Solutions
Engine
Core Pipeline
Intelligence Modules
Every engine version is a milestone in deterministic cinematic compilation. SRC, PAIL, COS, Frame0 — each module evolving with real production data.
25+ versions. Each solving problems discovered in real story compilation. Version-pinned outputs mean your stories compile the same way, every time.
Engine 7.2.8 — Anchor-Centric ACTION
Each engine release represents a milestone in StoryDirector's pursuit of deterministic, production-ready AI video generation.
Prevents video models from silently applying ACTION verbs to the anchor when the narrative agent is a different character or faction. Mechanical detector scans for the earliest-mentioned cast entity in narrative_intent; if the first mention is not the anchor, a low-temperature LLM call rewrites the action from the anchor's reactive / witnessing perspective.
New stories use the active engine. Existing stories remain pinned to their original engine.
Engine switches are audited and cache-invalidated to protect consistency.
Each engine version represents a milestone in our pursuit of deterministic, production-ready AI video generation.
Scene 8/1 Hessian (Washington Crossing) on Gen-4.5 post-7.2.7: wardrobe rendered perfectly (mitre cap + all 6 MUST WEAR items) but the Hessian anchor joined the Continental Army charge instead of defending against it. Root cause: ACTION text was scene-level narrative ("The Continental Army charges forward...") with a grammatical subject different from the shot's anchor. Video models bind ACTION verbs to the ANCHOR character regardless of stated subject — so the Hessian ran the "charges forward" verb. This failure mode would have contaminated 8.0 chained-i2v validation by mixing subject-assignment drift with chain-mechanic questions.
The launch gate before 8.0. 7.2.7 validated the compression layer. R3 (tension + full wardrobe + mitre cap) isolated the remaining quality gap to subject-assignment in ACTION. Most multi-character scenes have an anchor that is NOT the narrative agent — reaction shots, witness shots, fleeing/hiding/observing — so fixing this unblocks a large class of scenes from launch-quality rendering. Three-way review (Derek + Claude Code + Opus 4.7) converged on: mechanical faction-conflict detector as primary path, LLM judgment only for ambiguous cases, three-template dispatch to avoid low-temperature conditional-logic drift, feature flag for rollback safety, 6-shot regression battery including a same-faction guardrail to catch false-positive inversions.
Prevents video models from silently applying ACTION verbs to the anchor when the narrative agent is a different character or faction. Mechanical detector scans for the earliest-mentioned cast entity in narrative_intent; if the first mention is not the anchor, a low-temperature LLM call rewrites the action from the anchor's reactive / witnessing perspective.
Scene 8/1 post-7.2.6 Hessian rendered without the signature brass-fronted grenadier mitre cap because MUST WEAR was unconditionally stripped. Gen-4.5's training priors for "Hessian Mercenary" reconstruct "soldier in period coat" but not the distinctive Hessian grenadier silhouette. Multi-faction shots (Continental vs Hessian) lost visual differentiation. This regression on non-iconic content would have contaminated 8.0 validation signal by mixing compression-artifact identity drift with chain-mechanic questions.
Users' content distribution will skew heavily non-iconic (fictional characters, original settings, less-canonized historical figures). Washington-Crossing-the-Delaware-level training priors are the exception, not the rule. 7.2.7 makes the engine handle the rule case correctly while preserving the 7.2.6 quality gains on exceptional content. The prior-strength heuristic uses a curated list of high-prior historical figures (~30 names) + iconic visual-canon events (~20 phrases) + period-band scoring. Multiplicative threshold: character_prior × scene_prior >= 6000 → drop wardrobe; >= 2500 → compress to first + last 2 items; otherwise preserve in full.
Makes 7.2.6's smart compressor aware of character and scene training-data priors. Wardrobe now preserves for non-iconic content where models cannot reconstruct it from context; compresses for iconic historical figures where models can.
The April 22 7.2.2 reanalysis of the Washington Crossing draft surfaced four render-path regressions: Scene 10 Shot 1 and Shot 2 rendered with CINEMA and PHYSICS lines dropped entirely; Scene 9/2, 11/1+2, and 12/1+2 narratives ended with "…" truncation; Scene 10 ANCHOR rendered "For; 1776." because the narrative opens "For the first time in months..."; Scene 4 Shot 2 ANCHOR rendered "continental_soldier; ..." leaking the raw handle slug. All four share the same property: they did not appear in 7.2.1 but 7.2.1 did not introduce them — 7.2.2 simply produced canon shapes that hit them simultaneously.
Without line-terminator hygiene in buildBlockingLine and buildCinemaLine, PromptLengthGuard::ensureSafeEnding was chopping entire output lines back to the previous period with no log. Without the visual_bible.cast fallback in cascade step 2, any faction-kind cast entry leaked its slug into ANCHOR. Without the non-pronoun function-word blocklist, any narrative opening with a preposition or conjunction leaked that word into ANCHOR. The new Log::warning on ensureSafeEnding silent drops is the bigger structural win — this class of bug is invisible until you diff exports by hand, and the warning will surface the next instance at monitoring time.
Closes three latent render-path bugs exposed by the 7.2.2 Washington Crossing reanalysis — unterminated BLOCKING/CINEMA lines silently dropped, blocking-spec physics ignored, handle slugs leaking as subjects, function-word subjects leaking from narrative-intent cascade.
The April 22 Washington Crossing pass on 7.2.1 surfaced one cosmetic regression: Scene 1 ANCHOR rendered as "It; river; night; 1776." because the cascade's narrative_intent regex captured the leading pronoun "It" (the narrative opens "It was the night of December 25, 1776..."). Bare pronouns are meaningless as subjects but fooled the proper-noun-shaped extraction. Without an explicit blocklist, any narrative starting with "It"/"He"/"She"/"They" would silently degrade ANCHOR in the same way.
The pronoun and placeholder fixes are small, but they close the last visible "ugly subject" cases in the storyboard export. The regression test is the bigger structural win — Engine 8.0's chain-and-reset architecture will faithfully preserve any beat-drift or multi-faction-ordering bug into cohesive-looking but narratively wrong sequences. A programmatic guard at this layer catches regressions before they propagate through the chain.
Final polish on top of 7.2.1 — pronoun-aware subject extraction, generic-placeholder strip at render, plus the first automated regression test for beat distribution and multi-faction ordering. Sets the foundation for 8.0's chain mechanics.
The April 2026 Washington Crossing video pass surfaced three specific issues: Scene 11 and Scene 12 shots rendered content that actually belonged to Scene 7 and Scene 8 (beat drift — content shifted ~4 beats forward across the story). Scene 9 Shot 1 rendered with ANCHOR "soldier; river; 1776." and no IDENTITY LOCK, forcing the video model to invent a random soldier archetype. Scene 8 Shot 2 rendered with George Washington as the subject but no IDENTITY LOCK directive, leaving his wardrobe unenforced on that specific shot. All three were silent failures — the prompts looked reasonable but the orchestration had gaps.
Beat drift is the single most impactful story-to-shot issue for any narrative exceeding two scenes. When Scene 11 shows content from Scene 7, viewers immediately sense something is wrong even if they can't articulate what. Fixing it is the difference between "AI-generated clips strung together" and "a cohesive short film." The faction and named-character IDENTITY LOCK relaxations close the last remaining paths where wardrobe drift could sneak in — subjects that the distribution LLM tagged implicitly but not explicitly.
Hotfix release before Engine 8.0. Tightens story-to-shot mapping: beat content no longer drifts between scenes, faction wardrobe enforcement fires reliably across group subjects, and named character identity locks activate even when the distribution LLM forgets to tag them.
Earlier engines relied on generic prompt templates that inserted stock phrases like "An establishing shot captures the scene" and "Medium coverage of action/dialogue" between your story and the AI video model. These templates produced lifeless, repetitive output regardless of what your story actually described. Shot counts were inflated — simple moments were padded with 3-7 redundant angles. Every shot defaulted to the same wide framing with the same slow camera push, giving a uniform, directionless feel across every story. Character references leaked into prompts unresolved, and physical descriptions were one-size-fits-all regardless of environment.
Engine 7.0 is the first release with genuinely production-quality prompt output. The generated direction is specific enough that filmmakers no longer need to rewrite prompts before they produce watchable video. Fewer shots means higher-quality coverage per beat — the pipeline stops padding simple moments with redundant angles. Every shot earns its place through narrative context, not template assignment. Metadata and attribution lines are filtered out before they can contaminate the narrative, so the first beat is always real story content.
Direct narrative, zero templates. Every shot earns its place. The enriched beat narrative IS the prompt — no more generic boilerplate standing between your story and the AI. Shot count reduced 60-70%. Shot type and camera movement vary by beat content. First release that feels ready for real users.
Legacy JSON-based shot specs lacked relational integrity, cinematic enrichment, and narrative coverage guarantees. Exports showed empty prompts for v6.5 drafts.
Every shot now has a canonical prompt object enriched with cinematic metadata, ensuring full narrative arc coverage and consistent video generation across providers.
Engine 6.5 introduces the full coverage + continuity system: relational Shot model with PromptCanon, cinematic enrichment (act positioning, energy curves, emotional states), LLM-powered action distribution across the full narrative arc, motion translation with physics-aware verb dictionaries, and a video strategy resolver for provider-agnostic generation. Exports now include compiled video generation prompts per shot.
Previous engines lacked deterministic grammar rules for shot entry and scene structure, leading to inconsistent cinematic patterns.
Every shot now follows a cinematic grammar contract, producing professional-grade scene structure with deterministic camera and energy management.
Engine 6.0 introduces cinematic grammar intelligence: PAIL enforces scene entry contracts (ESTABLISH → PRE_ACTION → ACTION), Pro Detection respects author-written pre-action, intent density scoring quantifies narrative richness, and the director briefing system explains every inference decision. Built on the stable 5.2 orchestration foundation.
Engine 5.1 proved that cinematic intelligence (emotional arcs, camera grammar, motion contrast) makes stories feel professional. But the system still lacked global structure: no act-level pacing, no energy curves, no film grammar enforcement. Stories could peak randomly, aftermath shots could be absent, and physics language could trigger model guardrails. Engine 5.2 solves all of this with a deterministic orchestration layer.
StoryDirector transitions from a prompt assembly system to a cinematic orchestration engine. Every story now has professional film structure: acts with energy envelopes, escalation curves that build to a peak, physics grounded in cause→effect, and geometry that persists from thumbnail to video. This is the category shift from AI prompt generator to cinematic intelligence layer.
Engine 5.2 introduces deterministic film grammar: Act-driven structure, shot role classification, visual blocking generation, physics consequence modules, escalation curve enforcement, and thumbnail-video geometry binding. Every story is now orchestrated through professional film grammar principles.
Engine 5.0 proved end-to-end video works. But shots felt flat — actors posed independently, camera movements repeated between scenes, emotional arcs were absent, and detail density was uniform. Engine 5.1 addresses all four weaknesses with a deterministic intelligence layer.
For the first time, StoryDirector produces video where every shot has an emotional direction, actors react to shared events, cameras move differently between adjacent shots, and detail scales with duration and intensity. The system moves from "works" to "feels cinematic."
Engine 5.1 adds deterministic cinematic intelligence: emotional progression, multi-actor coherence, motion contrast, narrative density calibration, and scene camera profiles. Every shot now carries emotional context, reaction chaining, and camera grammar — all deterministic, all version-gated.
Unified the entire story-to-video pipeline into a single, production-safe workflow. Scene video generation, final cut stitching, and Director Output with View Video—all working end-to-end.
This release crosses the viability threshold. The system is no longer a set of experiments—it is a coherent creative workflow that produces a full story output.
Engine 5.0 establishes StoryDirector as a real production pipeline. Story → scene video generation → final cut stitching → Director Output with View Video. This release crosses the viability threshold: the system is no longer a set of experiments—it is a coherent creative workflow that produces a full story output.
Architectural cleanup certifying the engine for video generation. Eliminated split-brain authority, silent failures, dead code, and parallel builder families.
Established a clean, deterministic, and observable foundation that all future video generation work builds on.
Release notes coming soon.
Raw prompts lacked explicit physical action.
AI video models now receive clear motion directives.
Every clip now contains an actionable, animatable Action line.
Added Scene Reality Check (SRC)—a validation layer that verifies scene specifications against story intent before generation.
Catches specification drift and logical inconsistencies before they reach the generation pipeline, reducing failed outputs.
Release notes coming soon.
Rebuilt the prompt assembly layer to support layered narrative intelligence, separating story intent from visual specification.
Dramatically improved output consistency and reduced prompt drift in complex multi-scene stories.
Release notes coming soon.
Deterministic compilation. Version-pinned outputs. Production-grade shot specs from your story description.
