; unsupported qualification gaps remain gaps rather than requests to invent proof. Repairs receive the rejected typed candidate and exact unaffirmed source IDs as untrusted data, with the current source and coverage allowlists remaining authoritative.

Validation Layers

The canonical contact formatter keeps professional-link destination hosts and paths visible in both text and PDF labels. Two platforms may use the same handle; display text must still distinguish the saved links while preserving their exact hyperlink targets. Generated edits do not change contact facts.

1. Structured Output Schema

Strict schemas reject extra fields, unknown enums and missing inventory items.

2. JSON Field Validation

Code validates row/line membership, immutable fields, explicit selections and numeric/date/currency binding. Self-talk, named-technology claims and register are model decisions.

3. Resume Assembly

Rendering preserves stable line IDs through text, HTML and PDF layout boxes. No reader reconstructs sources by matching strings.

The shared HTML print renderer repeats the selected template padding on every physical page fragment. A4 and Letter continuation pages retain their top and bottom insets without shrinking the font or changing accepted text. Section headings stay with their following content across a page break. Content may still flow onto another page; inset preservation does not impose a page count or establish the accuracy of the pre-print layout-box approximation.

4. Rendered Resume Validation

Sections, layout and quote existence are mechanical. Meaning is assessed against actual final text.

5. Structural Tailoring Checks

Required IDs, pins, artifact budget and exact-value checks remain code. Representation of an achievement, seniority alignment, stuffing and style remain model judgments.

6. Claim Verification

Every final line receives a typed verdict, factual propositions, verbatim supporting evidence, source contributions and findings. Unsupported/uncertain factual claims fail. Generation anchors must reconcile with verifier-affirmed source IDs. The technology/title/employer lexicons and stemmer are deleted.

The factual response uses request-local short citation references, expanded into full canonical evidence before validation and persistence. Every line and claim is still reviewed. Whole-source selection is explicit and bounded; final-line claims and findings require verbatim spans. The Materials audit contract owns the format, cache and historical-reader guarantees.

7. Post-Generation Fit Gate

Coverage counts only requirements the verifier declares served by shipped line IDs. The requirement-fit policy owns arithmetic and bands over those typed verdicts. No keyword union launders provenance.

8. Structured Judge

The existing quality judge remains a separate call. Its typed pass/fail verdict controls quality acceptance; scores are diagnostic rather than a semantic threshold.

9. Adversarial Review

Any additional model review uses canonical sources, explicit line IDs and strict findings. Code never reclassifies its prose by keywords.

10. Optional Voice Pass

The model chooses prose changes by stable line identity. Skill/structural inventories remain unchanged. Final changed lines require fresh claim verification and quality review; no buzzword density or structural-variety proxy decides adoption.

Persistence And Audit Data

Accepted artifacts carry claim-verification and quality-determination IDs. artifact_line_anchors records line, evidence, requirement, transform and reason. semantic_determinations holds accepted envelopes; bindings fence entity versions. Blocked attempts are separate from accepted artifacts. Apply Review joins anchors/findings by line ID and displays “No recorded source” for unanchored lines.

Deterministic Versus LLM-Owned Work

Code owns formats, identity, schema, source binding, security and arithmetic. Models own support, relevance, voice, seniority and requirement demonstration. Literal user filters remain literal. No read-side lexical matching is allowed.

Current Constraints

A configured provider and spend allowance are required. The offline demo reports generation and verification unavailable. Catalog browsing, authored edits and historical reads remain available under their existing safety boundaries. Merge and release remain owner-authorized actions.

Common Failure Reasons

Provider unavailable, budget denied, malformed JSON, schema violation, foreign IDs, non-verbatim quotes, mismatched values and stale source versions are distinct actionable failures. An unsupported model verdict drives repair or failure with its cited reason. A refresh never erases the accepted generation.

How To Change Tailoring Safely

Update the owning schemas/prompts and source contracts together. Prove the model controls the result using opposite valid fake-model verdicts on the same sources. Exercise each failure without fallback, last-accepted preservation and actual product persistence. Do not add eval sets, sentence corpora, baselines or recorded-output replay fixtures.

Outreach Draft Gates (Reused Materials Stack)

Outreach calls claim verification through a port, using confirmed contact facts and canonical profile sources. An independent quality judge assesses tone. User edits receive the same model checks before approval. The user confirms a draft and records sending separately; generation never sends a message.

Outreach Publication Boundary

SqliteOutreachThreadRepository.load reads the complete tenant/thread row, all draft generations and all send logs in one SQLite snapshot. An immutable, exact canonical-state tuple accompanies the aggregate through its replacements; expected absence is distinct from overwrite authority. Generation captures this base before synthesis, and revision before verification. Read transactions end before model calls; model authority, lane/spend preflight and offer/profile source fences keep their existing owners.

At save, SQLite writer ownership precedes comparison. A changed or deleted base, or competing first creation for the same tenant/contact/job, raises OutreachPublicationConflict before canonical mutation. This migration-free comparison uses full canonical rows, not timestamp equality or a process-local counter. Unrelated tenants and threads do not invalidate the base. Concurrent revisions from one base can publish at most one next generation.

Persistence inserts new draft and send-log identities, updates existing draft lifecycle columns explicitly, and never deletes/recreates accepted history. Bodies, gate/provenance records and prior review evidence remain immutable. Canonical changes, new determination anchors and durable lifecycle events commit as one transaction; subscriber fanout runs best effort after commit. A failed publication rolls them all back. Completed model determinations have their own persistence lifecycle and can survive a conflict without becoming draft authority.

TypeScript approval/rejection, send logging and reminder transitions acquire writer ownership before loading or validating the current state, including offer bindings and approved-draft requirements. Their fact and event writes share the same transaction. The Python RPC transports a stale proposal as a separate result variant; the API maps it to HTTP 409. No accepted ID or generation is assigned to the returned private proposal. The web editor is mounted independently of the current review card, retaining text and completed results across refetch/SSE, and asks for deliberate reconciliation before another verification attempt. When a revision succeeds, the editor closes only if its current input still matches the submitted input snapshot. Newer typing stays available and unverified; the completed request never submits those later edits automatically. Open-page sessions partition the mounted review/editor state by tenant, contact and optional job. Inactive recipient/job sessions remain hidden and inert, including their reconciliation controls; returning to the original target resumes its work. Pending callbacks retain that target's own mutation context. Tenant changes reset these sessions. Generation also partitions by purpose and offer revision, retains its private conflict separately from the mutation error and presents the completed body, checks and provenance. Another generation requires explicit comparison, and a failed deliberate retry preserves the last proposal. Neither generation nor conflict recovery automatically retries a model call. This retention ends when the panel unmounts or the page reloads. The protocol does not add contact edit CAS, command replay identities or any send transport.

Key Code Pointers

• domain/materials/claim_verification.py and domain/ports/claim_verification.py.
• domain/materials/artifact_quality.py, generation.py, provenance_builder.py and claim_grounding.py.
• domain/enrichment/interpretation.py, domain/profile/canonical_sources.py.
• infrastructure/determinations.py, Materials repositories and apps/api/src/semantic-determinations.ts.

Legacy confirmed achievement rows retain saved cells and confirmation history, but an unresolved confirmation_review_required marker fails canonical source construction with legacy_evidence_review_required. A refresh cannot silently substitute an empty evidence inventory or replace accepted material IDs, bytes or anchors. Only a current-state-bound explicit human review resolves this structural authority; see Storage.